# Generative Gap Filling

Canonical page: https://works.battleoftheforms.com/papers/generative-gap-filling/

GENERATIVE GAP FILLING
Yonathan A. Arbel
David A. Hoffman
Contract law polices a line between interpretation, the recovery of
meaning a text already holds, and gap filling, the supply of terms the text
lacks. The boundary rests on an unchecked premise: that courts need to
fill gaps themselves because the document has run out of meaning.
We tested it. Borrowing a masking design from machine learning, we took
real contracts, hid a term the parties had negotiated, and asked three kinds
of readers to predict what we had removed: ordinary people, legally trained
ones, and several large language models.
Human interpreters guessed right a bit more than half the time, doubling
the rate of chance alone. Law students edged out lay readers, and
practicing lawyers were mostly better still. Then came the machines.
Given nothing but the rest of the contract, the models predicted the
missing term nearly nine times in ten. Contracts, we conclude, are like
radio signals: even when incomplete, enough of the message is carried
elsewhere that the missing part can be reconstructed with the right
receiver. These findings speak to a line that has long bedeviled contract
theory. The hypothetical bargain, treated for a century as invisible and
unruly, turns out to be statistically legible, and models can read it in the
open. Courts can weigh these predictions in adversarial settings, and
parties can discipline the practice by adopting “Choice of Model” clauses.

William Alfred Rose Professor of Law and Director, AI Legal Studies Initiative, The
University of Alabama.
††
William A. Schnader Professor of Law, University of Pennsylvania Carey School of Law. We
thank participants at the 10th Annual Empirical Contracts Workshop, and a faculty workshop
at Penn Law, and Lisa Bernstein, Jack Boeglin, Michael Bommarito, Cary Coglianese, Jean
Galbraith, Brian Feinstein, Paul Heaton, Bob Hillman, Emily Leslie, Shaun Ossei-Owusu,
Omri Ben-Shahar, and Tess Wilkinson-Ryan for comments. We are also grateful to all the
lawyers, judges, and law students who took part in the experiments described below.
†

!

Introduction ............................................................................................................. 3
I.

Gap Filling’s Empirical Gap ............................................................................ 9

II.

Some Experimental Evidence of Gap Filling .................................................. 18
A.

Ground Truth and Interpretation ............................................................... 18

B.

Three Scenarios .........................................................................................22
1.

Artist. .....................................................................................................22

2.

Contingency Fee .................................................................................... 25

3.

Bottles .................................................................................................... 27

C.

Measures and Methods ............................................................................. 28

D.

Results ........................................................................................................ 31

E.

Robustness Checks: Perturbed Contracts ................................................... 37

F.

Scaling Up: 119 Unseen Contracts ..............................................................39

G.

General Discussion of Results ................................................................... 43

III. The Practice and Perils of Generative Gap Filling ......................................... 47
A.

Generative Gap Filling Within Contract Litigation ................................... 48

B.

Choice of Model Clauses & Equilibrium Drafting Effects ......................... 54

C.

Generative Gap Filling in the Chambers ................................................... 60

D.

How Far Does This Go? ............................................................................. 61

E.

Gaps in Generative Gap Filling and the Limits of the Method ...................63

IV.

Conclusion..................................................................................................... 68

I NTRODUCTION
Contracts are bursting with meaning, imperfectly expressed.1 The
parties draft off of old templates which pull obscure boilerplate into
unexpected modern contexts,2 even as they over-engineer their deals so
that they are robust to small errors.3 They pay their lawyers richly to
include the right kind of legalese, seeking to ensure that their intended
meaning will prevail in future disputes.4 Yet for all this effort, what they
produce can still look importantly unfinished.5
When a contingency arrives that the agreement didn’t address—an
index vanishes,6 a music hall burns,7 a floating price has to be set8—
judges will sometimes fill the contractual gap.9 Modern jurists have a
Cf. Wood v. Lucy, Lady Duff-Gordon, 222 N.Y. 88, 118 N.E. 214 (1917) (arguing they are
“instinct with an obligation…”)
See, e.g., Stephen J. Choi, Mitu Gulati & Robert E. Scott, The Black Hole Problem in
Commercial Boilerplate, 67 DUKE L.J. 1 (2017) (noting how rote use of standard forms can
create interpretative problems).
See Cathy Hwang & Matthew Jennejohn, Deal Structure, 113 NW. U. L. REV. 279 (2018)
(arguing that modern contract construction relies on modular designs and other overlapping
techniques).
See generally Eric Martinez, Francis Mollica & Edward Gibson, Poor Writing, Not Specialized
Concepts, Drives Processing Difficulty in Legal Language, 224 COGNITION 105070 (2022)
(finding persistent and hard-to-read legalese in legal texts).
The classic treatments of the problem are Charles J. Goetz & Robert E. Scott, The Mitigation
Principle: Toward a General Theory of Contractual Obligation, 69 Va. L. Rev. 967 (1983)
[hereinafter Mitigation]; Charles J. Goetz & Robert E. Scott, The Limits of Expanded Choice:
An Analysis of the Interactions Between Express and Implied Contract Terms, 73 Cal. L. Rev. 261
(1985) [hereinafter Limits]; Eyal Zamir, The Inverted Hierarchy of Contract Interpretation and
Supplementation, 97 COLUM. L. REV. 1710, 1777–88 (1997); Ian Ayres & Robert Gertner,
Filling Gaps in Incomplete Contracts: An Economic Theory of Default Rules, 99 YALE L.J. 87
(1989).
Oglebay Norton Co. v. Armco, Inc., 52 Ohio St. 3d 232 (1990) (filling gap when key financial
price list was no longer published).
Taylor v. Caldwell, 3 B. & S. 826, 122 Eng. Rep. 309 (Q.B. 1863) (supplying an unstated term
that the music hall’s continued existence was an implied condition of both parties’ duties,
and discharging the contract when fire destroyed the hall before performance).
Sun Printing & Publ’g Ass’n v. Remington Paper & Power Co., 235 N.Y. 338 (1923)
See generally Alan Schwartz & Robert E. Scott, Contract Theory and the Limits of Contract Law,
113 YALE L.J. 541, 594–608 (2003) (describing the case for defaults to fill gaps and
undermining it). We say sometimes because parties will sometimes want other actors to fill
gaps—from arbitrators to seller-buyer committees to project managers. See generally Ronald
J. Gilson, Charles F. Sabel & Robert E. Scott, Contracting for Innovation: Vertical
Disintegration and Interfirm Collaboration, 109 COLUM. L. REV. 431 (2009) (joint buyer–
supplier governance committees and contractual referees); Lisa Bernstein & Brad Peterson,
Managerial Contracting: A Preliminary Study, 14 J. LEGAL ANALYSIS 176 (2022) (deal
managers).

#

name for this work. They call it construction, the supplying of terms the
text lacks, to set it apart from interpretation, the humbler business of
recovering meaning the text already holds.10 The line between them is
supposed to mark something important: the place where judges stop
reading and start deciding.11 It is also where, the worry runs, a court
unmoored from the page can write its own preferences into the deal and
call them the parties’.12
The story rests on an implicit premise that a silent text can’t and
hasn’t spoken to the parties’ intent.13 Scholars have long assumed, but
not proven, that when courts construe, they do so behind the veil.14
The gap filling literature is notable for its inconsistent terminology. We adopt the
construction/interpretation line, a tradition that stretches back nearly two centuries. SEE
FRANCIS LIEBER, LEGAL AND POLITICAL HERMENEUTICS 55–57, 62, 166 (Boston, Charles C.
Little & James Brown 1839) (construction is where interpretation “ceases to avail”), cited in
Gregory Klass, Contracts, Constitutions, and Getting the Interpretation-Construction
Distinction Right, 18 GEO. J.L. & PUB. POL’Y 13, 19 & nn.14–18 (2020). Corbin’s version
contrasts interpretation, determining meaning of symbols, and construction, determining
legal effect including when parties had not considered the problem. 3 ARTHUR L. CORBIN,
CORBIN ON CONTRACTS § 534, at 7–15 (1960). Larry Solum revived the distinction first in
constitutional theory. See, e.g., Lawrence B. Solum, The Interpretation-Construction
Distinction, 27 CONST. COMMENT. 95, 110 n.30 (2010). Greg Klass’s vision of the divide
retains the idea that gap filling is emblematically an exercise of construction because there is
no intent, but also points out that “interpretation of the parties’ intent never suffices to
identify their legal obligations.” Klass, supra, at 18–20. See also William Baude & Stephen E.
Sachs, The Law of Interpretation, 130 HARV. L. REV. 1079, 1128 (2017) (offering a slightly
different general conception). We explore the semantic and doctrinal complexities further
infra Part I.
On the related problems of construction and interpretation in public law, see Lawrence B.
Solum, Originalism and Constitutional Construction, 82 FORDHAM L. REV. 453 (2013)
(identifying an irreducible zone of construction in constitutional law).
The worry that an untethered court supplies its own preferences runs through the major
positions. See generally David Charny, Hypothetical Bargains: The Normative Structure of
Contract Interpretation, 89 MICH. L. REV. 1815 (1991); Schwartz & Scott, supra note 9; Edwin
W. Patterson, The Interpretation and Construction of Contracts, 64 COLUM. L. REV. 833 (1964).
See, e.g., Schwartz & Scott, supra note 9, at 595 (“Courts in [incomplete contracts] cases no
longer can simply engage in interpretation because, by definition, the contracts lack words to
interpret. The courts’ task thus shifts to the development of rules to resolve gap cases.”);
Joseph William Singer, Legal Realism Now, 76 CAL. L. REV. 465, 485 (1988) (reviewing LAURA
KALMAN, LEGAL REALISM AT YALE: 1927–1960 (1986)) (explaining that, according to the
realists, “Gaps in contract language are common. Where gaps exist, courts must determine
the rights of the parties with little or no guidance from the parties themselves.”).
For an argument suggesting that there is no fact-of-the-matter about parties’ intent when
filling gaps in contracts, see Robert A. Hillman, More Contract Lore, 94 TUL. L. REV. 903, 910
(2020); see also Alan Schwartz & Robert E. Scott, Obsolescence: The Intractable Production
Problem in Contract Law, 121 COLUM. L. REV. 1659, 1675 n.72 (2021) (“There is virtually no

That lack of proof is itself (unfortunately) a gap in the literature about
gap-filling. 15 If the premise is false, much of the worry about courts
unbound dissolves.
So we tested it.
We borrow a design from the machine learning literature.16 Taking
real, executed contracts, we masked—the technical term of art for
redaction—a term the parties had negotiated, typically a contingency
provision, of the sort that tells the court what to do if a stated event
occurs. We gave the masked agreement, together with a realistic
scenario that would trigger the missing clause, to three kinds of readers:
ordinary people, legally trained ones (law students and practicing
lawyers), and several large language models. Each was asked to predict
the gap we created. We knew what the parties wrote; our readers did
not. Either they recovered it or they did not.17
Before telling you what happened, we’d note that this masking
design provides a good answer to one of the core challenges to the use
of generative AI in law, and a way forward for the general problem of
grading interpretative methods. This is the second half of a project we
began in a paper called Generative Interpretation, where we argued that
evidence that courts, even those operating under the UCC’s invitation to broadly examine
context, ever conduct serious empirical investigations, and hence there is little reason to
imagine they could succeed if they did.”). For suggestions in the literature that we could
recover meaning from existing contracts, see Ayres & Gertner, supra note 5, at 116 n.122
(“existing contracts provide evidence for what the parties would have done, so mimicking
the market may be justified”). Cf. Omri Ben-Shahar & Lior Jacob Strahilevitz, Interpreting
Contracts via Surveys and Experiments, 92 N.Y.U. L. REV. 1753 (2017) (reporting the results of
three surveys about the meaning of consumer contracts).
For the canonical paper on law review’s trope of identifying and filling gaps, see Noah C.
Chauvin, Against Gap-Filling, 2024 CARDOZO L. REV. DE NOVO 1 (arguing that gap-filling for
its own sake is embarrassing).
See Jacob Devlin, Ming-Wei Chang, Kenton Lee & Kristina Toutanova, BERT: Pre-Training
of Deep Bidirectional Transformers for Language Understanding (arXiv, Working Paper No.
1810.04805, 2019), https://arxiv.org/abs/1810.04805 (introducing the masked-languagemodel pretraining objective); cf. Wilson L. Taylor, “Cloze Procedure”: A New Tool for
Measuring Readability, 30 JOURNALISM Q. 415, 416 (1953) (anticipating the design by
“mutilating [a message’s] language patterns by deleting parts” and scoring readers on their
ability to restore the deletions). Early AI work used similar techniques to train models to
parse contracts. See, e.g., Dan Hendrycks, Collin Burns, Anya Chen & Spencer Ball, CUAD:
An Expert-Annotated NLP Dataset for Legal Contract Review (arXiv, Working Paper No.
2103.06268, 2021), https://arxiv.org/abs/2103.06268.
See infra Part II (corpus, populations, masking protocol, and scoring).
$

language models could help courts parse contract text.18 That argument
drew many heated objections, one we half-conceded ourselves: we
could show a model’s reading was sensible, but not that it was correct,
because in a real dispute there is no answer key.19 One carefully-argued
response put it starkly: “no experiment can determine whether a
generative method yields correct results, because there is no accessible
source of ground truth for legal meaning.”20
Here, by redacting a term that was in fact negotiated, priced,
drafted, and signed, we can manufacture a ground truth and, for the
first time, grade interpretative predictions against it.21
In our masking experiment, humans were surprisingly good gap
fillers. Lay subjects got the hidden clause right about half the time,
twice as often as chance would predict. And legal experience honed
readers’ abilities.22 While law students barely edged out lay readers,
seasoned lawyers predicted 59% of the missing terms. Then came the
machines. Given nothing but the rest of the contract, the models were
right 88% of the time. And differently trained models, variously
prompted, kept returning the same answer.
We probed these results’ robustness by masking over 100 other
contracts of different types. Across subject matter areas, AI gap filling
was persistently accurate. We also perturbed the original contracts of
Yonathan Arbel & David A. Hoffman, Generative Interpretation, 99 N.Y.U. L. REV. 451 (2024).
See id. at 460 (“in most contract cases there is no ground truth at hand”). The objection is
pressed in James Grimmelmann, Benjamin Sobel & David Stein, Generative
Misinterpretation, 63 HARV. J. ON LEGIS. 229, 252 (2026). See also Jonathan H. Choi, Off-theShelf Large Language Models Are Unreliable Judges, J. EMPIRICAL LEGAL STUD. (forthcoming
2027), https://ssrn.com/abstract=5188865 (arguing that LLM judgments are “highly
sensitive to prompt phrasing, output processing methods, and choice of model”); Brandon
Waldon, Nathan Schneider, Ethan Wilcox, Amir Zeldes & Kevin Tobia, Large Language
Models for Legal Interpretation? Don’t Take Their Word for It, 114 GEO. L.J. (2026).
Grimmelmann et al., supra note 19, at 304 n.321.
We do not discuss gap filling in statutory interpretation, let alone specialized areas like patent
and tax law, though those would be fertile areas for research.
This finding adds to a small literature empirically measuring how legal training works. See,
e.g., Dan M. Kahan, David A. Hoffman, Danieli Evans, Neal Devins, Eugene Lucci &
Katherine Cheng, “Ideology” or “Situation Sense”? An Experimental Investigation of Motivated
Reasoning and Professional Judgment, 164 U. PA. L. REV. 349 (2016) (lawyers were less subject
to motivated reasoning reading legal texts than either lay people or law students); but cf.
Holger Spamann & Lars Klöhn, Justice Is Less Blind, and Less Legalistic, than We Thought:
Evidence from an Experiment with Real Judges, 45 J. LEGAL STUD. 255 (2016) (providing
evidence counter to the Situation Sense result).

our first experiment, showing that when we changed individual terms
in the deal, AI predictions about the masked term moved toward the
distribution of resources the modified contractual language implied.
We estimated in our sample that around two-thirds of AI success in gap
filling rested on general expectations about contract structure, and onethird came from specific contract language.
All of this carries two lessons for interpretation.
First, the finding that readers of all stripes can recover so much
from text, and so reliably, puts pressure on an empirical assumption
that has long structured contract theory. Jurists routinely assume there
is meaningful difference between what a text contains and what
surrounding materials suggest was intended.23 The dominant scholarly
view is that most of the time, courts are gap filling without good
evidence of party intent, and inject their policy preferences into broken
deals. We challenge that premise. When sophisticated readers—and
especially AI—can accurately predict specific contractual terms from
the rest of the document, the supposed boundary between text and
context becomes fuzzy.
That fuzziness is no accident. It reflects what a growing body of
evidence has revealed: the terms the parties actually negotiate are not
freestanding choices imposed on a blank slate.24 Rather, on the margin,
they follow from (and in some ways result from interactions among
terms within) the document they sit within. 25 The invisible
hypothetical bargain thus may often be statistically legible. This implies
that with the help of AI agents, jurists can predict what parties would
have said with less resort to free ranging inquiry, at least where the
parties did not deliberately disagree.26 If a term that the parties drafted
Cf. Adam Kramer, Implication in Fact as an Instance of Contractual Interpretation, 63
CAMBRIDGE L.J. 384, 385 n.11 (2004) (noting the importance of whether “pragmatic
inference” can fill gaps)
Cf. Vincent S.J. Buccola & David A. Hoffman, Precedent Terms, U. CHI. L. REV. (forthcoming
2026) (noting the role of precedent in determining commercial terms).
See Robert Anderson IV & Jeffrey Manns, The Inefficient Evolution of Merger Agreements, 85
GEO. WASH. L. REV. 57, 64–66 (2017) (showing that merger agreements can be traced into
inherited “family tr[ees]”).
See, e.g., Omri Ben-Shahar, “Agreeing to Disagree”: Filling Gaps in Deliberately Incomplete
Contracts, 2004 WIS. L. REV. 389 (discussing the problem of strategic ambiguity).

