# Propositions from How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem

**Citation:** Yonathan A. Arbel & Shmuel I. Becher, How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem, in The Cambridge Handbook of Emerging Issues at the Intersection of Commercial Law and Technology 336, 336–374 (Stacy-Ann Elvy & Nancy S. Kim eds., 2025).

**Source:** [working-paper PDF of published book chapter](https://works.battleoftheforms.com/papers/ssrn-4491043/paper.pdf)

**Review status:** 68 model-drafted, source-checked; 0 human-reviewed. Page references use the printed pagination and, separately, the 1-based PDF page number.

## 1. The no-reading problem weakens informed consumer choice and sellers' incentives to offer fair and efficient standard terms

**Location:** Abstract and Introduction, printed pp. 1-3 (PDF pp. 1-3)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 1–3, that the no-reading problem weakens informed consumer choice and sellers' incentives to offer fair and efficient standard terms. When consumers manifest assent without reading or understanding boilerplate, firms receive little demand-side discipline over the legal content of their products. This is significant because nonreadership is not merely a defect in individual consent but a structural problem in the market for contract terms. It connects to standard-form contracts, consumer assent, market discipline, boilerplate, informed choice, contract quality.

**Evidence anchor:** Pages 1-3 state the no-reading problem and connect it to informed decisions and seller incentives.

**Boundary:** The paper acknowledges disputes over the scope and implications of nonreadership and does not claim that reading alone cures every defect in consent.

**Connections:** standard-form contracts; consumer assent; market discipline; boilerplate; informed choice; contract quality

**Record:** `ssrn-4491043-p01` · `machine-drafted-source-checked`

## 2. Evaluating present-generation smart readers remains worthwhile despite rapid model improvement because the immediate question is whether existing tools have crossed a practical utility threshold

**Location:** Introduction, printed pp. 3-4 (PDF pp. 3-4)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 3–4, that evaluating present-generation smart readers remains worthwhile despite rapid model improvement because the immediate question is whether existing tools have crossed a practical utility threshold. If tools available now can materially empower consumers, scholarship and policy designed around a technologically fixed no-reading problem must adjust now rather than await perfect systems. This is significant because fast technological change can make current capability, not only future potential, legally consequential. It connects to technology assessment, utility thresholds, legal policy, model progress, consumer empowerment, paradigm change.

**Evidence anchor:** Pages 3-4 defend testing current models and explain the policy relevance of a utility threshold.

**Boundary:** Results tied to a 2023 model set are a dated capability snapshot and should not be treated as a permanent ranking of systems.

**Connections:** technology assessment; utility thresholds; legal policy; model progress; consumer empowerment; paradigm change

**Record:** `ssrn-4491043-p02` · `machine-drafted-source-checked`

## 3. The tested smart readers greatly shortened and simplified consumer legal texts without generally sacrificing their essential information

**Location:** Abstract and Introduction, printed pp. 1-4 (PDF pp. 1-4)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 1–4, that the tested smart readers greatly shortened and simplified consumer legal texts without generally sacrificing their essential information. Across the sample, outputs reduced length by 66.9 percent, saved about fourteen minutes and forty-one seconds of reading, and sometimes transformed college-level prose into language readable at roughly the fifth-grade level. This is significant because LLM assistance can change the practical accessibility of contracts by orders large enough to matter outside laboratory benchmarks. It connects to contract simplification, reading time, readability, LLM evaluation, consumer access, empirical legal studies.

**Evidence anchor:** The abstract and introduction report headline reductions in length, reading time, and grade-level difficulty.

**Boundary:** Average results conceal substantial variation across models and clauses, and readability formulas do not directly measure comprehension.

**Connections:** contract simplification; reading time; readability; LLM evaluation; consumer access; empirical legal studies

**Record:** `ssrn-4491043-p03` · `machine-drafted-source-checked`

## 4. Smart readers are not substitutes for qualified lawyers but can still be effective where consumers' realistic alternative is no review at all

**Location:** Abstract and Introduction, printed pp. 1-4 (PDF pp. 1-4)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 1–4, that smart readers are not substitutes for qualified lawyers but can still be effective where consumers' realistic alternative is no review at all. Occasional omissions, misleading language, and misuse of legal terminology are disqualifying for professional legal advice but need not erase large gains in routine mass-market transactions. This is significant because the proper benchmark is the consumer's actual choice set, not flawless expert performance. It connects to comparative baseline, access to legal information, lawyer substitution, routine transactions, consumer welfare, error tolerance.

**Evidence anchor:** Pages 1 and 4 distinguish ordinary consumer use from lawyer replacement and identify the realistic baseline.

**Boundary:** The authors do not claim the tools are safe for high-stakes individualized legal advice or that all errors are benign.

**Connections:** comparative baseline; access to legal information; lawyer substitution; routine transactions; consumer welfare; error tolerance

**Record:** `ssrn-4491043-p04` · `machine-drafted-source-checked`

## 5. Consumer nonreadership has multiple causes, including rational apathy, behavioral bias, take-it-or-leave-it terms, trust, social norms, and expectations of judicial protection

**Location:** Part II: LLMs and the No-Reading Problem, printed pp. 4-5 (PDF pp. 4-5)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 4–5, that consumer nonreadership has multiple causes, including rational apathy, behavioral bias, take-it-or-leave-it terms, trust, social norms, and expectations of judicial protection. Dense drafting is important, but a complete account must recognize that reading may remain unattractive when negotiation is impossible or the expected benefit of attention is low. This is significant because simplification addresses only some mechanisms behind the no-reading problem and should not be mistaken for a universal cure. It connects to rational apathy, behavioral bias, information overload, adhesion contracts, social norms, judicial expectations.

**Evidence anchor:** Pages 4-5 review competing explanations for why consumers do not read form contracts.

**Boundary:** The chapter surveys explanations rather than testing their relative causal weight in its sample.

**Connections:** rational apathy; behavioral bias; information overload; adhesion contracts; social norms; judicial expectations

**Record:** `ssrn-4491043-p05` · `machine-drafted-source-checked`

## 6. Firms can pursue a HIDE strategy by making terms Hardly Interpretable but Dependably Enforceable

**Location:** Part II: LLMs and the No-Reading Problem, printed pp. 5 (PDF pp. 5)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 5, that firms can pursue a HIDE strategy by making terms Hardly Interpretable but Dependably Enforceable. When buyers do not read and courts enforce, sellers can claim more transactional surplus through obscure self-serving clauses without losing demand. This is significant because unreadability can be a profitable design choice rather than an accidental byproduct of legal complexity. It connects to HIDE strategy, strategic drafting, contract enforcement, transactional surplus, dark patterns, seller incentives.

**Evidence anchor:** Page 5 defines HIDE and explains how nonreadership and enforcement jointly benefit firms.

**Boundary:** The paper presents HIDE as a strategic possibility and does not empirically prove that every complex form was intentionally designed to conceal.

**Connections:** HIDE strategy; strategic drafting; contract enforcement; transactional surplus; dark patterns; seller incentives

**Record:** `ssrn-4491043-p06` · `machine-drafted-source-checked`

## 7. Traditional responses to unreadable contracts rely on duties to read, plain-language laws, disclosure formatting, and readability requirements

**Location:** Part II: LLMs and the No-Reading Problem, printed pp. 5-6 (PDF pp. 5-6)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 5–6, that traditional responses to unreadable contracts rely on duties to read, plain-language laws, disclosure formatting, and readability requirements. Courts, legislatures, and agencies attempt to make assent more meaningful through rules such as conspicuousness, controlled language, and mandated scores. This is significant because smart readers enter an already crowded policy field rather than addressing a problem untouched by law. It connects to duty to read, plain-language laws, UCC conspicuousness, Magnuson-Moss Act, Truth in Lending Act, CFPB.

**Evidence anchor:** Pages 5-6 catalog judicial, statutory, and administrative approaches to readability and disclosure.

**Boundary:** The survey does not evaluate every doctrine or jurisdiction and treats these tools at a high level.

**Connections:** duty to read; plain-language laws; UCC conspicuousness; Magnuson-Moss Act; Truth in Lending Act; CFPB

**Record:** `ssrn-4491043-p07` · `machine-drafted-source-checked`

## 8. One-size-fits-all plain-language regulation poorly matches the diversity of consumer literacy, cognition, language, experience, and visual capacity

**Location:** Part II: LLMs and the No-Reading Problem, printed pp. 6 (PDF pp. 6)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 6, that one-size-fits-all plain-language regulation poorly matches the diversity of consumer literacy, cognition, language, experience, and visual capacity. A single average readability target can miss those most in need of protection while imposing drafting costs and still failing to make legal concepts genuinely accessible. This is significant because effective communication should adapt to users rather than merely simplify text for an abstract reasonable consumer. It connects to consumer heterogeneity, personalization, literacy, accessibility, plain language, reasonable consumer.

**Evidence anchor:** Page 6 explains why uniform plain-language rules may not serve a heterogeneous consumer population.

**Boundary:** Personalization can create privacy, consistency, and differential-treatment risks that are outside the empirical test.

**Connections:** consumer heterogeneity; personalization; literacy; accessibility; plain language; reasonable consumer

**Record:** `ssrn-4491043-p08` · `machine-drafted-source-checked`

## 9. Many readability reforms neglect contract length even though a plain but extremely long document may still go unread

**Location:** Part II: LLMs and the No-Reading Problem, printed pp. 6 (PDF pp. 6)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 6, that many readability reforms neglect contract length even though a plain but extremely long document may still go unread. Reducing sentence and word difficulty cannot overcome the time cost of dozens of pages, making compression a separate design objective. This is significant because accessibility requires managing both cognitive complexity and attention burden. It connects to contract length, attention costs, plain language, reading burden, disclosure overload, consumer time.