%

is recoverable from the rest of the contract at ninety-plus percent
accuracy, we argue that the same is probably true of many terms the
parties did not draft but might have, for whatever reason.
Second, we provide direct evidence that AI can help legal
decisionmakers reach the right results in a core contract interpretation
task. And by right we really mean it. The standing critique of generative
interpretation in all kinds of legal work is that its outputs cannot be
validated as the criterion of correctness is itself contested—that what
looks like accuracy may be an artifact of prompt construction, model
selection, or projection by the user.27 We cabin that worry by choosing
a task where correctness is knowable. Either the model recovers the
term the parties drafted or it does not. The frequency with which it
does so is a fact about the model, not a fact about our priors.
That generative AI is so good at this task shouldn’t necessarily
knock your socks off. After all, the models aren’t subject to cognitive
exhaustion, boredom, or motivated thinking: they are willing to parse
all the tokens in the contract we provided and pattern match them
against a vast dataset of other contracts they saw during training. But
knowing that the answers they come to are in fact correct is reassuring:
judges and lawyers should get increasingly comfortable with this
technology as a technique to debias their own over-confidence, extract
information from deals and illustrate a range of possible outcomes.28
Grimmelmann et al., supra note 19; see also James Toomey, Zombies, AI, and the “Objective”
Theory of Contracts, HARV. J.L. & TECH. (forthcoming 2026) (arguing that agentic models
cannot intend to conclude contracts, no more than a cloud that happens to spell out “if you
build it he will come” intends to communicate a message about a baseball field in Iowa);
Zachary Catanzaro, The Dead Law Theory: The Perils of Simulated Interpretation, FLA. L. REV.
(forthcoming 2027), https://ssrn.com/abstract=6164388 (contending that computational
legal interpretation “fails because it is a category error,” since LLMs “manipulat[e] symbols
without accessing what those symbols mean”).
On the mounting judicial imperative to integrate AI into courthouses and chambers, see
Yonathan A. Arbel, Judicial Economy in the Age of AI, 96 U. COLO. L. REV. 549 (2025). On
explicit AI use in case production, see Snell v. United Specialty Ins. Co., 102 F.4th 1208, 1221
(11th Cir. 2024) (Newsom, J., concurring) (urging interpreters to consider “whether and how
AI-powered large language models . . . might—might—inform the interpretive analysis”);
United States v. Deleon, 116 F.4th 1260, 1277 (11th Cir. 2024) (Newsom, J., concurring)
(concluding that “LLMs have something to contribute to the ordinary-meaning endeavor”).
For a comparative perspective, see Juzgado Primero Laboral del Circuito de Cartagena [First
Labor Circuit Court of Cartagena], Sentencia No. 032, Radicación No. 13001-41-05-0042022-00459-01 ( Jan. 30, 2023) (Colom.) (incorporating ChatGPT responses into a tutela

We argue that evidence produced by generative AI drawn from the text
itself can be helpful, arguably no less than extrinsic evidence, in
deciding how to fill gaps.
But how can jurists use AI without turning contract interpretation
disputes into empty contests over whose model has more parameters
and whose prompt won the race?29 Our answer borrows from existing
contract doctrine. Parties already choose their governing law and their
forum; they can also stipulate a choice-of-model clause—naming, in the
agreement, the system whose reading will be given weight. 30 Such a
clause does two things at once. It folds the technology into the ordinary
machinery of interpretation, and it answers the sharpest objection to
it—that its output bends to whoever picks the model after the dispute
arises—by fixing the model before the dispute exists.
The Article proceeds in three parts. Part I excavates the gap in the
gap filling literature by unpacking the empirical assumptions
underlying the divide between interpretation and construction. Part II
then sets up our masking experiment and relays its results. Part III
explains how jurists might go about using this method of filling gaps
day-to-day, introduces choice of model clauses and describes how their
use will change contract practice, and considers several objections.
I.

G AP F I LLI NG ’ S E MPI RI CAL G AP

Contract law’s gap-filling literature is extraordinary in its scope,
with dozens of major works in the last two generations.31 We necessarily
will draw with a very broad brush in describing it. Even what counts as
ruling on health coverage); Jaswinder Singh v. State of Punjab, CRM-M-22496-2022,
2023:PHHC:044541 (Punjab & Haryana High Ct. Mar. 27, 2023) (India) (Chitkara, J.) (using
ChatGPT on bail practice in a section titled “Post-Reasoning”). For survey evidence, see
Anika Jaitley, Daniel W. Linna Jr., Hon. Xavier Rodriguez, V.S. Subrahmanian & Siyu Tao,
Artificial Intelligence in Federal Courts: A Random-Sample Survey of Judges, 27 SEDONA CONF.
J. (forthcoming 2026) (reporting that 61.6% of federal judges responding to a December 2025
survey had used at least one AI tool in their judicial work); Gregg Wirth, 2026 AI in
Professional Services Report: AI Adoption Has Hit Critical Mass, Thomson Reuters Inst. (Feb.
9, 2026), https://www.thomsonreuters.com/en-us/posts/technology/ai-in-professionalservices-report-2026/ (reporting, from a survey of more than 1,500 professionals across 27
countries, that organization-wide AI use nearly doubled to 40% in 2026 from 22% in 2025).
On output sensitivity to prompt phrasing and choice of model, see Choi, supra note 19.
We floated the possibility in Generative Interpretation, supra note 18, at 455 (parties “may start
to include them in their choice-of-law repertoire”).
For a sampling see supra notes 5,9,11,12, and 14.
&

a “gap” and the work of “construction” in filling it is endlessly
contested. 32 Consider that the parties may have not spoken about a
contingency because they know the law will give them a default
term 33 —that’s not a “gap”, that’s efficient drafting!—or Richard
Posner’s observation that ambiguity can always be redefined as a gap in
expression.34
Beneath the definitional churn, the doctrine’s ambition has been
stable for a century. When a contract doesn’t speak with clarity to a
particular problem, courts say they are trying to give the parties the
deal they (hypothetically) intended. 35 In determining what that deal
would have looked like, decisionmakers often resort to a welter of
extrinsic evidence, text, and policy, trying to pick rules that maximize
ends varying from transaction cost minimization to fairness.36 The less
meaning the text provides, the more rests on context and non-text.
Scholarship has organized itself for generations around doubt that
courts can do this well. Beneath nearly every major position sits a
shared empirical assumption: terms stated explicitly about a dispute
express meaning, and in their absence, the rest of the document and its
context are poor inferential sources about what they say.37 Silence, on
this view, is informationally thin. The worry runs that, unmoored from
Ayres and Gertner argue that identification of gaps is itself an interpretative question that
sounds in the legal system’s altering rules, the rules that define how and when parties can
contract around a gap. See Ayres & Gertner, supra note 5, at 119–21. See also
Ian Ayres, Regulating Opt-Out: An Economic Theory of Altering Rules, 121 YALE L.J. 2032, 2044
(2012).
See Randy E. Barnett, The Sound of Silence: Default Rules and Contractual Consent, 78 VA. L.
REV. 821, 865–66 (1992) (arguing that “Silence in the face of default rules can constitute an
‘indirect’ consent to courts using these default rules to supply terms when a gap exists in the
parties’ expression of consent.”).
Richard A. Posner, The Law and Economics of Contract Interpretation, 83 TEX. L. REV. 1581,
1589 (2005).
See, e.g., Goetz & Scott, supra note 5, at 971 (“Ideally, the preformulated rules supplied by the
state should mimic the agreements contracting parties would reach were they costlessly to
bargain out each detail of the transaction. Using this benchmark raises two separable issues:
First, what arrangements would most bargainers prefer? And, second, what atypical
arrangements should be supported as benign alternatives?”).
See, e.g., Robert A. Hillman, Health Crises and the Limited Role of Contract Law, 85 LAW &
CONTEMP. PROBS. 19, 26–27 (2022) (canvassing approaches).
To be sure there are dissenters. See, e.g., Robin Bradley Kar & Margaret Jane Radin, PseudoContract and Shared Meaning Analysis, 132 HARV. L. REV. 1135, 1148–55 (2019) (adapting
Grice’s cooperative principle); see also Kramer, supra note 23.

the page, courts saying they are recovering meaning are in fact fishing
their own wishes. David Charny articulated a version of this judgesgone-wild claim three decades ago. He argued that interpretive
conventions like the hypothetical bargain should be evaluated for what
they accomplish as conventions, not for their descriptive fidelity to
party intent—fidelity being a lost cause.38
This doubt produced a shared research program, which offered
strategies for answering the question of what a court ought to do when
a contract doesn’t speak directly, without relying on intent recovered
from context clues. The default-rules project reads majoritarian terms
off what typical parties would want precisely because what these parties
wanted is thought to be unrecoverable case by case. 39 Its most
celebrated refinement—information-forcing penalty defaults that push
the better-informed party to disclose—spurns intent.40
Other scholars would supply trade usage, course of dealing, and
good-faith standards drawn from the surrounding commercial context:
importing content from the relationship because the writing is assumed
to carry too little.41 And a vigorous and skeptical formalist camp argues
that courts can’t get gap filling right at any tolerable error rate, and so

Charny, supra note 12, at 1820–35.
See, e.g., Goetz & Scott, supra note 5, at 971.
Ayres & Gertner, Filling Gaps, supra note 5, at 97–100; see also Ian Ayres & Robert Gertner,
Strategic Contractual Inefficiency and the Optimal Choice of Legal Rules, 101 YALE L.J. 729
(1992) (refining penalty-default theory in response to critics); Eric A. Posner, There Are No
Penalty Default Rules in Contract Law, 33 FLA. ST. U. L. REV. 563 (2006) (arguing that the
canonical illustrations are not in fact penalty defaults); Ian Ayres, Ya-Huh: There Are and
Should Be Penalty Defaults, 33 FLA. ST. U. L. REV. 589 (2006) (defending both the existence
and the desirability of penalty defaults). For a criticism of penalty default theory based on its
pernicious expressive content, see Tess Wilkinson-Ryan, David A. Hoffman & Emily
Campbell, Lessons in Contract, GEO. L.J. (forthcoming 2026) (manuscript at 52–53).
See, e.g., Robert A. Hillman, More Contract Lore, 94 TUL. L. REV. 903, 912 (2020) (arguing
that law reforms should avoid talking about party intent as a goal of gap filling but instead
focus more frankly on social goals).

!!

the law should opt for value-or policy-driven fillers,42 or refuse to fill
at all.43
What unites these positions is less obvious than what divides them.
They do differ, sometimes sharply, over what counts as a gap and over
the means of filling one. But they also share a premise about where the
document gives out: so long as the contract supplies meaning, that
meaning has gravity. When the document stops speaking, it is the
judge’s turn to opine. The judge is then licensed to import materials
from outside the four corners: e.g., inferences about what these parties
or typical ones would want given their prior dealings or negotiations,
penalties designed to make the next parties speak, the customs of the
trade, the court’s own sense of sound policy.44
For our specific purposes, and without making a more general
analytical claim, we will treat interpretation as recovering content the
document supplies, and construction as importing it from the outside.45
See Robert E. Scott, The Case for Formalism in Relational Contract, 94 NW. U. L. REV. 847, 848
(2000) (arguing that courts should “accept the limits imposed by legal formalism and
interpret the facially unambiguous . . . terms of disputed contracts literally”); Omri BenShahar, The Tentative Case Against Flexibility in Commercial Law, 66 U. CHI. L. REV. 781, 806–
20 (1999) (arguing that because parties draft anti-erosion provisions in anticipation of flexible
judicial gap-filling, the apparent welfare gains from such flexibility are largely illusory); see
generally Robert E. Scott, A Theory of Self-Enforcing Indefinite Agreements, 103 COLUM. L. REV.
1641 (2003) (documenting the persistence of the indefiniteness doctrine and arguing that
courts should decline to complete deliberately incomplete agreements); Varney v. Ditmars,
217 N.Y. 223, 111 N.E. 822 (1916) (holding a promise of a “fair share” of profits too indefinite
to enforce, and refusing to supply the missing term); Alan Schwartz & Robert E. Scott, The
Common Law of Contract and the Default Rule Project, 102 VA. L. REV. 1523, 1556–67 (2016);
Lisa Bernstein, Merchant Law in a Merchant Court: Rethinking the Code’s Search for Immanent
Business Norms, 144 U. PA. L. REV. 1765, 1796–1802 (1996) (arguing that some merchants
prefer strict, formalist enforcement of the written contract once a dispute reaches
adjudication, contrary to the UCC’s incorporation of trade usage, course of dealing, and
good faith); see also Lisa Bernstein, The Questionable Empirical Basis of Article 2’s Incorporation
Strategy: A Preliminary Study, 66 U. CHI. L. REV. 710, 751–60 (1999) (arguing that gap-filling
based on “trade usage” rests on a false empirical premise about trade norms’ authenticity).
In re El Paso Pipeline Partners, L.P. Derivative Litig., No. CIV.A. 7141-VCL, 2014 WL
2768782, at *17–18 (Del. Ch. June 12, 2014) (cleaned up) (“Not all gaps should be filled.”).
See, e.g., CHARLES FRIED, CONTRACT AS PROMISE: A THEORY OF CONTRACTUAL OBLIGATION
69–73 (1981) (gaps cannot be filled by the contract).
Patterson, supra note 12, at 835 (“Construction, which may . . . be usefully distinguished
from interpretation, is a process by which legal consequences are made to follow from the
terms of the contract and its more or less immediate context, and from a legal policy or
policies that are applicable to the situation.”); Lawrence A. Cunningham, Hermeneutics and

This is not a universally shared divide, and some parts of the gap filling
literature avoid both terms,46 but our arguments here do not turn on the
definitional contest. Whatever the vocabulary, on some margin all
participants assume that at some point at which the parties’ recoverable
intent has run out. Our question is empirical rather than taxonomic—
how much reliable information does the text contain about missing
pieces of its instruction set?
Each of these programs, whatever divides them, proceeds as if the
answer were “not a lot.” But what if the assumption is wrong?
Courts have always suspected it might be. Consider Wood v. Lucy,
often taught as the paradigmatic gap-filling case. 47 Judge Cardozo
implied a promise by Wood to use reasonable efforts in marketing DuffGordon’s wares, making enforceable an otherwise illusory contract.
Carefully read, Cardozo’s inference rests in part on his own common
sense, and in part on the text itself. In deciding that Wood owed some
effort, the court’s analysis extrapolates from the deal’s textual
exclusivity, its profit-share compensation structure, and the collateral
undertakings around it. 48 Only after reciting these facts about the
contract did Cardozo conclude it was “instinct with an obligation,
imperfectly expressed.”49

Contract Default Rules: An Essay on Lieber and Corbin, 16 CARDOZO L. REV. 2225 (1995)
(recovering the older hermeneutic tradition for the contract default-rule debate).
See, e.g., Klass, supra note 10, at 14–15 (arguing that construction is ubiquitous in fixing
contractual obligations); cf. Frederick Schauer, Constructing Interpretation, 101 B.U. L. REV.
103, 109 (2021) (arguing that “interpretation itself is often constructed”). For parallel
debates in constitutional theory, see generally Baude & Sachs, supra note 10 (arguing that legal
interpretation is governed neither by linguistics alone nor by policy choice, but by a “law of
interpretation”).
Wood v. Lucy, Lady Duff-Gordon, 222 N.Y. 88, 118 N.E. 214 (1917); see also CHARLES L.
KNAPP ET AL., PROBLEMS IN CONTRACT LAW: CASES AND MATERIALS 486 (10th ed. 2023)
(teaching Wood as a principal case on implied terms).
See Victor P. Goldberg, Reading Wood v. Lucy, Lady Duff-Gordon with Help from the Kewpie
Dolls, in FRAMING CONTRACT LAW: AN ECONOMIC PERSPECTIVE 43, 47–63 (2006)
(reconstructing the deal’s commercial context through Wood’s earlier Kewpie-doll license,
which contained an express best-efforts clause, and questioning whether Cardozo’s
implication squares with it).
Wood, 222 N.Y. at 91.
!#

That inferential doctrine is older and broader than Wood. 50 The
English business-efficacy line dates to The Moorcock, in which the court
inferred an undertaking of reasonable care to ensure a safe berth from
the structure of a wharf-rental agreement on the ground that the deal
would be unworkable without it. The companion officious-bystander
test from Shirlaw v. Southern Foundries is even more revealing. It
implies a term whenever a hypothetical bystander proposing the term
during the parties’ negotiations would have been testily suppressed
with “‘Oh, of course!’”51
A short century later, UK courts still debate these issues. Lord
Hoffmann sought to merge reading and implication, insisting that there
is “only one question: is that what the instrument, read as a whole
against the relevant background, would reasonably be understood to
mean?” 52 However the U.K. Supreme Court resisted this synthesis,
seeking to sharpen the line between interpretation and implication,
holding (in the American tradition) that “[interpreting] the words used
and implying additional words are different processes governed by
different rules.”53
American doctrine hews to the inferential line. The Restatement
(Second) of Contracts § 204 supplies a term “reasonable in the
circumstances” whenever an agreement omits one essential to
determining the parties’ rights, and the implied covenant of good faith
and the various implied-warranty doctrines do related work in
particular substantive areas.54 The underlying premise is that courts
can most legitimately insert a term when the rest of the document, plus
See, e.g., Larry A. DiMatteo, Cardozo, Anti-Formalism, and the Fiction of Noninterventionism,
28 PACE L. REV. 315 (2008) (reading Wood as text-derived rather than externally imposed);
cf. Goldberg, supra note 48, at 47–63 (noting problems with the imputation given surrounding
context).
The Moorcock (1889) 14 P.D. 64, 68 (Eng. CA); Shirlaw v. S. Foundries (1926) Ltd. [1939] 2
K.B. 206, 227 (Eng. CA), aff’d [1940] A.C. 701 (HL).
Att’y Gen. of Belize v. Belize Telecom Ltd. [2009] UKPC 10, [21], [2009] 1 W.L.R. 1988 (appeal
taken from Belize).
Marks & Spencer plc v. BNP Paribas Sec. Servs. Tr. Co. ( Jersey) Ltd. [2015] UKSC 72, [26],
[2016] AC 742 (appeal taken from Eng.).
Restatement (Second) of Contracts § 204 (Am. L. Inst. 1981); see also id. § 205; U.C.C. § 2314 (Am. L. Inst. & Unif. L. Comm’n 2022). Comment d to § 204 is two-staged: courts look
first to the document’s meaning and to the probability that a term would have been used, and
resort to community standards of fairness only “where there is in fact no agreement.” Id. §
204 cmt. d.