**Evidence anchor:** Page 6 identifies excessive length as a gap in many legal readability frameworks.

**Boundary:** Shorter text can omit nuance, so compression must be assessed together with substantive quality.

**Connections:** contract length; attention costs; plain language; reading burden; disclosure overload; consumer time

**Record:** `ssrn-4491043-p09` · `machine-drafted-source-checked`

## 10. Smart readers shift interpretive power toward consumers because they can parse and personalize seller-drafted text without the seller's cooperation

**Location:** Part II: LLMs and the No-Reading Problem, printed pp. 6-7 (PDF pp. 6-7)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 6–7, that smart readers shift interpretive power toward consumers because they can parse and personalize seller-drafted text without the seller's cooperation. Instead of forcing one official rewrite on everyone, the tool can translate the same contract for a user's specific literacy, questions, or concerns and thereby penetrate a HIDE strategy. This is significant because consumer-side technology can alter contractual power without first changing the seller's form or the legal rule of assent. It connects to consumer-side technology, personalization, contract interpretation, HIDE strategy, private empowerment, seller control.

**Evidence anchor:** Pages 6-7 describe smart readers as consumer-controlled tools able to tailor dense forms.

**Boundary:** The study tests simplification rather than full personalization and does not empirically measure seller responses.

**Connections:** consumer-side technology; personalization; contract interpretation; HIDE strategy; private empowerment; seller control

**Record:** `ssrn-4491043-p10` · `machine-drafted-source-checked`

## 11. Early GPT-2 and GPT-3 demonstrations established possibility but not reliability because outputs were sporadic, cherry-picked, meandering, or misleading

**Location:** Part II: LLMs and the No-Reading Problem, printed pp. 7 (PDF pp. 7)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 7, that early GPT-2 and GPT-3 demonstrations established possibility but not reliability because outputs were sporadic, cherry-picked, meandering, or misleading. The authors' earlier work openly relied on selected examples, which left unresolved whether a consumer could trust the technology across ordinary contracts. This is significant because proof of concept and dependable consumer utility are distinct evidentiary stages. It connects to proof of concept, GPT-2, GPT-3, cherry-picking, reliability, technology maturation.

**Evidence anchor:** Page 7 contrasts nascent 2021 demonstrations with the need for systematic present-day evaluation.

**Boundary:** The retrospective characterization does not reproduce the earlier output distribution in this chapter.

**Connections:** proof of concept; GPT-2; GPT-3; cherry-picking; reliability; technology maturation

**Record:** `ssrn-4491043-p11` · `machine-drafted-source-checked`

## 12. The early weaknesses of smart readers appeared to arise from limited data and compute rather than a missing conceptual breakthrough

**Location:** Part II: LLMs and the No-Reading Problem, printed pp. 7 (PDF pp. 7)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 7, that the early weaknesses of smart readers appeared to arise from limited data and compute rather than a missing conceptual breakthrough. On this view, inconsistency, scale, and context problems were transient engineering constraints likely to improve along a visible capability trajectory. This is significant because the source of a system's limitation affects whether law should treat it as permanent or anticipate rapid mitigation. It connects to scaling laws, compute, training data, technical bottlenecks, forecasting, legal adaptation.

**Evidence anchor:** Page 7 explains why the authors expected the salient early problems to be temporary.

**Boundary:** This is a technological judgment by the authors, not a controlled causal demonstration that every failure was compute-bound.

**Connections:** scaling laws; compute; training data; technical bottlenecks; forecasting; legal adaptation

**Record:** `ssrn-4491043-p12` · `machine-drafted-source-checked`

## 13. GPT-4's exam performance, mass adoption, low user cost, and ease of use justify treating smart readers as a present policy question

**Location:** Part II: LLMs and the No-Reading Problem, printed pp. 7-8 (PDF pp. 7-8)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 7–8, that GPT-4's exam performance, mass adoption, low user cost, and ease of use justify treating smart readers as a present policy question. The combination of much stronger language capability and consumer availability changes the relevant inquiry from speculative feasibility to practical limitations and governance. This is significant because deployment conditions matter alongside benchmark capability when assessing legal significance. It connects to GPT-4, bar exam, mass adoption, consumer access, deployment, technology policy.

**Evidence anchor:** Pages 7-8 use contemporary performance and adoption evidence to motivate the study.

**Boundary:** Exam performance is not direct evidence of contract-simplification accuracy and should be treated as contextual rather than dispositive.

**Connections:** GPT-4; bar exam; mass adoption; consumer access; deployment; technology policy

**Record:** `ssrn-4491043-p13` · `machine-drafted-source-checked`

## 14. The study uses eight contracts and privacy policies from major services across different industries, lengths, and complexity levels

**Location:** Part III: Dataset and Methodology, printed pp. 8-9 (PDF pp. 8-9)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 8–9, that the study uses eight contracts and privacy policies from major services across different industries, lengths, and complexity levels. Yahoo, the Wall Street Journal, Spotify, Snapchat, Netflix, Google, Amazon, and Airbnb supply recognizable mass-market documents for comparative simplification. This is significant because a diverse but bounded dataset tests more than a single favorable example while retaining close qualitative review. It connects to dataset design, consumer contracts, privacy policies, cross-industry sample, major platforms, external validity.

**Evidence anchor:** Pages 8-9 identify the eight documents and their selection rationale.

**Boundary:** Eight purposively selected documents are not representative of all contracts, jurisdictions, languages, or consumer products.