the type and shape of the deal, is informative enough to support a rough
outline of its content.
That gap-filling proceeds at least to some degree from text does not
make it obvious how much we can reliably extract from the contract
itself when it is silent on the key issue before the court. 55 The best
modern account of contract construction argues that it is continuous
with interpretation but does not say how reliable text-derivation can be
without express terms.56 Or to put it differently, there’s clearly a sense
in the literature that at some point the ice is too thin to walk out on, and
what’s needed is a normative theory that justifies and buttresses
courts’ intervention. But the actual method we use to make deductions
from the text is no better developed than it was in Cardozo’s day.
Doctrine’s implied-term tests—business efficacy, the officious
bystander’s testy “of course”—are, in effect, unvalidated proxies for
how much of an unwritten term the rest of the deal actually determines.
Whether silence really is informationally thin obviously also
depends on why the parties went silent. For decades, scholars have
noted that silence in commercial drafting captures at least three things,
often at once.57
It can be strategic disagreement: the parties knew their interests
diverged on a particular contingency and did not want to spend the
negotiating capital to resolve it before signing. Material-adverse-change
clauses and many best-efforts and reasonable-efforts terms are
paradigmatic—deliberately vague because precise terms could not be
agreed upon.58 Here, courts are in a pickle. Filling a gap may pick sides
The rare field evidence points the same way. See Yair Listokin, The Meaning of Contractual
Silence: A Field Experiment, 2 J. LEGAL ANALYSIS 397, 406–10 (2010) (providing estimates that
buyers price in contractual silence similar to default UCC warranties).
See generally Klass, supra note 10; cf. Ronald J. Gilson, Charles F. Sabel & Robert E. Scott,
Text and Context: Contract Interpretation as Contract Design, 100 CORNELL L. REV. 23, 30–45
(2014).
The core insight is from E. Allan Farnsworth, Disputes over Omission in Contracts, 68 COLUM.
L. REV. 860 (1968).
On strategic vagueness as a response to interest divergence, see Albert H. Choi & George G.
Triantis, Strategic Vagueness in Contract Design: The Case of Corporate Acquisitions, 119 YALE
L.J. 848 (2010) (showing that parties to corporate acquisitions deliberately use vague terms
to defer rather than resolve disagreements). On strategic non-negotiation as a separate
phenomenon driven by relative cost, see Robert E. Scott & George G. Triantis, Anticipating

!$

in a battle the parties chose not to fight, and consequently didn’t bear
the price of settling. 59 The same can be said about strategic punting,
where the parties envision some process, other than a judge or
arbitrator, that would fill their gap later.60
It can be strategic failure to reduce an agreement to writing: the parties
recognized the issue, shared an understanding about it, but judged the
expected cost of writing that deal down greater than the expected loss
from leaving it alone.61 Perhaps they like the default, or as in Shirlaw
take it for granted that the term is covered, or maybe they trust the
prudence of a future judge to ascertain it just right.62 Here, providing a
term certainly serves the parties’ own ends, because it helps them to
come to agreements by providing terms on the back end that front end
bargaining would have arrived at. Relatedly, on an influential account,
true gaps should be rare: broad standards are cheap to write, so the
paradigmatic “gap” is not an absence at all but a deliberate delegation
of specification to courts at the back end.63
And it can be something more like negligence: one or both sides
simply did not think about the contingency at all. These include the
wild unknowns—think of the Ever Given container ship getting
wedged in the Suez Canal—and the more pedestrian failure to
anticipate a storm.
Litigation in Contract Design, 115 YALE L.J. 814 (2006) (modeling drafting precision as an
investment with diminishing returns). On silence as a way to avoid negative signals, see
Bernstein, Merchant Law, supra note 42, at 1789–90 (“Transactors may also fail to include
written provisions dealing with a particular contingency because each may fear that the other
will interpret a suggestion that they do so as a signal that the transactor proposing the
provisions is unusually litigious or likely to resist flexible adjustment of the relationship if
circumstances change.”).
See Ben-Shahar, supra note 26, at 390–91 (proposal), 400–01 (rejecting definitive gap fillers
for deliberately incomplete contracts and proposing instead a pro-defendant default).
Bernstein & Peterson, supra note 9, at 185 n.31.
See Steven Shavell, On the Writing and the Interpretation of Contracts, 22 J.L. ECON. & ORG.
289 (2006) (modeling parties’ choice to leave terms unspecified when the cost of writing
them exceeds the expected benefit); Gillian K. Hadfield, Judicial Competence and the
Interpretation of Incomplete Contracts, 23 J. LEGAL STUD. 159 (1994) (analyzing when parties
rationally leave contracts incomplete in anticipation of judicial completion).
See Schwartz & Scott, supra note 42, at 1578 (“Parties also leave gaps when they accept the
legal default.”).
Robert E. Scott & George G. Triantis, Incomplete Contracts and the Theory of Contract Design,
56 CASE W. RES. L. REV. 187, 190, 197 (2005) (precise terms increase drafting costs, and vague
standards shift those to enforcement).

These three silences are only partially analytically distinct.
Negligence can always be reframed as the intentional choice not to take
a precaution: a strategically uninformed silence is hard to tell from a
strategically purposeful one without further information. But the claim
that silence is uninformative depends on treating the three kinds the
same, and each in fact implies something different. A silence after
disagreement tells the reader that the parties’ interests diverged on the
issue and that no resolution was reached. On this we agree with the
relational tradition: where silence records a fight the parties declined to
finish, there is no convergent preference to recover, and if judicial gap
filling is even appropriate, it should proceed without much regard to
what the parties would have or did think.64
But in the other two modes the document remains evidence—of
the parties’ priorities, their types, the deal’s structure, of what was
treated as worth bargaining over and what was not. A deliberate nondrafting tells the reader which contingencies the parties thought worth
writing down; even an inadvertent one tells us what kind of parties
these were and what kind of contract they were drafting. The question
is how much evidence the document can provide.
The literature’s central disputes—over the right filler, over rules
versus context, over what to call the judicial act—thus cash out to an
empirical claim about how much a silent document still says. Because
we lack an agreed method to extract that information reliably, jurists
have been driven to normative justifications for gap filling, which some
come to grips with better than others.65
What’s needed first is measurement. If silence is as thin as the
literature assumes, defaults and policy must carry nearly all the weight,
and the worries about improvising judges have full force. If it is not—if
the document reliably determines much of what the parties left
unwritten—then the domain where defaults, penalties, and policy must
operate is real but smaller than supposed, and the same measurement
that shrinks it can help sort the silences where intent survives from
those where it is truly missing. It’s to that project we now turn.
See, e.g., Ian R. Macneil, Contracts: Adjustment of Long-Term Economic Relations Under
Classical, Neoclassical, and Relational Contract Law, 72 NW. U. L. REV. 854 (1978).
For the paradigmatic normative account, see Zamir, supra note 5.

!%

II.

S OME E XPERIMENTAL E VI DENCE OF
G AP F I LLI NG

Our goal now is to test the hypothesis that contractual texts are
richer sources of contractual meaning for missing terms than the
literature has given us reason to expect. We start by introducing why
we need a technique like masking contract terms.
A. Ground Truth and Interpretation
When scholars ask whether textualism produces more reliable
interpretations than contextualism, 66 whether judges outperform
juries, 67 or whether sophisticated commercial arbitrators fare better
than generalists,68 they are challenged by an absence of an agreed-to
way of judging accuracy. And that’s true even though contract law—
unlike, say, statutory or constitutional law—generally agrees about
interpretation’s purpose. It is to figure out what the parties meant to
say at the time of the bargain and give it to them.69
Sometimes interpretative judgments are empirical: what do “most
people” think a sandwich is, and does a taco fit into that definition?70
But interpretative questions can also appear more intractable: you can’t
answer the question of what the drafting parties of a legal document

See, e.g., Jeffrey W. Stempel & Erik S. Knutsen, Rejecting Word Worship: An Integrative
Approach to Judicial Construction of Insurance Policies, 90 U. CIN. L. REV. 561, 600–01 (2021)
(defending contextualism and arguing that textualism is malleable); Kevin Tobia, Testing
Ordinary Meaning, 134 HARV. L. REV. 726 (2020) (experimentally finding that dictionaries,
corpus linguistics, and lay intuitions about “ordinary meaning” diverge, thus undercutting
claims that textualist tools yield more determinate readings).
Cf. Larry Heuer & Steven Penrod, Trial Complexity: A Field Investigation of Its Meaning and
Its Effects, 18 LAW & HUM. BEHAV. 29 (1994) (field study finding complexity did not
significantly increase judge–jury disagreement).
See, e.g., Theodore Eisenberg & Elizabeth Hill, Arbitration and Litigation of Employment
Claims: An Empirical Comparison, 58 DISP. RESOL. J. 44 (2003) (head-to-head arbitration-vscourt study finding no statistically significant difference in win rates or award size).
See Restatement (Second) of Contracts § 201 (Am. L. Inst. 1981) (directing that words be
interpreted in accordance with the meaning the parties attached to them).
For an insightful post on this problem, see Ilya Somin, Indiana Court Rules Burritos and Tacos
Qualify as Sandwiches, VOLOKH CONSPIRACY (May 19, 2024, 2∶28 PM),
https://reason.com/volokh/2024/05/19/indiana-court-rules-burritos-and-tacos-qualify-assandwiches/.

intended by surveying other people at a different time and place.71 And
if you ask them directly, their answers will be motivated.72
Empirical legal interpretation scholarship has in recent years made
progress where progress can be made. Existing solutions are generally
built around the wisdom of crowds: they ask how a particular
interpretative tool’s answer stacks up against the modal survey answers
of a large group of people. 73 Alternatively, some scholars compare
particular empirical findings against benchmarks like judges’ written
opinions,74 or the meaning of words derived from proximity clues in
large corpora of language.75 But again if the goal is to ascertain what the
parties themselves meant to say, it’s not obvious that these methods can
provide settled answers.
This uncertainty now infects questions about how to fit generative
interpretation into jurisprudence. Critics argue that LLM outputs lack
integrity because we can’t know if they are (in fact) correct.76 And, they
On the general point, recent work includes Mark Greenberg, What Makes a Method of Legal
Interpretation Correct?, 130 HARV. L. REV. F. 105 (2017) (arguing that absent consensus on a
criterion validating an interpretive theory, “it is indeterminate which theory of interpretation
is correct”); Ward Farnsworth, Dustin F. Guzior & Anup Malani, Ambiguity About
Ambiguity: An Empirical Inquiry into Legal Interpretation, 2 J. LEGAL ANALYSIS 257 (2010)
(showing even the threshold judgment of whether text is “ambiguous” tracks readers’ policy
preferences).
See generally Lawrence Solan, Terri Rosenblatt & Daniel Osherson, False Consensus Bias in
Contract Interpretation, 108 COLUM. L. REV. 1268 (2008) (noting this problem).
Other empirical legal scholarship has approached the ground-truth problem through
different routes. Many use surveys. See, e.g., Solan et al., id. (using surveys); Ben-Shahar &
Strahilevitz, supra note 14; Richard Craswell, Contract Law, Default Rules, and the Philosophy
of Promising, 88 MICH. L. REV. 489 (1989) (noting the utility of “a survey to determine what
people usually mean . . . when they make those noises in particular contexts”).
Arbel & Hoffman, supra note 18; Christoph Engel & Richard H. McAdams, Asking GPT for
the Ordinary Meaning of Statutory Terms, 2024 U. ILL. J.L. TECH. & POL’Y 235 (comparing
LLM aggregate predictions against those survey distributions). Kruse simulates the surveys
themselves with demographically profiled AI agents. Johannes Kruse, The Ordinary Meaning
Bot: Simulating Human Surveys with LLMs (Max Planck Inst. for Rsch. on Collective Goods,
Discussion Paper No. 2025/12, 2025), https://ssrn.com/abstract=5378203; Yonathan A.
Arbel, The Generative Reasonable Person (Feb. 17, 2026) (unpublished manuscript),
https://arxiv.org/abs/2508.02766. Each takes the relevant population’s beliefs as ground
truth.
Stephen C. Mouritsen, Contract Interpretation with Corpus Linguistics, 94 WASH. L. REV. 1337
(2019); Thomas R. Lee & Jesse Egbert, Artificial Meaning?, 77 FLA. L. REV. 2235, 2281 (2025)
(arguing that corpus linguistics is capable of contextual sensitivity and generative
interpretation is “rooted in (artificial) intuition”).
See generally supra note 27.

!&

contend, because LLM outputs can’t be proven to be correct, their
mysterious processes are particularly vulnerable to critique.
We borrow a solution to these foundational problems from the
machine learning literature which faced a version of them in reverse.
Training large language models requires “labeled” examples,
essentially, a set of questions and correct answers. 77 For the scales
necessary to train a modern language model, curating a large enough
dataset of labeled examples presents an impossible challenge, even for
tech behemoths the size of Google. The solution is a technique called
“masking.”78
The concept is simple to grasp. Take a statement, hide a portion of
it, and then ask the model to predict what the hidden part says.79 This
way, the researcher has access to the truth, which is the content of the
masked term. During training, the model tries a prediction and receives
feedback on whether it was correct. Run over billions of sentences, the
model increasingly learns prediction by finding patterns in the
sentences fed to it: from surface syntax to deeper level representations
of meaning.80 Calibration ends with a successful prediction engine: a
model able to pass the bar exam, find new answers to long-standing
math problems, and even provide relationship advice.81
See David Lehr & Paul Ohm, Playing with the Data: What Legal Scholars Should Learn About
Machine Learning, 51 U.C. Davis L. Rev. 653, 673 (2017) (explaining that supervised
algorithms “are given a labeled outcome variable . . . representing the true values to be
predicted on the basis of input data”); IAN GOODFELLOW, YOSHUA BENGIO & AARON
COURVILLE, DEEP LEARNING 103 (2016) (defining supervised learning algorithms as those
that “experience a dataset containing features, but each example is also associated with a
label or target”).
See Devlin et al., supra note 16, at 4174 (“In order to train a deep bidirectional representation,
we simply mask some percentage of the input tokens at random, and then predict those
masked tokens. We refer to this procedure as a ‘masked LM’”).
Id. at 4171 (“The masked language model randomly masks some of the tokens from the input,
and the objective is to predict the original vocabulary id of the masked word based only on
its context.”).
On scale, see id. at 4175 (pre-training on “the BooksCorpus (800M words) . . . and English
Wikipedia (2,500M words)”).
On the bar exam: Daniel Martin Katz, Michael James Bommarito, Shang Gao & Pablo
Arredondo, GPT-4 Passes the Bar Exam, 382 PHIL. TRANS. R. SOC. A 20230254 (2024) (GPT4 achieved a Uniform Bar Exam score of 297, passing in every U.S. jurisdiction); but see Eric
Martínez, Re-evaluating GPT-4’s Bar Exam Performance, 33 Artificial Intelligence & L. 581
(2025) (disputing the “90th percentile” gloss). On novel mathematics, see Nat Sothanaphan,

We apply the same logic to contract interpretation.82 We take real,
negotiated contracts and redact a consequential provision. We then ask
interpreters to recover the hidden language from the rest of the
document. Because we know what the parties wrote, we have a groundtruth benchmark against which to score answers.
Our task most directly tracks the case in which the parties reached
agreement on a term, priced it, and failed to commit it to writing. It
approximates the case where the parties would predictably agree on the
meaning of the term, had they just focused on it. It does not say much
about gap filling where the parties did not want the term filled by a court,
disagreed or would have disagreed on a term, or the case where the
inquiry is, by its own terms, frankly normative. It’s our sense that the
first two categories represent most live contract disputes, but that’s a
guess which could be empirically tested. We return to the boundaries
of the claim in Part III.

Resolution of Erdős Problem #728: A Writeup of Aristotle’s Lean Proof, (arXiv, Working
Paper No. 2601.07421, 2026), https://arxiv.org/abs/2601.07421 (formalizing the solution to
an open question that was “fully resolved autonomously by an AI system”); Alexander
Novikov et al., AlphaEvolve: A Coding Agent for Scientific and Algorithmic Discovery (arXiv,
Working Paper No. 2506.13131, 2025), https://arxiv.org/abs/2506.13131 (reporting an
algorithm for multiplying 4×4 complex-valued matrices in forty-eight scalar multiplications,
the first improvement over Strassen’s 1969 algorithm in that setting). On relationship advice:
go touch grass, and don’t do this.
To our knowledge, ours is the first paper to use masking to evaluate the accuracy of legal
interpretation. The nearest legal cousin uses a cloze-style task to probe models’ knowledge
of case law rather than their interpretation of agreements. Cf. Lucia Zheng et al., When Does
Pretraining Help? Assessing Self-Supervised Learning for Law and the CaseHOLD Dataset
(arXiv, Working Paper No. 2104.08671, 2021), https://arxiv.org/abs/2104.08671 (evaluating
language models on a multiple-choice task that masks the holding of a cited case and asks the
model to recover it). Recent work has developed expert-designed benchmarks to disaggregate
and measure different forms of legal reasoning performed by large language models. See Neel
Guha, Julian Nyarko, Daniel E. Ho, Christopher Ré, Adam Chilton, Aditya Narayana, Alex
Chohlas-Wood, Austin Peters, Brandon Waldon, Daniel N. Rockmore, Diego Zambrano,
Dmitry Talisman, Enam Hoque, Faiz Surani, Frank Fagan, Galit Sarfaty, Gregory M.
Dickinson, Haggai Porat, Jason Hegland, Jessica Wu, Joe Nudell, Joel Niklaus, John Nay,
Jonathan H. Choi, Kevin Tobia, Margaret Hagan, Megan Ma, Michael Livermore, Nikon
Rasumov-Rahe, Nils Holzenberger, Noam Kolt, Peter Henderson, Sean Rehaag, Sharad
Goel, Shang Gao, Spencer Williams, Sunny Gandhi, Tom Zur, Varun Iyer & Zehua Li,
LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large
Language Models, 36 NEURIPS DATASETS & BENCHMARKS TRACK (2023). Our method
complements that project by identifying a class of legal questions for which the underlying
document itself supplies a verifiable answer key.
'!

And by way of further warning, there’s an informational difference
between predicting the content of a term that the parties didn’t write,
and one that they did (but which has been masked by the
experimenter). Perhaps when parties write a term there is more data in
the contract outside of that term about its existence—the easy case is
cross-references or allusions in the text itself. Now that said, in none of
the cases we described above had the parties in fact written a term but
then (in effect) lost it, although such cases exist.83 And when those lostexhibit cases are resolved, the strategies courts use look remarkably like
ordinary gap filling. 84 We’ll return to the external validity of our
approach below in our discussion of the findings.
B. Three Scenarios
We implemented this masking method first across three scenarios
drawn from real contracts, each presenting a genuine interpretive
dispute with real stakes. We sourced the contracts from PACER, which
is paywalled, difficult to scrape, and consequently almost certainly
missing from the training corpora of frontier language models. 85 We
drafted the contractual disputes ourselves rather than rely on the
parties’ briefs.
1.

Artist.