**Connections:** dataset design; consumer contracts; privacy policies; cross-industry sample; major platforms; external validity

**Record:** `ssrn-4491043-p14` · `machine-drafted-source-checked`

## 15. A meaningful smart-reader evaluation must measure readability, length, and substantive quality separately

**Location:** Part III: Dataset and Methodology, printed pp. 9-10 (PDF pp. 9-10)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 9–10, that a meaningful smart-reader evaluation must measure readability, length, and substantive quality separately. Syntactic simplicity and compression can improve surface metrics while deleting rights, risks, obligations, context, or legally operative nuance. This is significant because optimization on readability alone can produce a polished but materially worse consumer disclosure. It connects to multi-metric evaluation, readability, compression, substantive fidelity, legal rights, quality assessment.

**Evidence anchor:** Pages 9-10 define the three criteria and explain why simplification can sacrifice meaning.

**Boundary:** The study's quality review is labor-intensive and partly subjective rather than a fully validated automated metric.

**Connections:** multi-metric evaluation; readability; compression; substantive fidelity; legal rights; quality assessment

**Record:** `ssrn-4491043-p15` · `machine-drafted-source-checked`

## 16. Standard readability formulas are limited and manipulable because they infer difficulty from surface features rather than meaning

**Location:** Part III: Dataset and Methodology, printed pp. 9-10 (PDF pp. 9-10)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 9–10, that standard readability formulas are limited and manipulable because they infer difficulty from surface features rather than meaning. Scores built from sentence length, syllables, and word rarity can vary dramatically across implementations and do not directly test whether a reader understands a legal concept. This is significant because headline grade levels should be interpreted as comparative signals, not literal measures of legal comprehension. It connects to Flesch-Kincaid, readability validity, metric manipulation, surface features, legal comprehension, measurement error.

**Evidence anchor:** Pages 9-10 explain common readability formulas and their validity and implementation problems.

**Boundary:** Despite the critique, the study still relies on these formulas for within-document comparisons.

**Connections:** Flesch-Kincaid; readability validity; metric manipulation; surface features; legal comprehension; measurement error

**Record:** `ssrn-4491043-p16` · `machine-drafted-source-checked`

## 17. The Combined Readability Measure reduces implementation gaming by averaging within and across several readability tests

**Location:** Part III: Dataset and Methodology, printed pp. 9-10 (PDF pp. 9-10)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 9–10, that the Combined Readability Measure reduces implementation gaming by averaging within and across several readability tests. CRM smooths the large score differences that can result from choosing one library or one formula while preserving a comparable grade-level signal. This is significant because measurement robustness requires aggregating noisy metrics rather than selecting the result most favorable to simplification. It connects to Combined Readability Measure, metric aggregation, robustness, implementation variance, grade level, empirical methods.

**Evidence anchor:** Pages 9-10 introduce CRM as a response to manipulability and inconsistent implementations.

**Boundary:** Averaging weak proxies can reduce variance without curing their shared lack of semantic validity.

**Connections:** Combined Readability Measure; metric aggregation; robustness; implementation variance; grade level; empirical methods

**Record:** `ssrn-4491043-p17` · `machine-drafted-source-checked`

## 18. Substantive quality is indispensable because a shorter, simpler contract may omit precisely the rights and risks consumers need to know

**Location:** Part III: Dataset and Methodology, printed pp. 10 (PDF pp. 10)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 10, that substantive quality is indispensable because a shorter, simpler contract may omit precisely the rights and risks consumers need to know. The evaluation therefore asks whether outputs preserve important legal aspects, obligations, remedies, and context rather than rewarding compression for its own sake. This is significant because consumer usefulness depends on faithful selection and explanation, not merely linguistic accessibility. It connects to legal fidelity, consumer rights, risk disclosure, summary quality, context preservation, evaluation design.

**Evidence anchor:** Page 10 defines quality as preservation of key facts, legal aspects, risks, obligations, and rights.

**Boundary:** Judgments about which provisions are important necessarily involve expert selection and may differ across users.

**Connections:** legal fidelity; consumer rights; risk disclosure; summary quality; context preservation; evaluation design

**Record:** `ssrn-4491043-p18` · `machine-drafted-source-checked`

## 19. Smart readers may reasonably use competent, inexpensive models rather than the most expensive state-of-the-art system

**Location:** Part III: Dataset and Methodology, printed pp. 10 (PDF pp. 10)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 10, that smart readers may reasonably use competent, inexpensive models rather than the most expensive state-of-the-art system. The study selected models from Anthropic and OpenAI that could operate cheaply, treating cost containment as a realistic feature of consumer deployment rather than a methodological defect. This is significant because accessibility at scale depends on the capability-cost frontier, not maximum laboratory performance. It connects to model selection, inference cost, Claude, OpenAI, deployment economics, consumer scale.