The Artist scenario is based on an engagement contract
between a talent agency, J. Noah, and the Chicago rapper known as
Polo G.86 The agreement was signed ahead of Polo G’s Europe tour and
promised him a $40,000 artist fee plus various travel expenses. The
agreement provided that the agency and local promoter would
organize, advertise, and support the local performance, including
venue and production arrangements and make efforts to accommodate
See, e.g., J.N.A. Realty Corp. v. Cross Bay Chelsea, Inc., 42 N.Y.2d 392, 396 (1977) (noting
that one party “claimed that they were not aware of the time limitation because they had
never received a copy of paragraph 58 of the rider.”).
Id. (engaging in a form of gap filling by concluding that applying the plain language of the
lease renewal would be unreasonable).
See, e.g., Adam R. Pah, David L. Schwartz, Sarath Sanga, Zachary D. Clopton, Peter DiCola,
Rachel Davis Mersey, Charlotte S. Alexander, Kristian J. Hammond7, Luís A. Nunes
Amaral, How to Build a More Open Justice System, 369 SCIENCE 134, 134 (2020) (bemoaning
the PACER paywall.)
Ex. A, Bartlett v. J. Noah B.V., No. 1∶23-cv-10345-JMF (S.D.N.Y. Nov. 15, 2024), ECF No.
46-1.

the artist’s requirements, from food and beverage to other performance
logistics. The artist, in turn, promised to perform and to participate in
promotional activities ahead of the show, including “1 x Tweet from
Artist announcing Show.”87 The agreement also stated that the agency
was authorized to collect the artist fee from the promoter, with the
artist fee defined as net of the agency’s commission.
We wrote a scenario in which the Artist (removing Polo G’s name)
had to cancel the show after a slip-and-fall injury shortly before the
concert. We gave Respondents the entire contract (again, changing the
names and some location details), including a part of Paragraph 10.2. It
reads:
10.2 Artist acknowledges and agrees that Agency has provided
valuable services with respect to the Performance and that
Agency’s obligations to Artist will be deemed satisfied upon
Agency’s successful brokerage of an agreement with Promoter for
the Performance.
[MASKED: Agency is entitled to Commission even if the
Performance has not taken place for whatever reason, unless
such non-Performance is the result of Agency’s own gross
negligence.]
We then asked subjects to decide whether “Based on your best
understanding of the contract and your estimation of the missing
clause, is the agency . . . entitled to its commission even though the
show was cancelled due to the artist’s injury?” They were presented
with four options88 (in randomized order):89
Id.
We chose a multiple-choice format to ensure rigor in scoring: grading open-text responses
would require judgment calls about partially correct answers, implications, and levels of
abstraction—effectively reintroducing the very debates about interpretation this design
exists to escape. The format also matches the institution our proposal serves. Adjudication
is about preferred interpretations, not open-ended generation: each side proffers a reading,
and the court selects among them.
The randomized order was important because models exhibit a general preference for
options in particular positions in multiple-choice questions. See Chujie Zheng et al., Large
Language Models Are Not Robust Multiple Choice Selectors (arXiv, Working Paper No.
2309.03882, 2024), https://arxiv.org/abs/2309.03882 (finding that LLMs “prefer to select
specific option IDs as answers” and are accordingly vulnerable to option-position changes);

'#

a. No, because the show didn’t happen, so there’s no
artist fee from which to pay commission
b. No, because force majeure events (like medical
emergencies) void all payment obligations
c. Yes, the agency is entitled to full commission because
it successfully brokered the deal
d. Yes, but only to partial commission (such as 50%) to
account for the cancellation
Given the masked term, answer C is correct: the agency is entitled
to its commission because the slip-and-fall did not result from its gross
negligence.
The attentive reader would have gotten some hints about the right
outcome from contractual text outside of the masked clause. Section 6
states that if the artist cancels due to an accident and no show can be
rescheduled, the artist must return paid deposits. It even recites that
“A cancellation does not affect the right of Agency to Commission.”
However, this does not fully resolve the issue, as it does not spell out
how much of the commission is owed or the limitation in cases of gross
negligence. Still, an attentive reader could use this information to rule
out some of the options.
We then asked Respondents a follow up question if they had
previously answered correctly:
Question 2: “Assuming the agency IS entitled to full commission,
under what circumstances would the agency LOSE its right to
commission?”
a. If the show cancellation was due to any fault or error by
the agency
b. Only if the show cancellation resulted from the agency’s
serious misconduct or extreme carelessness

Pouya Pezeshkpour & Estevam Hruschka, Large Language Models Sensitivity to the Order of
Options in Multiple-Choice Questions (arXiv, Working Paper No. 2308.11483, 2023),
https://arxiv.org/abs/2308.11483 (documenting performance gaps of thirteen to seventy-five
percent when answer options are merely reordered). Randomization also induced nonsemantic variation in our prompts, in case that would make life harder for the LLMs. See
generally Choi, supra note 19, at 13–19.

c. Only if the agency failed to use “best efforts” to ensure
the promoter met its obligations
d. Under no circumstances-once the deal is brokered,
commission is guaranteed regardless
The correct answer is B, consistent with the (masked) gross
negligence carve out.90
2.

Contingency Fee

The Contingency Fee scenario results from a fee agreement
between an injured worker and a law firm retained to pursue a
workplace injury claim.91 The worker, who became a client, suffered a
serious injury at work and the law firm advised him that similar cases
resulted in judgments of around $210,000, but that litigation might take
two to three years. The agreement assigned the firm a contingency fee
of 33.3% before filing of a complaint and 40% after filing.
We wrote a dispute where, one day after signing the agreement, the
client received a take-it-or-leave-it $60,000 settlement offer directly
from the employer’s legal counsel. Facing financial pressure after being
out of work for three months, the client accepted the offer without
consulting his attorneys and then told the attorneys that he would not
need their services.
We again gave the respondents the entire contract to read, with one
part marked as masked, reproduced below:
Settlement. [MASKED: No settlement shall be made without
consent of both parties. If Clients settle without Attorneys’

In a separate branch of the experiment, to test the potential effect of political concordance,
we use two variants of the same background scenario while holding the relevant contract
terms constant. In the first variant, the artist is called T-Real, a hip-hop artist from Atlanta
who has “become a voice for urban youth empowerment and social justice.” Other scenario
indicators suggest progressive leanings. In the second variant, the artist is identified as
country singer Cal Walker, whose songs play at GOP rallies. Other indicators suggest
Republican leanings. This branch of the experiment did not produce results that conformed
with our preregistered hypothesis, which we tentatively attribute to unwanted social
desirability effects. We will return to the question showing models and motivated
interpretation in future work.
Ex. 4, Halks v. Kindley, No. 3∶25-cv-00560-DMS-AHG (S.D. Cal. Mar. 10, 2025), ECF No.
1-5.

'$

knowledge or consent, Clients owe the full contingency fee plus
any costs advanced.]
We asked respondents:
“Based on your best understanding of the contract and your
estimation of the missing clause, are the attorneys entitled to
payment even though they didn’t participate in obtaining the
settlement?” they can choose among the following randomized
options:
a. No, because the attorneys did not participate in
obtaining the settlement
b. No, because the family can terminate the
agreement within a reasonable period after signing
c. Yes, to their contingency fee as calculated in the
agreement
d. Yes, but only if the attorneys obtained a written
settlement offer
The correct answer is C. The masked clause directly anticipates a
scenario of a settlement negotiated without the lawyers and provides
that lawyers are nonetheless entitled to contingency fees.
Again, if you read closely, you might infer the content of the masked
clause in other contractual provisions. Section 4(a) states the
contingency percentages and defines gross recovery broadly. Section
4(d) grants the attorneys a lien on claims and sums received to secure
fees and costs, and also provides special protection if clients discharge
the attorneys after a written settlement offer has been obtained.
The follow-up question here was
Question 2: “If the attorneys ARE entitled to their contingency fee,
how much would that be?”
a.

The pre-complaint percentage rate specified in Paragraph
4(a)
b. The post-complaint percentage rate as a penalty for
unauthorized settlement
c. A reduced percentage to reflect minimal work performed
d. An hourly rate for the consultation time only

The correct answer is A, as indicated by the language “full
contingent fee.”92
3.

Bottles

The Bottles scenario is based on a requirement contract between a
bottle manufacturer, CKS, and a beverage company, Milo’s Tea
Company.93 Under the agreement, Milo’s Tea Company agreed to buy
100% of its requirements for specified plastic containers from CKS. The
contract listed bottle types, destinations, prices per thousand units,
estimated annual volumes, minimum production runs, and related
commercial terms.
We generated a dispute by having Milo’s send CKS a message
saying that it expects to need “something on the order of an extra
110,000 (a full truckload) of the 20oz bottles” for an upcoming
marketing campaign, explaining that it anticipated grocery-chain
orders. CKS ordered materials, produced the additional 110,000
bottles, held the finished bottles for several months, and incurred
storage costs. Milo’s then said it did not need the bottles and refused
to take them. CKS, objecting because resale is hard, sent the bottles
back in a truck, asking to be paid.
The relevant, masked contract provision, stated:
Inventory. [MASKED: If inventory or materials held for Milo’s
Tea as a result of a purchase order or forecast become obsolete
during the Term, or have been held by CKS for more than three
months, or are on hand after expiration or termination, CKS
may deliver such inventory to Milo’s Tea and invoice at thencurrent prices or indexed raw material prices plus storage costs,
payable within 30 days.]
We asked respondents to consider the entire contract and to decide
whether under the masked clause CKS can require Milo’s “to purchase
and pay for the excess bottles produced.” The possible answers were:
As in the Artist scenario, we altered the document only to fit the background scenario that
uses fictional names. We again had two variants to test political or ideological hypotheses
that we will discuss in a separate paper.
Ex. 1, C.K.S. Packaging, Inc. v. Milo’s Tea Co., No. 3∶25-cv-00897 (M.D. Tenn. Aug. 7, 2025),
ECF No. 1-1 (2020 Purchase & Supply Agreement; masked § 12).

'%

a. No, the parties did not agree to use forecasts for inventory
planning
b. No, the parties only allowed for payment based on signed
purchase orders
c. Yes, because Milo’s Tea Company provided a projection
that CKS relied upon
d. Yes, because the bottles have been held for an extended
period without purchase
The correct answer is C. Under the masked clause, CKS may
invoice Milo’s for inventory or materials held as a result of a purchase
order or forecast. Given the Milo’s gave a projection, it is liable for the
price of the bottles.94
The follow-up question here was
Question 2: “If [CKS] CAN require payment, at what price may
[CKS] invoice [Milo’s]?”
a. At [CKS]’s option, either the current market price
($8,950) or the raw materials plus storage costs
($6,400)
b. At [CKS]’s option, either the contract price
($8,516) or the current market price ($8,950)
c. Only the contract price specified in Section 1
($8,516)
d. Only the raw material costs plus storage ($6,400)
The correct answer is A, as indicated by the language “then-current
prices or indexed raw material prices plus storage costs.”95
C. Measures and Methods
We pre-registered certain hypotheses that cluster around three
questions: whether humans and machines interpret above chance and
Answer B would be a sensible choice but it’s atextual. Answer D is also tempting because the
masked clause also refers to inventory that has been held for more than three months. But C
is the better answer because the operative fact in the scenario is that Milo’s forecast caused
CKS to produce the extra bottles. The masked clause clarifies, although its grammar is
awkward, that CKS may charge Milo’s for extra inventory tied to Milo’s own purchase orders
or forecasts, rather than for arbitrary inventory CKS chose to produce for other buyers.
Unlike the Artist and Contingency scenarios, the Bottles scenario was a control and did not
contain politically coded personal identity conditions for the purposes of our experiment.

relative to each other (H1); whether responses shift in a plaintifffavorable direction under political concordance (H2a); and whether
that shift differs in magnitude between humans and machines (H2b).
Secondary hypotheses examine legal training (H3), scenario
complexity (H4), and racial concordance as an exploratory extension
(H5).96
We collected data from lay respondents, law students, and lawyers,
as well as a suite of LLM models.
We use Prolific as our data collection platform, in part because of
its robust protections against the use of AI bots by its human subjects.97
We included attention and time-to-complete checks, the former meant
to filter out bots and the second to raise flags of AI assistance. We also
invited respondents to write comments. The AI detection software
Pangram V3.3.2 classified all 456 submitted comments and marked
them as 100% human written.98
We gathered responses from about 500 Prolific respondents, paying
each the equivalent of about $12/hour, and an additional bonus of two
gift cards of $100 raffled among those who answered the survey
correctly. 99 Consistent with best practices, we also implemented an
attention check, leaving us with 465 participants.

We specified robustness checks, including alternative operationalizations of political identity,
alternative quartile cutoffs, and a placebo analysis on the excluded middle quartiles, in
advance. The full pre-registration document appears in the supplemental materials. We leave
the racial and political concordance discussion to a different paper.
Prolific employs a host of detection tools meant to pick up on AI signatures. A recent study
benchmarked human respondents against 240 AI-agent runs, and found that 98% of Prolific
respondents passed a video attention check that only 17% of agents could pass, and that
virtually no Prolific respondents carried the server-IP signatures characteristic of AI agents.
Can Çelebi, Christine Exley, Sören Harrs, Hannu Kivimaki, Marta Serra-Garcia & Jeffrey
Yusof, Mission Possible: The Collection of High-Quality Online Data (Mar. 10, 2026)
(unpublished
manuscript),
https://github.com/survey-data-quality-lab/missionpossible/blob/main/paper/DataQuality_March2026.pdf.
Overall, we had 424 unique comments from 250 commenting respondents. The audit is on
file with the authors, available on request, and will be hosted online after this paper is
finalized.
We had overall 501 completions, after we removed two duplicates and 34 attention failures,
the N=495.

'&

Table 1. Lay Respondent Demographics Compared with U.S.
Benchmarks100
Sample

U.S.

Sample
Difference

Median age

39.1-2.1 years

Female

52.5%

50.5%

+2 pp

White

66.9%

57.5%

+9.4 pp

Bachelor’s degree or
higher

59.1%

35.7%

+23.4 pp

Democratic Party

39.6%

27.0%

+12.6 pp

Republican Party

24.7%

27.0%

-2.3 pp

Feature

Our sample is fairly representative of the overall US population.
The most important skew is that our respondents are more educated
and more liberal than the population at large; the former might bias our
results toward more accurate answers.101
For law students, we recruited a mix of seventy-seven students at
the end of their 1L year, close to their final exam in contracts, and 2Ls
and 3Ls sent an invitation to participate by email.
For lawyers, we recruited participants in a convenience sample,
leveraging connections to school alumni. Ultimately, forty-eight
lawyers completed the survey. Median years in practice were 18.5, and
48% reported 20 or more years of practice. The sample was weighted
toward law firm lawyers and in-house counsel, with smaller numbers
from government, judges, and solo/independent practice.

This sample focuses on those respondents who passed the attention check.
See generally Krin Irvine, David Hoffman & Tess Wilkinson-Ryan, Law and Psychology Grows
Up, Goes Online, and Replicates, 15 J. EMPIRICAL LEGAL STUD. 320, 326 (2018) (arguing that
online pools are generally useful tools for conducting law and psychology experiments).

For LLMs, we used the highest performing models at the time we
recruited the lay sample (the “frontier models”).102 We chose not to
update our work with newer models as the writing progressed, to
minimize the possibility that newer models may have been exposed to
the source data or even indirect possibility of learning of our study from
workshop announcements on social media chatter.
D. Results
Let’s start with the bottom line: did respondents correctly predict
the hidden contract term? Figure 1 summarizes our findings with the
large sample of lay respondents:

Figure 1: Unmasking Accuracy of Lay Respondents

Averaging across all three scenarios, lay respondents got it right 55%
of the time. This is a bit over double the rate that chance would predict.
This was a frankly relieving result (though one we had pre-registered):
lay subjects were able to fill a gap in a contract at rates better than
chance, based on its immediate context and their own background
knowledge. That said, we observe considerable variation by scenario.
In the artist scenario, respondents were barely better than chance, with

We selected six models, with twenty runs per model, resulting in 120 runs and 240 scenario
exposures. Temperature was set to zero (variation came from answer ordering), no browsing
tools were enabled.
#!

accuracy rate of 32.3%, but in the contingency fee scenario, prediction
was 71.1% accurate.
Figure 2 illustrates law student performance. Again, respondents
were better than random guessers. And law students were marginally
better than lay respondents, though the gap is not statistically
significant.

Figure 2: Lay and Law Student Respondents.

Figure 3 adds lawyers to the mix.

Figure 3: All Three Groups of Human Respondents.

Lawyers are capable contract interpreters: they correctly predicted
how to fill gaps nearly 60% of the time. Disaggregating by scenario,
however, reveals a different picture: while lawyers got it right (much)
more often than other human respondents on the Artist and
Contingency Fee scenarios, the converse was true for the Bottles
scenario.
Most lawyers who got the bottles wrong concentrated on option B,
which holds that no payment was due because the parties could only
have permitted recovery based on signed purchase order, not
projections. This was, notably, a quite sensible guess. But it does not
reflect the specific arrangement that the parties arrived at in this case.
At the same time, lawyers performed exceedingly well on the
Contingency Fee scenario, getting it right 82.6% of the time. This result
suggests that lawyers were benefiting from their broad familiarity with
contingency fee agreements. Pattern matching based on domain
familiarity, then, might be the explanation for both cases: the
contingency fee agreement aligned with industry norms while the
bottles scenario likely deviated from it.

##

Lastly, we arrive at the LLM models, summarized in Figure 4
below.103

Figure 4: Adding frontier models

As a group, the LLMs were 88.3% accurate in unmasking the
agreements. This rate far exceeded human performance. The LLMs
were essentially strong interpreters across each of the scenarios, with
some small variation.
On Artist, where lay respondents struggled most (32.3% accuracy,
only seven points above chance), the leading models cleared 90%. On
Contingency Fee, where lay respondents performed best at 71.1%, the top
models answered every unmasking question correctly. Bottles, the
longest and most technical packet, was also the most challenging for
the LLMs, though they were still ahead of the human respondents.
Figure 5 breaks down this performance by model.

We present this data using a weighted average that accounts for the protocol showing certain
scenarios more often than others.

Figure 5: Comparing frontier models

The LLM panel is variably capable. Opus 4.6 scored 100% on
scenario-balanced unmasking; Gemini 3.1 Pro, Grok 4.2, and GPT 5.4
ranged between 88% and 97%; Qwen 3.6 reached 82%; and GLM 5.1
trailed the panel at 70.0%, still well above the human rate. For context,
the two trailing models are generally thought to be the weakest of the
bunch.104
The headline measure asks whether the respondent identified the
correct meaning of the redacted provision. We probed that result with
a follow-up question designed to test a more specific implication of the
same hidden language. That is, did the respondents get the right
answer for the right reasons?

See, e.g., LMArena, Text Arena Leaderboard, https://lmarena.ai/leaderboard (last visited
July 15, 2026) (crowd-sourced pairwise-preference rankings placing GLM 5.1 more than
twenty ranks below the leading models in our panel); Artificial Analysis, LLM Leaderboard,
https://artificialanalysis.ai/leaderboards/models (last visited July 15, 2026) (composite
benchmark index scoring the GLM 5.1 and Qwen 3.6 model families well below the panel’s
frontier models).
#$

The odds of answering both multiple choice questions correctly are
1/16 (6.25%). The follow-up therefore gives us a tougher measure of
interpretive accuracy: it helps distinguish respondents who recovered
the operative meaning of the provision from those who arrived at the
correct headline answer by chance, by elimination, or by relying on only
a coarse feature of the surrounding contract.

Figure 6: Strict Reconstruction results

Roughly 26% of human respondents answered both questions
correctly when chance would predict only 6.25%. And legal training
again monotonically improved performance. But LLMs exceed all
groups. While LLM accuracy fell by this stricter measure, it was still
about double that of human interpreters.
Digging deeper, Figure 6 below disaggregates by scenario.

Figure 7: Strict reconstruction with LLMs

For human interpreters, the artist and bottle follow up questions
presented a significant challenge, while the contingency fee follow up
was significantly easier.
LLMs had a different nemesis. They breezed through the artist and
contingency fee questions, but were simply unable to correctly answer
the follow-up bottle scenario question: among the eighty-four runs that
correctly answered the headline bottles question, none selected the
correct follow-up answer; eighty-one chose option C and three chose
option B. Recall that this option held that CKS was only entitled to the
contract price, rather than current market price or raw materials and
storage costs. Presumably, the models over-indexed on section 1 to the
contract, which stipulates product prices.
E. Robustness Checks: Perturbed Contracts
One possible explanation for LLMs’ strong performance on these
tasks may be that they gathered contextual clues from the contracts and
accurately predicted the gaps we manufactured. But two different
possibilities are plausible: (a) extrapolation about what the gap would
say based only on it looking like other contracts; and (b) test-hacking,
where the models infer the right answer from subtle drafting cues of
the options presented to them.

#%

To investigate these possibilities, we ran two analyses that perturbed
each contract to test how models derived their answers. We first took
the original contracts and redrafted them to flip their direction. If the
original contract was pro-buyer, the perturbed agreement was proseller. Our idea was to see if models predicted a different set of answers
given this changed data internal to the text itself.
Second, we asked the models to predict the masked clause blind. The
prompt instructed that due to user error the contract was not attached
and that the model should infer the right answer to the best of its ability.
We provided only the contract’s title. The perturbed scenarios
otherwise replicated the question stems from our original study.
Figure 8 summarizes the findings of the perturbation analysis.

Figure 8: Perturbation Analysis

The main result is that, in line with our expectations, perturbation
reduced accuracy: from 88% to 60% for the flipped condition, and to 68%
in the no-contract condition.
Figure 9 considers LLM performance on each of the contract types.

Figure 9: Perturbation Analysis, By Contract Type

Stripped of the contract, model responses appear to measure
general contract expectations. In both the Bottles and Contingency Fee
scenarios, omitting the contract ironically did not reduce accuracy, and
indeed raised it in a statistically insignificant fashion. This implies that
baseline expectations can be powerful decoders of some contracts,
presumably when those deals match industry norms. For the Artist
Contract, however, accuracy collapsed to a hair above chance: that deal
overall was apparently distinct from the normal brokerage contract, and
the model was unable to infer the gap without the particularized
contract that generated it.
F. Scaling Up: 119 Unseen Contracts
A different concern is that our accuracy results may be the
artifact of picking three contracts that happen to be friendly to the work

#&

of language models. To what extent do our results extend to diverse,
unseen agreements?
To answer that question we collected 119 commercial contracts
from SEC EDGAR exhibits, predominantly from 2026 filings: 87 from
2026, 31 from 2025, and one from 2024. 105 From each we masked a
single material clause and repeated a similar experiment: given the
surrounding contract, a short scenario, and four options, predict the
agreement made by the parties. (Saying “we” here elides that in fact
we had the models do the work.)106
Across the 119 unseen contracts and six models, the models
recovered the masked clause 87% of the time, every model landing
between 84% and 92%. This result is very close to the accuracy results
in the main study. The following figure summarizes these results:

While the publication of the exhibit was made in 2026, it is possible that some contracts were
made public in some form before then. It is not feasible to audit the models for specific
knowledge of these contracts, but it is also highly unlikely that many of these contracts
featured heavily in the training data. We consider contamination to be a generalized
background risk.
You might worry that the AI designed and passed its own tests. But the masked clauses are
genuine text drafted by practicing lawyers, not invented by a model. We iterated several times
on the drafting of the questions and randomized answer order. And because we use a diverse
body of models, any AI self-preferencing in some latent ways would be unlikely to carry over
model families.

Figure 10: Testing 119 2026 EDGAR contracts

To better understand model accuracy, we split the various
contracts by type. Consistent with the discussion above, the models
best predict gaps in templated, market standard documents:
promissory notes (98%), securities and subscription agreements (96%),
and credit facilities (95%). By contrast, they fail more (though not most
of the time!) in predicting negotiated or bespoke terms:
indemnification provisions (61%), registration rights (67%), and,
tellingly, employment agreements (76%).

(!

Figure 11: LLM performance by contract type

Overall, the models were highly accurate: all six models selected
the correct answer on 87 of the 119 contracts. Errors were concentrated
rather than uniformly distributed. The fifteen trickiest contracts (12.6%
of the sample) accounted for about 75% of the errors and all six models
missed the same five contracts. An exploratory, model-assisted audit
classified each of those five clauses as reversing a conventional market
or legal default.107 For example, one license agreement permitted the

We asked a model to classify each of the 119 masked clauses by whether it conformed to or
departed from the applicable market or legal default. Even expert reviewers might disagree
on the true classification, so we treat this as suggestive evidence. When the masked clause
matched the default, models answered correctly 93.6% of the time. When the masked
provision was anti-default—where a model that simply recites the boilerplate would score
essentially zero—models answered correctly 59.8% of the time. This tracks our perturbation
results, where flipping the contracts’ orientation likewise changed accuracy to about 60%.

licensee to sublicense its rights to any third party without the licensor’s
prior consent.
Wrong answers also tended to converge. 108 On two of the five
shared misses, all six models selected the same wrong answer; on the
remaining three, their wrong answers split 4–2 or 5–1. Interestingly,
model disagreement was a sign that something was off. In about 94% of
the contracts where model made mistakes, they also disagreed among
themselves.

Figure 12: Hard and Easy Contracts

G. General Discussion of Results
Lay subjects filled contractual gaps about twice as well as chance
alone would predict. Working from the four corners of the agreement,
and their life experiences, ordinary readers successfully predicted
This suggests that models combine boilerplate expectations with contract-specific
information. We also note that accuracy was not related to contract length, clause length, or
mask position. Computations on file with authors, available on request.
Pairwise correlations between the six models’ binary error indicators averaged .54 across the
fifteen model pairs, ranging from .36 to .75.
(#

masked terms. Readers with greater legal expertise did just as well, if
not better, inching towards 60% accuracy on interpretative tasks.
We hypothesize that respondents generate gap filling predictions
via two modes of inference. The first is general domain knowledge,
common sense, and an understanding of the goals the parties sought to
achieve. Deductions from that knowledge base explain lawyers’
superior ability to predict a counsel fee provision compared to lay
respondents.
At the same time, respondents learned from the contracts they
read. Text that surrounded the masked gap carried what theory calls
“mutual information.” 109 The clauses of an agreement are not
independent of one another: the price reflects the risk allocation, the
risk allocation reflects the terms on excuse and termination, and so on
down the document.110 Because the provisions are linked, the ones a
contract lays out carry information about the agreement’s missing
terms. It is the same property that lets a radio signal remain intelligible
over distance and through interference: enough of the message is
carried elsewhere that the lost part can be reconstructed.111
From these sources, human inferences were decent, but AI
predictions were extraordinary. Frontier models were 88.3% accurate in
predicting the masked terms, and 54% on the stricter measure that
See Claude E. Shannon, A Mathematical Theory of Communication (pts. 1 & 2), 27 BELL SYS.
TECH. J. 379, 623 (1948) (founding the mathematical theory of information); THOMAS M.
COVER & JOY A. THOMAS, ELEMENTS OF INFORMATION THEORY 19–20 (2d ed. 2006)
(defining the mutual information of two random variables as “the reduction in the
uncertainty of one random variable due to the knowledge of the other”).
See Albert Choi & George Triantis, The Effect of Bargaining Power on Contract Design, 98 VA.
L. REV. 1665, 1670–71 (2012) (unpacking the “irrelevance proposition” that bargaining power
moves only the price term and showing how price and nonprice terms adjust to one another);
see generally Goetz & Scott, Limits, supra note 5 (analyzing the interactions between express
and implied contract terms).
See Shannon, supra note 109, at 398–99 (defining redundancy, estimating that ordinary
English is “roughly 50%” redundant, and observing that one can “delete a certain fraction of
the letters from a sample of English text and then let someone attempt to restore them”);
Claude E. Shannon, Prediction and Entropy of Printed English, 30 BELL SYS. TECH. J. 50, 54
(1951) (estimating the entropy of English from experiments in which human subjects guessed
concealed letters from the surrounding text—the same prediction task we provide in this
article for lawyers and machines); see generally JOHN R. PIERCE, AN INTRODUCTION TO
INFORMATION THEORY: SYMBOLS, SIGNALS AND NOISE (2d rev. ed. 1980) (providing an
accessible account of entropy, redundancy, and the noisy channel).

measures whether they could guess the term’s limits. These rates well
exceed chance (25% and 6.25% for the stricter measure), and also exceed
every human group by margins that are large and significant. We
observe that differently trained models, run under different settings,
with different prompts and randomized answer orders, converged on
the same answers and with less spread than human readers.
Consistency brought with it its own complications. We found that,
across the 119 unseen contracts, models tended to be wrong together.
Of the 90 errors models made, 79 fell on the most common wrong
answer. Still, model disagreement was a valuable sign that something
was off: model disagreement was predictive of 30 out of the 32
contracts on which models made errors. 112
Readers may worry that the model performance we’ve described
may be an artifact—that is, the results lack external validity. One
version of this worry is that the models were just predicting what courts
would do, even though we instructed them to tell us what the contracts
themselves said. Another is that pattern matching is to contract
Another is that pattern matching is to contract type, not something
specific about the deal itself. Our robustness tests should alleviate both
concerns. Perturbed contracts and missing contracts resulted in
weaker performance, just as theory would predict, at least with respect
to deals that were less anchored to the market standard. And when we
fed models over a hundred unseen agreements, the main results
persisted, suggesting that cherry picking is not by and large a concern.
We thus would conclude that AI models extract more textual
information from contracts than humans. This is no surprise. Models
are trained to detect patterns in vast troves of data, so large that no
human can ever read them in a lifetime. And this was a cognitively
demanding, relatively long, task for an online survey.
Our third finding relates to legal experience. Lawyers beat the
accuracy of law students and lay people, though these differences were
not statistically significant in our sample. (The significance finding is in
part an artifact of the lack of precision in the multiple choice estimates

Pairwise correlations between the models’ error indicators average .54, and joint errors
occur 4.1 to 6.2 times as often as independence would predict; all fifteen pairwise associations
survive correction for multiple comparisons.
($

we created). But, lawyers underperformed both students and lay
readers on Bottles.
The data is most consistent with the literature on expertise, which
holds that it results from a form of pattern matching, 113 and pattern
matching cuts both ways. It supplies real inferential leverage when a
deal is typical and the pattern holds; it misleads when the deal is
bespoke and the parties have contracted around the pattern. The
lawyers who got the Bottles contract wrong concentrated on the answer
holding that recovery ran only to signed purchase orders, a sensible,
common arrangement, and very likely the industry default.
Unfortunately, the parties had in fact tied the obligation to forecasts.
Legal training pulled them toward the standard term and away from the
one in front of them.114
AI models share this vulnerability. On the Bottles follow-up, the
models overwhelmingly collapsed onto a single wrong answer, the
option holding that the manufacturer could invoice only the contract
price stated in Section 1, rather than the market-or-replacement-cost
figure the masked clause actually specified. They were pulled by a
conspicuous number elsewhere in the document and away from the
operative local language.
When we omitted contract language entirely, model accuracy
dropped. But those changes were not universal, and in the context of
the highly familiar contingency agreement, the models were just as
accurate without seeing any of the language. This is both a strength and
a weakness of the models. When we made the contracts flip their
valence, today’s models were excessively obstinate in their predictions.
All of this goes to a general question about expertise and domain
familiarity. Do the fruits of deep domain knowledge pay for the
occasional failure to notice a particular idiosyncratic agreement? And
does it make sense to occasionally install low-information
Karl N. Llewellyn, On Reading and Using the Newer Jurisprudence, 40 COLUM. L. REV. 581,
589–91 (1940) (arguing that judges can be expected to converge in their rulings based on a
shared apprehension of the pertinent features of recurring “situation types.”).
Cf. Elisabeth S. Clemens & James M. Cook, Politics and Institutionalism: Explaining
Durability and Change, 25 ANN. REV. SOC. 441 (1999) (noting how certain procedures come
to be seen as the standard operating procedure).

decisionmakers, such as juries, to ward off excessive reliance on
expertise?
Finally, we return to the question of the generalizability or external
validity of our findings. An important limitation of our experiment is
that our gaps were manufactured, not (as in nature) exposed by
contingent events. The worry would be that recovering manufactured
gaps is simply a qualitatively different activity from guessing what the
parties would have said had they deigned to write something down.
But for an unwritten term to be qualitatively unrecoverable from
the text, it would have to bear no patterned relationship to the rest of
the deal. That assumption is substantively implausible. Drafters work
within deal types, allocate risk consistently across provisions, and
adjust price and other margins as terms move. A discontinuity between
the predictability of written and unwritten terms would require all of
that structure to break off exactly at the boundary of what was
committed to paper.
A related concern is that the terms parties choose to write are not
selected at random. We agree they are not—but the selection runs in
our favor. At the margin, parties write what does not follow obviously
from the rest of the deal and economize on what does. The terms we
masked are therefore the harder cases, and our estimates likely
understate how much of an ordinary gap is recoverable from text.
III.

T HE P R ACTI CE AND P ERI LS OF G ENER ATI VE
G AP F I LLI NG

We’ve shown that AI agents are pretty good at uncovering what a
contract says behind a curtain we’ve drawn. At the very least, they are
better than humans, even lawyers. Does that mean they are ready for
deployment in the real world? To what ends and with what limits? This
section takes on the project of turning our experiment into a workable
set of recommendations for courts and practitioners, who are
increasingly using this new technology in their day-to-day professional
lives.115

See sources cited supra note 28.
(%

A. Generative Gap Filling Within Contract Litigation
Imagine a televised court proceeding interpreting a contract of
wide interest—the Katrina flood exclusion, 116 Donald Trump’s
NDAs, 117 Pepsi’s jet prize 118 —that consisted of nothing more than
feeding the contract to an oracle on a mountaintop which then, by
resort to magic or the divine, produced a distribution of probabilities.
Even in today’s fallen moment for the rule of law, this kind of
interpretive procedure would not fly. And we don’t recommend it
here.119
Critics worry that oracular interpretation stands on the wrong kind
of reasons, and contract law lives or dies by the kind of reasons it
produces. As Robert Cover explained, courts are “jurispathic”: they
extinguish real meanings that were available in the world to advance
state objectives. 120 Legal meaning follows from this process of
selection, and is legitimate only if someone is “prepared to live by it . . .
The transformation of interpretation into legal meaning begins when
someone accepts the demands of interpretation and, through the
personal act of commitment, affirms the position taken.”121
A version of generative gap filling would be to treat it like our
hypothesized oracle. Indeed, some have already proposed just that: to
retire judges and turn over interpretation to endlessly correct, efficient,
tireless, Babbage engines.122 The problem with those proposals is that
See, e.g., In re Katrina Canal Breaches Litig., 495 F.3d 191, 199 (5th Cir. 2007) (denying relief
to homeowners).
Denson v. Donald J. Trump for President, Inc., 530 F. Supp. 3d 412 (S.D.N.Y. 2021)
(declaring NDA unenforceable).
Leonard v. Pepsico, Inc., 88 F. Supp. 2d 116 (S.D.N.Y. 1999) (denying recovery to plaintiff
who probably knew better).
In fact, Arbel is notably more optimistic about the ability of these models to produce
sociologically legitimate results than Hoffman, and was dragged kicking and screaming into
agreeing to these paragraphs.
Robert M. Cover, The Supreme Court, 1982 Term—Foreword: Nomos and Narrative, 97
HARV. L. REV. 4, 44 (1983).
Id. at 44–45.
The two best known examples are by practitioners who offered their judge-replacement
theories in papers that have almost no engagement with the relevant literature about judging.
See Adam Unikowsky, In AI We Trust, Adam’s Legal Newsletter ( June 8, 2024),
https://adamunikowsky.substack.com/p/in-ai-we-trust; Kimo Gandall, Jack Kieffaber &
Kenny McLaren, We Built Judge.ai. And You Should Buy It, SSRN (Mar. 27, 2025),
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5115184.

citizens believe that legal interpretation, to produce sociologically
legitimate results, must result from human judgment, both in reality
and in how it is publicly defended. 123 Identifying the parties’ predispute expectations is a goal of contract interpretation—perhaps even
the most important one. And yet getting that “right” through an
illegitimate method would be folly.
But nothing in what we have shown requires or even implies the
replacement of judges by models.124 The natural posture for generative
gap-filling is as an input to adjudication—evidence introduced by the
parties, contested through the ordinary adversarial machinery,
weighed and explained by the judge in an opinion that gives reasons in
the ordinary way, and open to appeal.125
Consider a dispute closely modeled on our Bottles scenario, which
so tripped up the lawyers in our sample.126 Imagine that a manufacturer
produces a quantity of glass containers in reliance on what it
characterizes as a binding rolling forecast from its buyer. But the buyer
characterizes the forecast as informational only and contends that its
obligation runs only to product covered by signed purchase orders.127
See, e.g., Benjamin Minhao Chen, Alexander Stremitzer & Kevin Tobia, Having Your Day in
Robot Court, 36 HARV. J.L. & TECH. 127, 131 (2022) (noting that “[p]roceedings conducted by
human judges were seen as fairer than those conducted by AI judges”); see also generally Anna
Fine, Emily R. Berthelot & Shawn Marsh, Public Perceptions of Judges’ Use of AI Tools in
Courtroom Decision-Making: An Examination of Legitimacy, Fairness, Trust, and Procedural
Justice, 15 BEHAV. SCI. 476 (2025), https://doi.org/10.3390/bs15040476 (experimental study
finding “[j]udicial legitimacy … was significantly higher for judges who relied on their
expertise than those who incorporated AI”).
AI arbitrators are a different kettle of fish, in part because parties opt-in (at least notionally)
to that process, and because the stakes of some arbitrations are so low that an algorithmic
decisionmaker would offer real welfare gains. See generally Michael J. Broyde & Yiyang Mei,
Don’t Kill the Baby! The Case for AI in Arbitration, 21 N.Y.U. J.L. & BUS. 119, 119 (2024) (“This
article examines the integration of AI into arbitration, arguing that the Federal Arbitration
Act (FAA) allows parties to contractually choose AI-driven arbitration, despite traditional
reservations.”).
Even the humble dictionary—the most familiar interpretive aid—is routinely fought over
through exactly this machinery. See, e.g., Taniguchi v. Kan Pacific Saipan, Ltd., 566 U.S. 560,
566–69 (2012) (canvassing competing dictionary definitions on whether an “interpreter”
includes a document translator).
See supra Section II.B.3 (describing the Bottles scenario).
Disputes of this shape are a staple of supply-chain litigation. See, e.g., Empire Gas Corp. v.
Am. Bakeries Co., 840 F.2d 1333, 1335, 1338–41 (7th Cir. 1988) (Posner, J.) (resolving a
requirement contracts dispute); Simcala, Inc. v. Am. Coal Trade, Inc., 821 So. 2d 197, 202–
05 (Ala. 2001) (holding a purchase order’s stated 17,500-ton estimate binding).