**Evidence anchor:** Page 10 explains the decision to test strong but inexpensive models from two providers.

**Boundary:** The historical model set cannot establish the optimal present-day cost-quality tradeoff.

**Connections:** model selection; inference cost; Claude; OpenAI; deployment economics; consumer scale

**Record:** `ssrn-4491043-p19` · `machine-drafted-source-checked`

## 20. Context-window limits required a custom pipeline that split contracts into full-sentence chunks, simplified each chunk, and recombined the outputs

**Location:** Part III: Dataset and Methodology, printed pp. 10-11 (PDF pp. 10-11)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 10–11, that context-window limits required a custom pipeline that split contracts into full-sentence chunks, simplified each chunk, and recombined the outputs. The code handled different APIs and input limits through LangChain while avoiding sentence fragments, though it could still sever dependencies across clauses. This is significant because system architecture can create errors that are not inherent in the language model itself. It connects to context windows, chunking, LangChain, API orchestration, sentence boundaries, cross-clause dependencies.

**Evidence anchor:** Pages 10-11 list the custom pipeline steps and explain chunking's semantic risks.

**Boundary:** Sentence-preserving chunks still lose inter-clause context, and a clause-aware method would require additional technical work.

**Connections:** context windows; chunking; LangChain; API orchestration; sentence boundaries; cross-clause dependencies

**Record:** `ssrn-4491043-p20` · `machine-drafted-source-checked`

## 21. Prompt wording is a critical treatment variable because small changes can radically alter simplification quality

**Location:** Part III: Dataset and Methodology, printed pp. 11 (PDF pp. 11)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 11, that prompt wording is a critical treatment variable because small changes can radically alter simplification quality. Lacking robust prompt-optimization tools, the authors used trial and error to balance brevity, low readability, plain wording, short sentences, and preservation of legal meaning. This is significant because performance claims about LLMs are inseparable from the instructions and examples used to elicit the output. It connects to prompt engineering, treatment specification, few-shot examples, capitalization, trial and error, reproducibility.

**Evidence anchor:** Page 11 reproduces the prompt and explains its construction and emphasis choices.

**Boundary:** The chosen prompt was not proven globally optimal and may have induced some later terminology errors.

**Connections:** prompt engineering; treatment specification; few-shot examples; capitalization; trial and error; reproducibility

**Record:** `ssrn-4491043-p21` · `machine-drafted-source-checked`

## 22. The prompt itself embodied a tradeoff because adding detail could improve guidance while consuming scarce context space

**Location:** Part III: Dataset and Methodology, printed pp. 11 (PDF pp. 11)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 11, that the prompt itself embodied a tradeoff because adding detail could improve guidance while consuming scarce context space. The authors kept the instruction shorter than desired, used capitalization to prioritize meaning, and supplied a few word-replacement examples known to improve model performance. This is significant because input limits force designers to allocate context between the source document and instructions about how to process it. It connects to prompt length, context budget, few-shot prompting, instruction hierarchy, model performance, design tradeoffs.

**Evidence anchor:** Page 11 discusses why the prompt was short and why it used capitalization and examples.

**Boundary:** Longer-context models may reduce this particular constraint but do not eliminate conflicts among simplification goals.

**Connections:** prompt length; context budget; few-shot prompting; instruction hierarchy; model performance; design tradeoffs

**Record:** `ssrn-4491043-p22` · `machine-drafted-source-checked`

## 23. The analysis combines document-level metrics, a trap-based quality audit, and close review of eight difficult clauses

**Location:** Part III: Dataset and Methodology, printed pp. 11-12 (PDF pp. 11-12)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 11–12, that the analysis combines document-level metrics, a trap-based quality audit, and close review of eight difficult clauses. The three phases move from scalable quantitative comparison to expert checking of important terms and then to granular diagnosis of particular simplification failures. This is significant because triangulation reveals errors that neither aggregate readability scores nor isolated anecdotes can identify alone. It connects to mixed methods, quality audit, trap detection, clause analysis, quantitative metrics, triangulation.

**Evidence anchor:** Pages 11-12 describe the three analytical phases.

**Boundary:** The qualitative phases rely on author judgment and do not include blinded independent raters or consumer-comprehension experiments.

**Connections:** mixed methods; quality audit; trap detection; clause analysis; quantitative metrics; triangulation

**Record:** `ssrn-4491043-p23` · `machine-drafted-source-checked`

## 24. Across contracts and models, simplified outputs averaged about thirty percent of the original word count

**Location:** Part IV.A.1: Text Length, printed pp. 12-13 (PDF pp. 12-13)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 12–13, that across contracts and models, simplified outputs averaged about thirty percent of the original word count. Estimated reading time fell from roughly twenty minutes and forty-five seconds to six minutes and six seconds, a reduction of about fourteen minutes and thirty-nine seconds. This is significant because the time burden of reading mass-market terms can be reduced enough to alter the consumer's practical decision whether to engage. It connects to word-count reduction, reading time, attention costs, contract length, summarization, consumer engagement.

**Evidence anchor:** Pages 12-13 report the average word and reading-time reductions across all agreements and models.

**Boundary:** Reading-time estimates use population-average rates and do not measure whether consumers actually read or understand the outputs.

**Connections:** word-count reduction; reading time; attention costs; contract length; summarization; consumer engagement

**Record:** `ssrn-4491043-p24` · `machine-drafted-source-checked`

## 25. Models varied dramatically in compression even under the same prompt

**Location:** Part IV.A.1: Text Length, printed pp. 13 (PDF pp. 13)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 13, that models varied dramatically in compression even under the same prompt. The least compressive model reduced length by forty-nine percent, the most compressive by 88.4 percent, and the longest output remained about three times the shortest. This is significant because “an LLM summary” is not a stable treatment, and model choice materially changes the consumer artifact. It connects to model heterogeneity, compression ratio, output variance, prompt consistency, model selection, reliability.

**Evidence anchor:** Page 13 reports model-level variability in length reduction.

**Boundary:** The comparison does not isolate whether architecture, training, provider defaults, or sampling settings caused the differences.

**Connections:** model heterogeneity; compression ratio; output variance; prompt consistency; model selection; reliability

**Record:** `ssrn-4491043-p25` · `machine-drafted-source-checked`

## 26. Smart readers reduced the number of difficult words by an average of 328, or sixty-one percent

**Location:** Part IV.A.2: Text Complexity, printed pp. 13-14 (PDF pp. 13-14)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 13–14, that smart readers reduced the number of difficult words by an average of 328, or sixty-one percent. Even the weakest tested model removed thirty-six percent of words outside the Dale-Chall familiar-word list, though performance varied substantially. This is significant because simplification changed vocabulary as well as overall length. It connects to Dale-Chall, difficult words, lexical complexity, vocabulary simplification, model variability, readability.

**Evidence anchor:** Pages 13-14 report the difficult-word methodology and results.

**Boundary:** The Dale-Chall list is not tailored to legal jargon and supports relative, not absolute, claims about difficulty.

**Connections:** Dale-Chall; difficult words; lexical complexity; vocabulary simplification; model variability; readability

**Record:** `ssrn-4491043-p26` · `machine-drafted-source-checked`

## 27. The average model improved Flesch-Kincaid readability by 1.47 grade levels, while Claude-001 improved it by 5.6 grades to about grade 5.4

**Location:** Part IV.A.3: Text Readability, printed pp. 14-15 (PDF pp. 14-15)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 14–15, that the average model improved Flesch-Kincaid readability by 1.47 grade levels, while Claude-001 improved it by 5.6 grades to about grade 5.4. The strongest output therefore reached a level below the sixth-to-eighth-grade band often recommended for public materials, even though the original contracts required ten to fourteen years of schooling. This is significant because some models can produce accessibility gains far larger than the cross-model average. It connects to Flesch-Kincaid, grade level, Claude-001, public communication, accessibility, model comparison.

**Evidence anchor:** Pages 14-15 report original grade levels, mean improvement, and Claude-001's best result.

**Boundary:** Grade-level labels are formula outputs, not direct findings that fifth-grade children understand the legal consequences.

**Connections:** Flesch-Kincaid; grade level; Claude-001; public communication; accessibility; model comparison

**Record:** `ssrn-4491043-p27` · `machine-drafted-source-checked`

## 28. The more robust CRM measure showed a smaller average gain of about one grade but still found a roughly five-grade improvement for Claude-001

**Location:** Part IV.A.3: Text Readability, printed pp. 15-16 (PDF pp. 15-16)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 15–16, that the more robust CRM measure showed a smaller average gain of about one grade but still found a roughly five-grade improvement for Claude-001. Convergence on a large best-model effect alongside a modest mean effect suggests both genuine capability and consequential model selection. This is significant because robustness checks qualify the headline without eliminating the central improvement. It connects to CRM, robustness check, Claude-001, grade-level reduction, average treatment, model selection.

**Evidence anchor:** Pages 15-16 compare CRM results with the Flesch-Kincaid findings.

**Boundary:** CRM inherits shared limitations of formulaic readability measures and does not test legal understanding.

**Connections:** CRM; robustness check; Claude-001; grade-level reduction; average treatment; model selection

**Record:** `ssrn-4491043-p28` · `machine-drafted-source-checked`

## 29. Some models made contracts more complex despite an explicit simplification prompt

**Location:** Part IV.A.3: Text Readability, printed pp. 16 (PDF pp. 16)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 16, that some models made contracts more complex despite an explicit simplification prompt. Claude-001 and Text-Davinci-003 were the most consistent, but the direction of change was not uniformly beneficial across systems. This is significant because deployment requires model-level validation rather than confidence in the general category of LLMs. It connects to performance regressions, model validation, instruction following, deployment risk, model heterogeneity, quality assurance.

**Evidence anchor:** Page 16 notes adverse complexity results and identifies the most consistent models.

**Boundary:** The result is tied to the tested models and prompt and does not predict every later system.

**Connections:** performance regressions; model validation; instruction following; deployment risk; model heterogeneity; quality assurance

**Record:** `ssrn-4491043-p29` · `machine-drafted-source-checked`

## 30. Both ChatGPT-Turbo and Claude made the sampled Spotify terms substantially simpler while retaining most important consumer information

**Location:** Part IV.B: Quality Assessment, printed pp. 16-17 (PDF pp. 16-17)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 16–17, that both ChatGPT-Turbo and Claude made the sampled Spotify terms substantially simpler while retaining most important consumer information. The authors' expert audit compared the original first three clauses with two outputs and checked eleven preidentified traps rather than relying on stylistic impression alone. This is significant because substantive usefulness can coexist with compression when important terms are audited directly. It connects to Spotify terms, expert audit, consumer traps, ChatGPT-Turbo, Claude, substantive quality.