(&

The supply agreement, as may be typical of such agreements,
addresses minimum quantities, pricing tiers, and quality specifications
in considerable detail. It is silent on the legal effect of the rolling
forecast itself. The parties are now in court, and the question is
whether the forecast obligates the buyer to pay for the excess
inventory.
In current doctrine, this is a paradigmatic gap-filling case. At least
platonically, a court (not a jury)128 would try to figure out what to do
using some stew consisting of (at least) Restatement § 204, 129 the
UCC’s parallel provisions,130 the implied covenant of good faith and
fair dealing as developed in the relevant jurisdiction, 131 course of
dealing under § 1-303 if there was a prior course,132 and trade usage if
either party puts on a witness.133
Given this varying set of inputs, is there really any reason to think
that jurists will reliably get to the “right” result, if by “right” we mean
what the parties would have done? Our experimental evidence suggests
they won’t, as the real deal apparently runs afoul of ordinary
commercial norms. The problem is hard, and the more degrees of
freedom that the decisionmakers have the less their judgments will
cohere. This is precisely the long-standing critique of gap filling
doctrine. 134 Nor is the hypothetical-bargain inquiry disciplining,
Restatement (Second) of Contracts § 204 cmt. d (Am. L. Inst. 1981) (the court supplies the
omitted term “which comports with community standards of fairness and policy”); id. §
212(2) (interpretation is a question of law for the court unless it turns on the credibility of
extrinsic evidence or a choice among reasonable inferences from it).
Restatement (Second) of Contracts § 204 (Am. L. Inst. 1981) (“Supplying an Omitted
Essential Term”).
U.C.C. § 1-304; U.C.C. § 2-204(3) (a contract does not fail for indefiniteness where there is
“a reasonably certain basis for giving an appropriate remedy”); see also U.C.C. §§ 2-305, 2309 (supplying a reasonable price and a reasonable time).
Steven J. Burton, Breach of Contract and the Common Law Duty to Perform in Good Faith, 94
HARV. L. REV. 369, 369 (1980) (observing that American jurisdictions, the Restatement
(Second), and the UCC “now recognize the duty to perform a contract in good faith as a
general principle of contract law”).
U.C.C. § 1-303(b) (defining “course of dealing”); id. § 1-303(d) (course of dealing “may give
particular meaning to specific terms of the agreement, and may supplement or qualify the
terms of the agreement”).
U.C.C. § 1-303(c) (defining “usage of trade”).
Bernstein, Questionable, supra note 42, at 715 (findings “suggest that ‘usages of trade’ and
‘commercial standards,’ as those terms are used by the Code, may not consistently exist,
even in relatively close-knit merchant communities.”).

because any two judges might reach different answers from the same
record.135
Especially in commercial cases where both sides can afford to reach
for whatever evidence is at hand, the parties’ arguments are
recognizably inferential claims about the text.136 Neither side is required
to produce evidence directly in support of the inference.137 The court
will issue an opinion that picks one inference over the other, but all that
they can hope to offer is a limited explanation why they found the
chosen inference more probative than its rival, because there is no
ultimate truth on which such an explanation could be grounded.138
Generative gap-filling, deployed as an input rather than a
judgment, could help. Return again to the Bottles hypothetical. Under
the regime we propose, the manufacturer’s counsel would run the
Goetz & Scott, Limits, supra note 5, at 320 (describing the result of interpretation disputes
as a “lottery”); id. at 263 (observing “widespread judicial uncertainty over the proper
method of interpreting agreements that intermingle express and implied terms”); see also id.
at 269 n.14 (“Courts do perform this function, but their use of interpretive criteria is
problematic.”). For empirical demonstrations that identical records yield divergent
judgments, see Spamann & Klöhn, supra note 22, at 255 (giving U.S. federal judges the same
case file and finding that a legally irrelevant change in defendant identity moved affirmance
rates by forty-five percentage points, while a weak precedent had no detectable effect);
Farnsworth, Guzior & Malani, supra note 71, at 257–58 (finding that respondents asked
whether the same statutory text is “ambiguous” give answers “strongly biased by their policy
preferences”).
Gilson et al., supra note 56, at 37 (explaining that theories of contract “require courts to find
out, as far as is possible, what the parties meant by the words they used,” so that
interpretation proceeds by inference from the text); see also id. at 40 (describing the risk that
courts will “erroneously infer the parties’ preference for any particular contextual
interpretation”).
Id. at 60 (observing that the relevant inputs are “fully observable by the contracting parties
even if not verifiable to a court”); see also id. at 56 (noting that only where terms and
performance are “observable and verifiable” is “the likelihood of a court making a mistake
in interpreting the contract … reduced”); Michael S. Pardo, The Nature and Purpose of
Evidence Theory, 66 Vand. L. Rev. 547, 597 (2013) (on the explanatory conception of proof,
“[t]he primary explanations at issue are those provided by the parties,” which the factfinder
then assesses by “inference to the best explanation”).
See Bernstein, Questionable, supra note 42, at 717 (the usage courts purport to find is “a legal
fiction rather than a merchant reality”); Pardo, supra note 137, at 552 (factfinders “assimilate
evidence into competing narratives of the events and select the most plausible … of the
available accounts”); Charles Nesson, The Evidence or the Event? On Judicial Proof and the
Acceptability of Verdicts, 98 HARV. L. REV. 1357, 1358 (1985) (factfinders “see only evidence
of the act, not the act itself,” so the process “must somehow accomplish an inductive leap
from the evidence presented to a statement about a past event”; the object of factfinding is
“acceptable verdicts,” not demonstrated truth).

$!

unmasked contract through an AI model, with the disputed silence
identified, and ask it to predict what a term governing the legal effect
of forecast quantities would say in a contract structured this way. The
output—say, that the model assigns 87% probability to a bindingforecast reading and 11% to a nonbinding one, with the residual on a
third option—is disclosed to opposing counsel with the model
identified, the version specified, the prompt reproduced, and the
system instructions logged. Opposing counsel would then run its own
query, perhaps with a different prompt, perhaps with a different model,
and produce its own output.139
The court now will face a contested evidentiary record about what
the document itself implies. It can evaluate which prompt better
characterized the dispute, whether either prompt was loaded, whether
the chosen model has known weaknesses in commercial-contract
inference, whether the corpus on which the model implicitly draws is
appropriate to the deal type at hand. The judge can require the parties
to run sensitivity analyses,140 or even appoint its own neutral expert
under Rule 706 to perform an independent query.141 The efficacy of the
truth-seeking function of adversarial presentation might be contested,
but at the very least there’s nothing unfamiliar about it.142
Nothing about what we’ve proposed would replace the inferential
work that Judge Cardozo did in Wood v. Lucy, or the deductive
techniques that Judge Friendly used in Frigaliament to figure out the
meaning of chicken.143 What it changes is whether judicial deductions
See Learned Hand, Historical and Practical Considerations Regarding Expert Testimony, 15
HARV. L. REV. 40, 53–54 (1901) (objecting that to adversarial expert presentation); see also
sources cited infra notes 140–142 (documenting adversarial bias among party-retained
experts and courts’ tools for policing it).
See Daniel L. Rubinfeld, Econometrics in the Courtroom, 85 COLUM. L. REV. 1048, 1070–75
(1985) (proposing to “require experts to report, as a standard practice, … the sensitivity of
the results” to alternative specifications).
See Fed. R. Evid. 706; Joe S. Cecil & Thomas E. Willging, Accepting Daubert’s Invitation:
Defining a Role for Court-Appointed Experts in Assessing Scientific Validity, 43 EMORY L.J. 995,
998 (1994) (endorsing appointment of a court’s expert where “an independent source of
information is necessary for a principled resolution of a conflict”).
See generally David E. Bernstein, Expert Witnesses, Adversarial Bias, and the (Partial) Failure
of the Daubert Revolution, 93 IOWA L. REV. 451 (2008) (arguing the Daubert revolution has
only partly succeeded on its own terms and that adversarial expert presentation introduces
conscious, unconscious, and selection biases).
Frigaliment Importing Co. v. B.N.S. Int’l Sales Corp., 190 F. Supp. 116 (S.D.N.Y. 1960).

are performed in the open, on a contestable record, with reasons that
can be examined and challenged, or, rather in the mind of a single
decisionmaker drawing on intuitions she cannot fully articulate, against
a record of selective argument from the parties.
In fact, because we see contestation as part of the recipe, the
opposing worry looms: generative gap filling will be just as costly and
indeterminate as battles of experts or dueling dictionaries. But we do
not think that such full-fledged model battles would be common,
because our data suggests the method would prove stable enough in
run-of-the-mill cases. And just as the tool resolves some litigation
questions, it will also reduce the propensity to litigate on familiar
litigation-selection grounds.144
We might even suggest some rules of the road for those who are
excited about the promise of the method but worried about its
pathologies.
To start, the proponent of a model output must disclose the model,
its version, the prompt, and any system instructions and run settings.
Essentially, all that is needed to replicate the method of query. This is
the natural successor to the disclosure proposals our prior work
sketched, though tuned to gap-filling rather than dispute
interpretation.145 The asymmetry that helps here is that the opposing
party has access to the same surrounding document and can run a
competing query, which makes the disclosure regime self-enforcing in
a way that prompt disclosure for general litigation is not.
Second, the proponent must specify information about the model
“harness,” the term of art for the model’s mode of deployment and
access to tools. A model that has no access to the web and that had
finished training in some past date will produce different results than a
model that has access to the internet (which may include details about
the parties and their dispute), private transaction data, or past chat
details.
See George L. Priest & Benjamin Klein, The Selection of Disputes for Litigation, 13 J. LEGAL
STUD. 1, 4–5 (1984). Note that we make a partial equilibrium argument; the general
equilibrium of litigation and the selection of disputes in the age of AI requires separate
analysis.
See Arbel & Hoffman, supra note 18, at 509.

$#

Third, sanctions for fabricated outputs must be severe and visible,
because the integrity of the entire mechanism depends on it. Recent
episodes of lawyers submitting fake citations are a preview of what not
doing this looks like.146 We use fabricated rather than hallucinated for a
particular reason. There are subtle ways to steer models towards
desired ends, and lawyer ingenuity knows no end. Litigants should be
wary of risks and have the benefit of judicial penalties when opposing
parties manipulate the tools in underhanded ways – much like the case
of bribing an expert.
Fourth, courts must retain a robust gatekeeping role for the
threshold question of whether model output is probative at all given the
kind of gap at issue. This point is one we develop across the rest of Part
III.
B. Choice of Model Clauses & Equilibrium Drafting Effects
If courts treat AI output as evidence about textual inference,
contestable through the ordinary adversarial machinery, then it’s fair
to expect that sophisticated parties will start drafting around it. That
would fit the pattern of other deal terms generated in response to
publicly provisioned rules about interpretation and adjudication. Thus,
merger clauses cabin the unwieldy effects of the parol evidence rule,147
choice of forum clauses locate disputes despite background rules
privileging plaintiff choice,148 while choice-of-law clauses respond to
conflict-of-laws doctrine.149
The new contractual technology deserves a name. Choice of Model
clauses would be provisions specifying the terms of engagement for
See generally Matthew Dahl, Varun Magesh, Mirac Suzgun & Daniel E. Ho, Large Legal
Fictions: Profiling Legal Hallucinations in Large Language Models, 16 J. LEGAL ANALYSIS 64
(2024) (finding hallucination rates between 58% and 88% when models are asked verifiable
questions about federal court cases).
Eric A. Posner, The Parol Evidence Rule, the Plain Meaning Rule, and the Principles of
Contractual Interpretation, 146 U. PA. L. REV. 533, 537 (1998) (explaining that “parties can, in
principle, contract out of the parol evidence rule by including a ‘merger’ or ‘integration’
clause”).
John F. Coyle, Interpreting Forum Selection Clauses, 104 IOWA L. REV. 1791, 1791 (2019)
(examining forum selection clauses, whose construction determines “whether litigation
arising out of a particular contract must proceed in a given place”).
See generally John F. Coyle, A Short History of the Choice-of-Law Clause, 91 U. COLO. L. REV.
1147 (2020) (tracing how contracting parties developed choice-of-law clauses as a privateordering response to conflicts doctrine).

judicial integration of generative interpretation and gap filling tools.
Those include which model, which harness, with which prompt
protocol, governs disputes about contractual silence. 150 An earlygeneration version might read, simply:
Any dispute concerning the meaning or effect of an
omitted or ambiguous term in this Agreement shall be
resolved in the first instance by reference to inferences
drawn by [Model X, version Y or any successor],
applied to this Agreement under [a specified protocol].
But there’s plenty of room for more sophisticated variants. Parties
might designate a panel of models with weighting rules, require
sensitivity analyses, or carve out particular categories of gap for
submission to a court rather than a model. Our own data counsel
modesty about what a panel buys: because the models’ errors today are
strongly correlated, a panel of models purchases contestability and
replication rather than independent verification. A clause naming a
panel should therefore specify an aggregation rule and a tie or
abstention rule—our strict-majority panel returned no answer on eight
of the 119 contracts—and should treat model divergence as a trigger for
ordinary adversarial scrutiny rather than treating unanimity as a safe
harbor. As this area of practice develops, we imagine that parties will
also pair choice of model clauses with arbitral ones.151
Choice of Model clauses are not exactly like choice-of-law or
merger clauses. They select an interpretive method, and it’s not obvious
that all courts would accept their force.152 But to get a sense of how it
We floated the possibility in Generative Interpretation, supra note 18, at 455, but did not name
the choice of model clause itself.
The AAA Legal Context Protocol might include this kind of language. Legal Context
Protocol, https://legalcontextprotocol.org (last visited July 2, 2026).
We build on Shawn Bayern’s argument that parties have preferences about which
interpretive regime—textualism, contextualism, some combination—governs their
contract, and that contract law should treat the selection of an interpretive regime as a
question prior to substantive interpretation. See Shawn J. Bayern, Contract MetaInterpretation, 49 U.C. DAVIS L. REV. 1097 (2016); cf. Yair Listokin, Bayesian Contractual
Interpretation, 39 J. LEGAL STUD. 359 (2010) (proposing Bayesian inference as a framework
for contract interpretation, but without the party-selection element); but cf. Charny, supra
note 12, at 1819 (“Most fundamentally, no text can completely specify its own means of
interpretation. A contractual statement that purported to be such a complete specification
would itself have to be interpreted by some set of rules of interpretation.”).

$$

would work, we’d urge courts to consider an analogical practice: the
incorporation-by-reference of technical standards.
Parties have been incorporating technical resources by reference
for as long as there have been technical industries, and for just as long
courts have given those choices deference.153 In derivative contracts,
ISDA Definitions govern the meaning of terms. 154 In construction
contracts, the AIA conventions reign. 155 ISO forms govern terms in
commercial insurance.156 Class-action settlements pin diagnostic terms
to particular editions of the DSM. 157 Some contracts even specify
which dictionary controls disputed words.158 Choice of Model clauses
would extend this interpretative control to how courts employ
generative AI.159
Accepting a choice of model clause may create a versioning issue:
what if ChatGPT 6 is replaced by ChatGPT 7, and the original model
(though picked by the parties) is no longer available. The default rule
for incorporation-by-reference of an undated resource is that the
See Bibb v. Allen, 149 U.S. 481, 491–92 (1893) (holding that the rules and regulations of the
New York Cotton Exchange “enter into and form part of” contracts for future delivery of
cotton executed on that exchange); see also Lisa Bernstein, Private Commercial Law in the
Cotton Industry: Creating Cooperation Through Rules, Norms, and Institutions, 99 MICH. L. REV.
1724, 1724–30 (2001) (describing the cotton industry’s private commercial-law system of
trade rules, merchant tribunals, and industry norms); see generally Emily S. Bremer,
Incorporation by Reference in an Open-Government Age, 36 HARV. J.L. & PUB. POL’Y 131 (2013)
(surveying the parallel regulatory practice of giving privately drafted technical standards
legal force by incorporating them by reference into federal regulations).
See generally M. Konrad Borowicz, Contracts as Regulation: The ISDA Master Agreement, 16
Cap. Mkts. L.J. 72, 72 (2021) (discussing the ISDA).
Kevin E. Davis, The Role of Nonprofits in the Production of Boilerplate, 104 MICH. L. REV. 1075,
1078–79 (2006) (discussing the AIA’s dominance).
See Kenneth S. Abraham, The Legal Architecture of Insurance, MICH. ST. L. REV. (forthcoming
2026).
Special Master Ruling on Alzheimer’s Disease Diagnostic Criteria, In re Nat’l Football
League Players’ Concussion Injury Litig., No. 2∶12-md-02323-AB (E.D. Pa. Mar. 10, 2025),
https://www.nflconcussionsettlement.com/ViewDoc.aspx?dp=alz_diagnosis_criteria_sm.
pdf.
See Bayern, supra note 152, at 1104 (treating the choice among “relying on a dictionary, or
admitting trade usage or course of dealing” as a meta-interpretive question the parties can
resolve); see, e.g., Agreement § 9.6, Olaregen Therapeutix Inc., Ex. 10.5 (Apr. 2021)
(providing that words not defined in the agreement “shall have the meaning found in
Merriam-Webster’s
Dictionary”),
https://www.sec.gov/Archives/edgar/data/772320/000149315221009535/ex10-5.htm.
Cf. Alan Schwartz & Robert E. Scott, Contract Interpretation Redux, 119 YALE L.J. 926, 941
(2010) (arguing that courts should obey party interpretive instructions).

contract picks up subsequent revisions, particularly where parties
drafting in a domain with regularly-updated standards are presumed to
know that standards update.160 The same rule should apply to Choice
of Model clauses. A reference to “Model X” without version
specification ought to pick up subsequent versions. Parties who want
the model frozen at signing must say so explicitly.
So far, so good: Choice of Model clauses are not an innovation in
kind. The more interesting problem comes next.
Once parties can anticipate that courts will enforce choice of model
clause, the rational move will be to run the contract through the
selected model before signing. Imagine, then, that both sides run the
model and both jointly see what it predicts about silences, particularly
when the AI itself generates contingencies to test against. This
combination of a choice of model clause and pre-testing has
unexpected consequences for resolution of contractual disputes.161
Recall the taxonomy from Part I. We distinguished three sources of
silence: strategic disagreement, strategic non-drafting, and pure
inadvertence. Current doctrine generally treats them alike because
courts have no principled way to tease them apart.
Now, under a legal rule that enforces Choice of Model clauses, the
third category dissipates for more sophisticated parties, because
inadvertently leaving a gap is less likely once both parties have run the
model and seen what each silence will be taken to mean. Contracts will
always be incomplete, but the degree of incompleteness will fall,
perhaps precipitously. Leftover silences in Choice of Model contracts
are more likely than before to be deliberate. But it turns out that there
are two entirely different versions of what they imply.