**Evidence anchor:** Pages 16-17 describe the Spotify quality-assessment procedure and overall result.

**Boundary:** The audit covers only the first 2,360 words of one contract and uses author-selected traps.

**Connections:** Spotify terms; expert audit; consumer traps; ChatGPT-Turbo; Claude; substantive quality

**Record:** `ssrn-4491043-p30` · `machine-drafted-source-checked`

## 31. ChatGPT-Turbo preserved nine of eleven identified traps, and partially different model omissions support ensemble review

**Location:** Part IV.B: Quality Assessment, printed pp. 17 (PDF pp. 17)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 17, that ChatGPT-Turbo preserved nine of eleven identified traps, and partially different model omissions support ensemble review. Because one system may catch an issue another misses, combining outputs or detectors can improve coverage and reveal categories of systematically neglected information. This is significant because model diversity can be used as a safety mechanism rather than treated solely as inconsistency. It connects to ensemble methods, omission errors, trap coverage, model diversity, quality assurance, systematic blind spots.

**Evidence anchor:** Page 17 reports nine-of-eleven trap coverage and draws the ensemble implication.

**Boundary:** The overlap and complementarity of two models in one sample do not establish the reliability or optimal design of an ensemble.

**Connections:** ensemble methods; omission errors; trap coverage; model diversity; quality assurance; systematic blind spots

**Record:** `ssrn-4491043-p31` · `machine-drafted-source-checked`

## 32. Chunking can cause omissions and presentation problems by disrupting the flow and interdependence of contractual provisions

**Location:** Part IV.B: Quality Assessment, printed pp. 17 (PDF pp. 17)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 17, that chunking can cause omissions and presentation problems by disrupting the flow and interdependence of contractual provisions. Even when chunks end at sentence boundaries, a qualification or definition in another clause may be absent from the model's local context. This is significant because some apparent model errors are pipeline errors that longer context or better segmentation may reduce. It connects to chunking artifacts, cross-reference loss, context windows, contract interdependence, pipeline design, transient limitations.

**Evidence anchor:** Page 17 attributes some quality problems to the chunking strategy and predicts improvement as input limits expand.

**Boundary:** Longer contexts reduce but do not necessarily eliminate attention, retrieval, or document-structure failures.

**Connections:** chunking artifacts; cross-reference loss; context windows; contract interdependence; pipeline design; transient limitations

**Record:** `ssrn-4491043-p32` · `machine-drafted-source-checked`

## 33. Hard-case selection is more informative than random-clause accuracy for assessing consumer protection

**Location:** Part V: Specific Clauses, printed pp. 17-18 (PDF pp. 17-18)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 17–18, that hard-case selection is more informative than random-clause accuracy for assessing consumer protection. The authors intentionally chose eight provisions involving cancellation, data sharing, liability, dispute resolution, and unilateral modification because errors there could materially affect unsuspecting users. This is significant because average performance on trivial language can conceal failure on the terms consumers most need explained. It connects to hard-case evaluation, consumer traps, clause selection, risk-sensitive testing, legal salience, benchmark design.

**Evidence anchor:** Pages 17-18 explain why the close review focuses on eight consequential provisions and uses GPT-4.

**Boundary:** Purposive hard-case selection cannot estimate the prevalence of error across all clauses.

**Connections:** hard-case evaluation; consumer traps; clause selection; risk-sensitive testing; legal salience; benchmark design

**Record:** `ssrn-4491043-p33` · `machine-drafted-source-checked`

## 34. GPT-4 effectively simplified the Wall Street Journal's unilateral-change clause but blurred notice as a condition of effectiveness

**Location:** Part V.A: WSJ Changes to Subscriber Agreement, printed pp. 18-19 (PDF pp. 18-19)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 18–19, that GPT-4 effectively simplified the Wall Street Journal's unilateral-change clause but blurred notice as a condition of effectiveness. The output's statement that the service will “let you know” can sound like a courtesy even though the original makes communication part of the mechanism by which amended terms bind the subscriber. This is significant because a small modal or structural change can alter the consumer's understanding of when contractual obligations arise. It connects to unilateral modification, notice, continued use, contract formation, semantic nuance, consumer understanding.

**Evidence anchor:** Pages 18-19 compare the original and simplified WSJ modification clause and identify the notice issue.

**Boundary:** The authors view the difference as subtle and acknowledge that the simplified sentence remains somewhat ambiguous rather than clearly false.

**Connections:** unilateral modification; notice; continued use; contract formation; semantic nuance; consumer understanding

**Record:** `ssrn-4491043-p34` · `machine-drafted-source-checked`

## 35. The WSJ modification output cut 26 percent of the words and approximately eight grade levels while preserving most operative meaning

**Location:** Part V.A: WSJ Changes to Subscriber Agreement, printed pp. 19-20 (PDF pp. 19-20)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 19–20, that the WSJ modification output cut 26 percent of the words and approximately eight grade levels while preserving most operative meaning. It replaced three long sentences with six shorter ones, used direct second-person phrasing, and made the clause readable at roughly the eighth-grade level. This is significant because substantial readability gains need not require wholesale loss of content in a relatively tractable clause. It connects to unilateral modification, grade-level reduction, sentence splitting, second-person drafting, word count, plain language.