See Constellation Power Source, Inc. v. Select Energy, Inc., 467 F. Supp. 2d 187, 205–08
(D. Conn. 2006) (applying New York law and holding that a contract’s reference to industry
rules encompassed subsequent revisions to those rules where the parties were aware that
revisions would occur and used unqualified, catch-all language); but cf. KENNETH A. ADAMS,
A MANUAL OF STYLE FOR CONTRACT DRAFTING (absent “as in effect at any given time,” a
reference captures the version existing at signing).
For a gesture toward this argument, see Omri Ben-Shahar, Towards the End of Normative
Interpretation of Contracts, JOTWELL (Nov. 2, 2023), https://contracts.jotwell.com/towardsthe-end-of-normative-interpretation-of-contracts/.

$%

Perhaps it’s endorsement. The parties ran the model, liked what it
predicted, and chose not to speak. In so doing, the silence acts like an
incorporation by reference.162 The autonomy-based defense of model
inference is at its strongest here, as the model is not imposing a
hypothetical bargain. Rather, the parties have adopted the model’s
inference their own.
Or maybe it’s strategy. The parties ran the model. At least one of
them did not like the prediction. But raising the issue at the bargaining
table would have surfaced a disagreement the parties could not resolve,
or could resolve only at a cost greater than the deal was worth. The
disfavored party still wishes to sign the agreement, perhaps hoping that
contingency won’t come to pass, but they do not agree to the model’s
interpretation.
Endorsement and strategy have doctrinally distinct implications.
An endorsed silence invites enforcement.163 A strategic silence on the
other hand requires resort to theory. One may take the view that, just
as in the case of legal defaults that a party may dislike but still not
override in their contract, even thin consent is consent. But others will
disagree, turning to the doctrines we developed precisely to handle
non-convergence: the implied covenant of good faith,
unconscionability, penalty default, and older equity-derived tools that
retain force when formal consent overstates actual agreement.164
Specifying which doctrine applies ought to depend on the deal
type, the parties, and the substantive area. The structural point is that
under a choice of model regime, courts are going to have to focus their

11 Samuel Williston & Richard A. Lord, A TREATISE ON THE LAW OF CONTRACTS § 30∶25, at
234 (4th ed. 1999) (“[T]o uphold the validity of terms incorporated by reference, it must be
clear that the parties to the agreement had knowledge of and assented to the incorporated
terms”).
See, e.g., Barnett, supra note 33 (silence against a known background default rule can itself
constitute consent to that default, so enforcing it rests on the parties’ own choice).
On deliberately incomplete agreements, see Ben-Shahar, supra note 26, at 404 (parties who
leave a gap because they cannot resolve a disagreement may merit special pro-defendant
defaults rather than ordinary majoritarian ones); cf. Omri Ben-Shahar, Contracts Without
Consent: Exploring a New Basis for Contractual Liability, 152 U. PA. L. REV. 1829, 1836 (2004)
(proposing a no-retraction principle and complicating any claim that strategic silence is
simply non-agreement).

inquiry on what to do about silence in a different way than they did
previously.
The evidentiary apparatus for doing so is the same as what we
developed earlier in this Section, pushed back to the drafting stage. A
proponent of model inference at litigation can show that the same
model was used in drafting, that the prediction was bilaterally
accessible, that the silence was not the subject of a failed negotiation.165
Discovery into drafting practice—what models were used, what they
predicted, what the negotiation record shows about whether the
silence was raised and abandoned—can and probably should be used to
diagnose which kind of silence is in front of the court.166
At some partial equilibrium, the choice of models should not
expand or contract the room that courts have to interpret normatively.
That hard judicial work is irreducible. Choice of models gives judges a
different evidentiary base, applied to a slightly different set of
questions. The hypothetical bargain question from gap filling gets an
estimated answer from the model, which may (or may not) reflect the
actual bargain. What remains is much harder: should we care about that
output, do we have overriding public policy concerns, did these parties
have access to the model’s prediction, and what should we make of
their silence given that access?
There are broader normative questions hovering over all of this that
we flag without resolving. If Choice of Model clauses become standard,
the choice of model becomes a non-trivial economic and political fact.
Which models dominate? Which providers? With access to what data?
Who has access, and at what price? 167 These are antitrust and
administrative-law problems, and ones that test the public-private
Cf. Ben-Shahar, supra note 26, at 413 (discussing the evidentiary nature of the question of
why there are silences).
Such discovery would run into the parol evidence rule, with its various exceptions.
Restatement (Second) of Contracts § 214(c) (Am. L. Inst. 1981); cf. Posner, supra note 147
(offering an economic framework for when courts should look beyond the writing, including
to the parties’ negotiations, in interpreting contracts).
On the foundation-model market’s tendency toward concentration with the attendant stakes
for access and price, see Jai Vipra & Anton Korinek, Market Concentration Implications of
Foundation Models: The Invisible Hand of ChatGPT (Brookings Ctr. on Regul. & Mkts.,
Working Paper No. 9, 2023) (high fixed training costs and near-zero marginal costs push
toward a few dominant providers, potentially warranting public-utility-style oversight).

$&

boundary in the production of law.168 The legal community would do
well to begin thinking now about the institutional design questions that
history will pose.
C. Generative Gap Filling in the Chambers
We’ve described a method of using AI centered on parties and their
lawyers. But AI is also available to judges acting on their own. We have
already seen judges experimenting with generative interpretation,
querying models for the ordinary meaning of contested terms,169 and
even for what common knowledge holds about the world. 170 Is this
desirable?
We’d distinguish three cases.
In the first, the judge is using a model as a language resource, i.e.,
generative interpretation.171 A judge who asks how ordinary speakers
use “landscaping” or “physically restrained” is doing something
structurally similar to opening up the dictionary on her shelf: the
inquiry is general rather than case specific, a probe of how a linguistic
community talks rather than of what these parties agreed. This looks
quite a bit like discussions about use of corpus linguistics and probably
falls out similarly: critics argue that the method is too novel, complex,
or opaque, and defenders highlight the gains from accuracy and
defending sua sponte use by the judge. 172 At least, we’d urge an AIcurious judge to be candid and transparent about the tool, the query,
See generally Tejas N. Narechania & Ganesh Sitaraman, An Antimonopoly Approach to
Governing Artificial Intelligence, 43 YALE L. & POL’Y REV. 95, 128–43 (2024) (raising NeoBrandesian antitrust concerns with AI); but see Simon Goldstein & Peter N. Salib, AI Is Not
a Natural Monopoly, 110 MINN. L. REV. HEADNOTES 121 (2026) (responding directly to
Narechania and Sitaraman, and arguing that fast-following dynamics undercut the naturalmonopoly diagnosis and that antimonopoly interventions could paradoxically raise prices
and reduce quality).
See cases cited at supra note 28.
See Ross, supra note 28, at 229 n.2, 236 n.4 (majority and dissent trading competing ChatGPT
queries on whether the danger of a hot car is common knowledge); id. at 229–31 (Howard,
J., concurring) (weighing the institutional risks of judicial AI use).
See generally Arbel & Hoffman, supra note 18, at 509 (arguing for transparency).
Compare State v. Rasabout, 2015 UT 72, 356 P.3d 1258 (criticizing, as unfair to the parties, a
concurrence’s sua sponte corpus analysis), with id. (Lee, A.C.J., concurring in part and
concurring in the judgment) (defending corpus methods as ordinary-meaning research), and
Wilson v. Safelite Grp., Inc., 930 F.3d 429 (6th Cir. 2019) (Thapar, J., concurring in part and
in the judgment) (urging corpus linguistics as an additional interpretive tool).

and the weight it received, so that the method can be contested on
appeal.
The second case involves inferring intent in the face of silence, i.e.,
generative gap filling. Here, the judge is going beyond establishing
general facts about the world into generating case specific evidence. A
judge who runs the contract through a model in chambers has, in
substance, commissioned an expert report that no party retained or can
cross-examine. The only way forward in such cases is for the judge to
disclose to the parties, ahead of time, their intended use of the model,
including model type, version, prompt, and settings; or route the task
through a neutral expert, as we suggested above.173
Finally, some judges or arbitrators will naturally try to push the
boundaries and use models to draft their decisions. 174 While we
appreciate that some may use the experiments we described here to
support such practices, this is decidedly not what we recommend nor
would that practice be supported by our evidence.
D. How Far Does This Go?
We’ve described some places where generative gap filling makes
most sense: commercial contracts where parties have the opportunity
to bargain for a Choice of Model and can amortize drafting costs across
many deals. But what about consumer contracts, or deals between oneoff players like small businesses?
Let’s start with consumer contracts. It would be natural to think
that using AI for these deals is particularly problematic where we are
fairly sure that no adherent read any of the text at all.175 “What would
See Fed. R. Evid. 201(e) (guaranteeing, on timely request, an opportunity to be heard on the
propriety of taking judicial notice); cf. Fed. R. Evid. 706.
See Petition to Vacate Arbitration Award, LaPaglia v. Valve Corp., No. 3∶25-cv-00833 (S.D.
Cal. Apr. 8, 2025) (alleging that the arbitrator exceeded his powers under 9 U.S.C. § 10(a)(4)
by “outsourcing his adjudicative role to artificial intelligence”—having ChatGPT ghostwrite
portions of the award); see also Order Granting Motion to Dismiss, LaPaglia v. Valve Corp.,
No. 3∶25-cv-00833 (S.D. Cal. Dec. 9, 2025) (dismissing the petition on jurisdictional grounds
without reaching the AI allegations).
See Yannis Bakos, Florencia Marotta-Wurgler & David R. Trossen, Does Anyone Read the
Fine Print? Consumer Attention to Standard-Form Contracts, 43 J. LEGAL STUD. 1, 1 (2014)
(tracking 48,154 online software shoppers and finding that roughly one or two in a thousand
access the standard-form license, and that those who do read only a small portion); David A.

)!

the parties have bargained for” is an odd question to ask about a
contract created without bargaining or reading! And of course
consumer contracts are exactly the sort of deals where background
questions about fairness, distribution of resources, and the thinness of
consent sap the case for intent-based interpretation.
And yet we’d note that extrapolating intent from text using a large
language model is not obviously worse than its competitor: deriving
consumers’ “reasonable expectations” using some mix of hunches and
surveys. That is after all the approach adopted by the new Restatement
of Contracts’ section on interpretation, and which presumably also
applies to gap filling.176 The problem is that consumers’ expectations
about what contracts will contain are shaped by the contracts they are
in.177 If you ask them what’s reasonable to fill a gap, it’s unlikely that
they’ll report a term that protects them. And the process of generating
those inferences is going to be an expensive survey, further tilting
outcomes away from consumers.178
Generative gap filling may produce gap-filling answers that benefit
firms. But it does so at lower cost than a survey, and in a way that is
more easily contestable by consumers in litigation. And, crucially,
courts will be able to ground their decisions in the text of the contracts

Hoffman, Defeating the Empire of Forms, 109 Va. L. Rev. 1367 (2023) (arguing that evercheaper, unread forms now blanket even the lowest-stakes transactions and proposing to
make many such forms unenforceable); cf. Yonathan A. Arbel, The Readability of Contracts:
Big Data Analysis, 21 J. EMPIRICAL LEGAL STUD. 927 (2024) (finding, from a corpus of two
million agreements, that consumer contracts’ median reading scores—a dubious but
common metric—approximate those of daily news articles).
Restatement of the Law, Consumer Contracts § 4(d) (Am. L. Inst. 2024) (“[S]tandard
contract terms are interpreted in a manner that effectuates the reasonable expectations of the
consumer.”); see id. § 4 cmt. 5 (directing an empirical inquiry into “the ordinary behavior
and perspective of consumers engaged in the type of transaction at issue”).
See David A. Hoffman, Consumers’ Unreasonable Textual Expectations, 15 HARV. BUS. L. REV.
43, 51, 56–58 (2025) (marshaling evidence that consumers form their interpretive
expectations from their own experience with firms’ contracts); see also generally Tess
Wilkinson-Ryan et al., supra note 40 (providing evidence from large scale survey of
Americans of differential experiences with contract and offering a follow-up experiment
suggesting that those experiences are constitutive of judgments).
Cf. Hoffman, supra note 177, at 49 (surveying consumers “may be a solution that is too
expensive and uncertain to get traction in courts”); Ben-Shahar & Strahilevitz, supra note 14
(proposing the survey method).

themselves, and the majoritarian inferences that the forms’ provisions
generate.
A different, and harder, problem comes when considering one-off
deals where the parties are not a part of majoritarian linguistic
communities. Consider, for example, a neophyte painting contractor
entering its first substantial construction subcontract,179 a new Hawaii
carpeting subcontractor unfamiliar with local trade usage, 180 a Swiss
buyer and New York seller disputing whether “chicken” means
broilers or also includes stewing fowl,181 or an Orthodox cantor and a
Miami Beach hotel disputing whether a Passover Seder engagement,
against the background of differing Orthodox and Reform practices,
obligated the hotel to hold a second Seder.182
Here, there is a real danger that generative gap filling will simply
get it wrong, because its textual orientation will miss important social
context. Although it is possible to develop models calibrated to
linguistic subcommunities, doing so adds complexity and expense, and
at least at first glance would be challenging for busy trial court judges.
If judges must hire OpenAI coders to appropriately resolve a summary
judgment motion, the method is of no use at all.
To put it differently—and in normative terms—when there’s good
reason to think that parties’ contracts are singular or bespoke, this
should lead us to be less interested in ordinary meaning textualism, and
in technological methods that make textualism more accurate. The
parties to such deals would have to tell us that they really do want to
use AI—through a Choice of Model Clause—before it would be
appropriate to resort to it.
E. Gaps in Generative Gap Filling and the Limits of the Method
We recognize several important limitations to the discussion
above.

Flower City Painting Contractors, Inc. v. Gumina Constr. Co., 591 F.2d 162 (2d Cir. 1979).
United States ex rel. Union Bldg. Materials Corp. v. Haas & Haynie Corp., 577 F.2d 568 (9th
Cir. 1978).
Frigaliment Importing Co. v. B.N.S. Int’l Sales Corp., 190 F. Supp. 116 (S.D.N.Y. 1960).
Tucker v. Forty-Five Twenty-Five, Inc., 199 So. 2d 522, 523–24 (Fla. Dist. Ct. App. 1967).

)#

We start with a cluster of concerns about model reliability and
bias. 183 Beyond the familiar concern with hallucinations, researchers
worry about issues of prompt sensitivity, sycophancy, and randomness
in outputs.184 A core worry is that using these tools will lead judges and
lawyers to be overconfident that they are right, crowding out norms of
humility that would otherwise make legal decisions more sociologically
legitimate. 185 And that overconfidence would rest on a tool that is
deceptively plausible. As every regular user knows, generative AI is
wrong some of the time, but weirdly so, in ways that are challenging to
identify through the swamp of sycophantic, plausible, feedback that it
sends a reader’s way. We agree that this is a real problem, particularly
for busy and hubristic judges.
That said, at least some of this concern is reducible to operational
protocols and the adversarial posture of the system. Courts need
See, e.g., Grimmelmann et al., supra note 19; Jonathan Scher, Note, Beyond Words: The Risks
of Generative Interpretation, 99 S. CAL. L. REV. Postscript 64 (2026) (use of GenAI creates
risk of overconfidence); Zachary Catanzaro, The Dead Law Theory: The Perils of Simulated
Interpretation,
FLA.
L.
REV.
(forthcoming
2027)
(manuscript),
https://ssrn.com/abstract=6164388 (GenAI cannot provide semantic meaning); Abhishek
Purushothama, Junghyun Min, Brandon Waldon & Nathan Schneider, Prompting from the
Bench: Large-Scale Pretraining Is Not Sufficient to Prepare LLMs for Ordinary Meaning Analysis,
arXiv:2510.25356 (2025) (forthcoming, 2026 ACM Conf. on Fairness, Accountability &
Transparency) (prompts are not robust); Susan Tanner, Prediction, Indeterminacy, and the
Architecture of Legal Meaning in the Age of Generative AI (Apr. 2, 2026) (unpublished
manuscript), https://ssrn.com/abstract=6511563 (interpretation requires normative
judgment); Frank Pasquale, The Non-Delegable Duty to Think: Judicial Legitimacy and the
Limits of Generative AI, 74 UCLA L. REV. (forthcoming 2026) (Cornell Legal Studies Rsch.
Paper No. 26-03) (arguing that humans are likely to be better, less fragile, and more just
interpreters).
On prompt sensitivity, see Choi, supra note 19; Purushothama et al., supra note 183; Melanie
Sclar et al., Quantifying Language Models’ Sensitivity to Spurious Features in Prompt Design or:
How I Learned to Start Worrying About Prompt Formatting (arXiv, Working Paper No.
2310.11324, 2024), https://arxiv.org/abs/2310.11324 (finding performance swings of up to
seventy-six accuracy points from semantically irrelevant formatting changes). On
sycophancy, see Mrinank Sharma et al., Towards Understanding Sycophancy in Language
Models (arXiv, Working Paper No. 2310.13548, 2023), https://arxiv.org/abs/2310.13548
(finding that five state-of-the-art assistants “consistently exhibit sycophancy” and tracing
the behavior to human-feedback training). On randomness, see Berk Atil et al., NonDeterminism of “Deterministic” LLM Settings (arXiv, Working Paper No. 2408.04667, 2024),
https://arxiv.org/abs/2408.04667 (observing accuracy variation of up to fifteen percent
across repeated runs of identical prompts under settings expected to be deterministic).
See generally Dan M. Kahan, Foreword: Neutral Principles, Motivated Cognition, and Some
Problems for Constitutional Law, 125 HARV. L. REV. 1, 62 (2011) (introducing and defending
expressions of complexity and engagement in opinion writing).

training on evaluating model outputs. And more generally, bias and
error are relative facts. There are decades of empirical research on how
human judgment varies based on “hidden parameters” such as mental
energy levels, mood, ambient temperature, time of day, and the
sequence of cases just decided. 186 Humans are, well, human
decisionmakers.187 It’s not obvious that using AI adds more error or
variance to an already imperfect system.
General reliability is also contested. The leading skeptical work in
this area shows that off-the-shelf models, untuned and uncalibrated,
given truncated legal scenarios and forbidden to reason, sometimes
produce large spreads over their judgments of somewhat similar
wording of the questions.188 But these differences were smaller than
what that same paper finds for human respondents.189 So it’s true that

See, e.g., Shai Danziger, Jonathan Levav & Liora Avnaim-Pesso, Extraneous Factors in
Judicial Decisions, 108 Proc. Nat’l Acad. Sci. 6889 (2011) (finding that the share of favorable
parole rulings falls steadily over a decision session and rebounds after judges’ meal breaks);
Ozkan Eren & Naci Mocan, Emotional Judges and Unlucky Juveniles, 10 AM. ECON. J.:
APPLIED ECON. 171 (2018) (finding that upset losses by the Louisiana State University
football team increased the sentences judges imposed on juveniles in the following week);
but see Keren Weinshall-Margel & John Shapard, Overlooked Factors in the Analysis of Parole
Decisions, 108 PROC. NAT’L ACAD. SCI. E833 (2011) (attributing much of the meal-break
pattern to nonrandom case ordering). See generally DANIEL KAHNEMAN, OLIVIER SIBONY &
CASS R. SUNSTEIN, NOISE: A FLAW IN HUMAN JUDGMENt (2021) (surveying unwanted
variability in professional judgment, including judging).
See generally Jeffrey J. Rachlinski & Andrew J. Wistrich, Judging the Judiciary by the Numbers:
Empirical Research on Judges, 13 ANN. REV. L. & SOC. SCI. 203 (2017) (reviewing the empirical
literature on extralegal influences on judicial decisionmaking).
See Choi, supra note 19, at 16–19 tbls.2–3 & figs.1–2. Choi reports dispersion but not the
legally operative statistic: how often the bottom-line verdict flips. By our analysis of his
reported numbers, GPT-4.1’s verdict flips on at most 1.8% and 3.0% of 2,000 rephrasings in
two of his five scenarios, in 2–9% in a third, and approaches a coin flip only in the remaining
two (28–52% and 31–47%) — both drawn, as all five are, from the most contested questions
appellate litigation produces. Computation on file with authors.
Models and human responders, in one part of his study, were asked to evaluate the ordinary
meaning of various scissor statements (is a taco a sandwich). When scored for dispersion,
the typical individual respondent misses the crowd mean by 21–23 points, which is above
GPT-4.1 (19.7) and on par with Claude Opus 4.1 (22.9). See Choi, supra note 19, at 26 tbl.4;
the dispersion figure for individual respondents reflects our own calculations from Choi’s
replication data, on file with authors. Other work has validated that models can replicate
ordinary judgments quite accurately (Kruse, supra note 74), and that models have
internalized latent reasoning schemas that ordinary people use to make legal judgment
(Arbel, The Generative Reasonable Person, supra note 74).

)$

the models have problems, but as always it’s important to ask compared
to what.
We too show that levels of inter-human disagreement are much
higher than those between models. 190 We offer direct evidence that
model results are convergent, despite using different model families,
model settings, and prompt variations. This finding coheres with
evidence from a variety of recent papers that shows that AI models can
accurately simulate human answers to legal tasks. 191 Ultimately, we
would frame even the best skeptical work as demonstrating a point
about relative legibility. Model uncertainty can be measured. 192 And
because it can be measured, legal actors can choose when to use and
when to be skeptical.

See Choi, supra note 19, at 26 tbl.4 (reporting deviations from the mean human response of
19.7 to 24.1 percentage points across GPT-4.1, Claude Opus 4.1, and Gemini 2.5 Pro, against
the 21–23-point deviation of the typical individual human respondent, per our calculations
from Choi’s replication data, on file with authors).
Johannes Kruse was even able to accurately simulate the answers provided by 2,835 human
respondents regarding ordinary-meaning. Kruse, supra note 74. Importantly, the
performance on this benchmark is heavily correlated with the overall performance of the
underlying model. See Guha et al., supra note 82 (introducing the LegalBench benchmark);
Vals AI, LegalBench, https://www.vals.ai/benchmarks/legal_bench (last visited July 15,
2026) (live leaderboard of frontier-model performance on LegalBench tasks, on which
standings largely track the models’ general capabilities).
More work here is needed, as Choi used two different measures of confidence and they
sharply disagreed with each other, making it difficult to interpret his results. E.g., in his
Scenario 3, the statistics imply the token method flips at least 28% of the time while the
confidence method flips at most 3.5%, an order of magnitude difference about the confidence
for the same question. Choi, supra note 19, at 16 tbl.2, 18 tbl.3. In addition, he focuses on firsttoken probabilities, but those are known to diverge from the written answers, see Xinpeng
Wang, Bolei Ma, Chengzhi Hu, Leon Weber-Genzel, Paul Röttger, Frauke Kreuter, Dirk
Hovy & Barbara Plank, “My Answer Is C”: First-Token Probabilities Do Not Match Text
Answers in Instruction-Tuned Language Models, in FINDINGS OF THE ASSOCIATION FOR
COMPUTATIONAL
LINGUISTICS:
ACL
2024,
at
(2024),
https://aclanthology.org/2024.findings-acl.441/ (finding mismatch rates above sixty percent
between first-token probabilities and models’ text answers); Ari Holtzman, Peter West,
Vered Shwartz, Yejin Choi & Luke Zettlemoyer, Surface Form Competition: Why the Highest
Probability Answer Isn’t Always Right, in PROCEEDINGS OF THE 2021 CONFERENCE ON
EMPIRICAL METHODS IN NATURAL LANGUAGE PROCESSING 7038 (2021),
https://aclanthology.org/2021.emnlp-main.564/ (showing that surface-form competition
makes the highest-probability answer an unreliable guide to model judgments).

A final cluster of concerns revolves around epistemics.193 Sure, we
can trust LLMs in domains where we can verify their answers like math
and coding, but why should we trust them in areas like interpretation
where there is no pre-agreed answer? At least with judges we can test
their reasoning. But with LLMs, the absence of a true cognitive process
means that the reasoning supplied and the operative reasons might be
entirely distinct.194
Our hope is that this Article provides a response. Rather than
relying on brute intuitions about language, dictionaries, surveys, or
previous judicial decisions as previous studies did, we select a legal
domain where there is consensus over what the goal of interpretation
is—the term the parties would write—and then identify the correct
answer to such questions. Our method allows us to evaluate whether
different legal actors and interpretative approaches can accurately
recover actual parties’ intent from surrounding context. It turns out
that they can.
Along the way, we offer a way to test the performance of, and
ultimately vindicate, lawyers’ skill. Are lawyers actually competent at
interpretation, or do they only project their own priors onto the page?
While legal professionals may give us reasons to support their
reasoning, many doubt those reasons just as much as AI-skeptics doubt
those provided by LLMs. 195 Having a method of resolving these
See Grimmelmann et al., supra note 19, at 280–82, 300 (arguing that “any attempt to calibrate
LLMs empirically depends on having some external benchmark to calibrate against,” and
that legal interpretation—unlike rote, verifiable tasks—offers no such ground truth); Waldon
et al., supra note 19, at 153 (demonstrating that LLMs’ metalinguistic judgments “are highly
sensitive to subtle prompting variations” and “can be easily ‘gamified’ to reflect a user’s
preconceived biases”); Lee & Egbert, supra note 75 (arguing that LLM outputs supply a form
of artificial intuition rather than transparent, replicable empirical evidence of ordinary
meaning).
See, e.g., Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison,
John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, Vlad Mikulik,
Samuel R. Bowman, Jan Leike, Jared Kaplan & Ethan Perez, Reasoning Models Don’t Always
Say What They Think (arXiv, Working Paper No. 2505.05410, 2025),
https://arxiv.org/abs/2505.05410 (finding that frontier reasoning models verbalize the hints
they actually relied on in fewer than twenty percent of cases in most settings).
See, e.g., JEROME FRANK, LAW AND THE MODERN MIND 100–01 (1930) (arguing that
“[j]udicial judgments, like other judgments, doubtless, in most cases, are worked out
backward from conclusions tentatively formulated”); see generally RICHARD A. POSNER, HOW
JUDGES THINK (2008) (contending that the legalist reasoning of opinions largely rationalizes
decisions reached on intuition, experience, and preconception).

)%

epistemological gaps, including showing whether lawyers are better
able to put ideological biases aside, would represent a real advance.

I V.

C ONCLUSION

When we’ve workshopped this Article’s dramatic main finding, a
question we’ve repeatedly gotten is existential in tone: what role can
humans possibly retain in the contract law of the future? Our readers
are lawyers and future ones, and we feel and half-share their despair
about our collective prospects. Like them, each time we hear a pundit
opine about the coming death of the knowledge-production economy,
we itch to grab some power-loom destroying tool and start smashing
the nearest data center. 196 And it’s fair enough to wonder about AIaligned proposals like the one we’ve offered: each individual generative
use case may seem defensible, scientific, rational, and efficient, but on
the whole we’re drifting to the bad place. 197
One answer—the one we have pressed throughout Part III—holds
onto optimism. Lawyers can put machines to work as our agents, not
our masters: disciplining contractual interpretation through Choice of
On gradual disempowerment, see Jan Kulveit, Raymond Douglas, Nora Ammann, Deger
Turan, David Krueger & David Duvenaud, Gradual Disempowerment: Systemic Existential
Risks from Incremental AI Development (arXiv, Working Paper No. 2501.16946, 2025),
https://arxiv.org/abs/2501.16946 (arguing that the incremental handoff of societal functions
to AI can erode human influence, competence, and control even absent any machine powerseeking); Richard M. Re & Alicia Solow-Niederman, Developing Artificially Intelligent Justice,
22 STAN. TECH. L. REV. 242, 275–78 (2019) (warning that AI adjudication risks “alienation,”
as humans “cease participating in the legal system and even lose interest in its operations”);
Pasquale, supra note 183 (arguing that core adjudicative judgment is a duty that cannot be
delegated to machines).
See, e.g., Chen, Stremitzer & Tobia, supra note 123 (examining whether adjudication by
“robot judges” can give litigants their day in court); Eric A. Posner & Shivam Saran, Judge
AI: A Case-Study of Large Language Models as Judges, 3 J.L. & EMPIRICAL ANALYSIS 179 (2026)
(evaluating GPT-4o as a substitute appellate decisionmaker and finding it more formalist
than the federal judges whose experiment it replicated). On the related worry that human
decisionmakers overtrust machine outputs, see Linda J. Skitka, Kathleen L. Mosier & Mark
Burdick, Does Automation Bias Decision-Making?, 51 INT’L J. HUM.-COMPUTER STUD. 991
(1999) (finding that reliance on automated aids produces errors of omission and
commission); Jennifer M. Logg, Julia A. Minson & Don A. Moore, Algorithm Appreciation:
People Prefer Algorithmic to Human Judgment, 151 ORG. BEHAV. & HUM. DECISION Processes
90 (2019) (finding that laypeople weight identical advice more heavily when they believe it
comes from an algorithm).

Model clauses, reducing AI pathologies through adversarial
presentation and disclosure, and writing opinions that are narrower,
more contestable, and more likely to give parties what they would have
wanted. This is an improvement on the status quo, as the next best
option for legal interpretation isn’t the platonic search for truth, but
rather unreflective recourse to dictionaries.198
Better still, because parties settle disputes when they can predict
outcomes, the likely result of more precision using AI is fewer
interpretation disputes. The ones that remain will concentrate on the
questions that were always the hard ones: whether to honor what the
parties would have said, not merely to determine what was intended.
That normative residue represents the irreducible core of the judicial
function. Because normative interpretation is hard to do within arbitral
tribunals, using AI will ironically buttress public, transparent, law.
But even such medium term optimism needs to grapple with the
long-term future of transactional practice.
Every argument in this Article rests on an assumption so basic that
we’ve ignored it for the last 24,000 words: that the contract was written
by people. Our central empirical fact—that the visible terms of an
agreement carry information about its hidden ones—is not a truth
about text as such. It’s a statement about how human bargains are
made. Contractual texts results from some social set of tradeoffs
occurring off the page. The mutual information we measured is the
residue of that process, a fossil record of the deal. That is why
predicting a masked term could serve as a proxy for recovering the
parties’ intent, and why our method could do something interpretation
scholarship has rarely managed: test interpreters against a ground
truth, the term the parties actually wrote.

See James J. Brudney & Lawrence Baum, Oasis or Mirage: The Supreme Court’s Thirst for
Dictionaries in the Rehnquist and Roberts Eras, 55 WM. & MARY L. REV. 483 (2013)
(documenting the Justices’ surge in dictionary reliance since the late 1980s); Jennifer L.
Mnookin, Scripting Expertise: The History of Handwriting Identification Evidence and the
Judicial Construction of Reliability, 87 VA. L. REV. 1723 (2001) (tracing how courts came to
credit forensic handwriting identification as reliable evidence).
)&

Contracts drafted and assembled by AI agents break both halves of
that foundation.199 Such documents may or may not continue to read as
contracts,200 but they result from a different process.201 While models
might be just as capable in filling gaps left by their brethren, they will
not be tracking intent anymore, at least not human intent.
If that’s where we’re headed, jurists lack good doctrinal vehicles to
manage the problems we’ll face. Every tool in the interpretive kit—
intent, assent, the hypothetical bargain, reasonable expectations,
contra proferentem—presupposes a human principal whose mental
states are the target of the inquiry. Agency law, the doctrine to which
courts will instinctively reach, attributes acts and authority.202 But it
has nothing to say about attributing clause-level meaning to a principal
who never read, drafted, or contemplated the clause.
Perhaps the answer is that intent migrates upstream: the
principals’ instructions become the operative expression of will, and
the prompt becomes the new parol evidence, controlled through
merger clauses that apply to models themselves.203 If so, the disclosureand-contestation apparatus we sketched in Part III.A provides an early
draft of the procedure such disputes will demand. But we flag the

On the increasing use of AI-agents in assembling contracts, see Bridget McCormack & David
Hoffman, Agentic Commerce Needs Legal Infrastructure, and The Courts Are Coming, American
Arbitration Association (Apr. 23, 2026), https://www.adr.org/news-and-insights/when-aiagents-transact-what-happens-next/.
See Jacob Andreas, Anca Dragan & Dan Klein, Translating Neuralese, in PROCEEDINGS OF
THE 55TH ANNUAL MEETING OF THE ASSOCIATION FOR COMPUTATIONAL LINGUISTICS 232,
232–33 (2017), https://aclanthology.org/P17-1022/ (calling agents’ automatically induced,
non-natural-language communication protocol “neuralese”).
Agentic contracts will drastically expand the “transaction frontier,” the scope of actions for
which bargaining is profitable. On the broader implications, see Yonathan A. Arbel, On the
Scales of Private Law: Nano Contracts, 37 HARV. J.L. & TECH. 151, 153–57 (2023) (explaining
that automated, near-zero-cost bargaining will open up a frontier of smaller transactions).
See Yonathan A. Arbel, Peter N. Salib & Simon Goldstein, How to Count AIs: Individuation
and Liability for AI Agents 3, 7–16 (Feb. 24, 2026) (unpublished manuscript),
https://arxiv.org/abs/2603.10028 (distinguishing thin identification that connects AI actions
to human principals from thick identification of AI agents as durable entities); Noam Kolt,
Governing AI Agents, 101 NOTRE DAME L. REV. (forthcoming 2026) (manuscript at 17–30),
https://ssrn.com/abstract=4772956 (using agency law and principal-agent theory to analyze
information asymmetry, authority, loyalty, and delegation problems presented by AI agents).
See David A. Hoffman, Cross-Examining Agentic Commercial Agents, Contracts’ Empire
( June 29, 2026), https://profhoffman.substack.com/p/the-contract-with-no-mind-behind.

continuity without claiming to have solved the problem. Someone will
need to build the doctrine, and they will need to start roughly now.
This does not trouble the project of this Article, because the stock
of human-drafted paper is titanic. Parties will be litigating human
agreements for a while, and those deals are precisely where generative
gap filling works and where our recommendations apply. But
something stranger looms.
This Article has treated contracts to radio signals, redundant
enough that a listener can rebuild what was lost from what came
through. That is why interpretation is possible at all. The metaphor
rests on a premise as important as the redundancy—someone sent the
message. For the contracts humans have written, the machines turn
out to be superb receivers, and the law should learn to use them. For
the contracts machines will write, the signal may come through
perfectly, and yet no one may ever have been on the other end. It is not
obvious why courts should keep listening.

%!