**Evidence anchor:** Pages 19-20 report the qualitative strengths and quantitative changes for the first WSJ clause.

**Boundary:** The notice ambiguity shows that strong quantitative and stylistic performance does not guarantee full legal fidelity.

**Connections:** unilateral modification; grade-level reduction; sentence splitting; second-person drafting; word count; plain language

**Record:** `ssrn-4491043-p35` · `machine-drafted-source-checked`

## 36. The arbitration simplification improved accessibility but made the term “arbitration” less salient by retitling the clause “Solving Disputes”

**Location:** Part V.B: WSJ Agreement to Arbitrate, printed pp. 20-22 (PDF pp. 20-22)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 20–22, that the arbitration simplification improved accessibility but made the term “arbitration” less salient by retitling the clause “Solving Disputes”. A friendlier heading communicates function yet may obscure the waiver of jury trial, individual-only process, and special procedural regime that make arbitration legally consequential. This is significant because plain language can reduce the warning function of legal labels even when it explains their general subject. It connects to mandatory arbitration, heading salience, jury-trial waiver, class waiver, access to justice, consumer notice.

**Evidence anchor:** Pages 20-22 reproduce the arbitration output and criticize its replacement heading.

**Boundary:** The body still uses the word arbitration, so the critique concerns prominence rather than total omission.

**Connections:** mandatory arbitration; heading salience; jury-trial waiver; class waiver; access to justice; consumer notice

**Record:** `ssrn-4491043-p36` · `machine-drafted-source-checked`

## 37. Replacing “intellectual property” with “ideas,” “class actions” with “group actions,” and award entry with award “use” changes legal meaning

**Location:** Part V.B: WSJ Agreement to Arbitrate, printed pp. 20-22 (PDF pp. 20-22)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 20–22, that replacing “intellectual property” with “ideas,” “class actions” with “group actions,” and award entry with award “use” changes legal meaning. These substitutions sound simpler but expand, blur, or distort the categories of disputes, procedures, and judicial enforcement described by the original. This is significant because legal terms of art cannot always be translated through ordinary-language synonym replacement. It connects to intellectual property, class actions, arbitration awards, terms of art, semantic distortion, legal translation.

**Evidence anchor:** Page 22 identifies several legally inaccurate lexical substitutions in the WSJ arbitration summary.

**Boundary:** Some lay readers may understand the simplified wording better, creating a real tradeoff between doctrinal precision and immediate accessibility.

**Connections:** intellectual property; class actions; arbitration awards; terms of art; semantic distortion; legal translation

**Record:** `ssrn-4491043-p37` · `machine-drafted-source-checked`

## 38. Omitting the consumer's right to bring issues to government agencies is a substantively important loss

**Location:** Part V.B: WSJ Agreement to Arbitrate, printed pp. 20-22 (PDF pp. 20-22)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 20–22, that omitting the consumer's right to bring issues to government agencies is a substantively important loss. The simplified arbitration clause preserved many private-process details but deleted a route through which federal, state, or local agencies might seek relief. This is significant because summaries can disproportionately hide exceptions and residual rights that qualify a broad waiver. It connects to agency complaints, residual rights, arbitration clauses, omission error, consumer remedies, government enforcement.

**Evidence anchor:** Page 22 calls the missing agency-complaint language a worrisome omission.

**Boundary:** The study identifies the omission but does not test whether consumers would act differently if the agency route were included.

**Connections:** agency complaints; residual rights; arbitration clauses; omission error; consumer remedies; government enforcement

**Record:** `ssrn-4491043-p38` · `machine-drafted-source-checked`

## 39. Translating “equitable relief” as “fair changes” illustrates a genuine conflict between doctrinal precision and lay comprehension

**Location:** Part V.B: WSJ Agreement to Arbitrate, printed pp. 22-23 (PDF pp. 22-23)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 22–23, that translating “equitable relief” as “fair changes” illustrates a genuine conflict between doctrinal precision and lay comprehension. The legal category can include injunctions, in-kind remedies, apologies, and other nonmonetary relief, while the simpler phrase is both inaccurate and potentially more intelligible than the original jargon. This is significant because some legal concepts may require explanation rather than a shorter substitute. It connects to equitable relief, legal terminology, lay translation, injunctions, remedies, explanatory drafting.

**Evidence anchor:** Pages 22-23 analyze the “fair changes” substitution and the quantitative improvement in the clause.

**Boundary:** The authors call the choice disputable but not irrational and do not identify one universally adequate short replacement.

**Connections:** equitable relief; legal terminology; lay translation; injunctions; remedies; explanatory drafting

**Record:** `ssrn-4491043-p39` · `machine-drafted-source-checked`

## 40. GPT-4 halved and greatly simplified Airbnb's data-collection clause while preserving its overall architecture

**Location:** Part V.C: Airbnb Third-Party Information, printed pp. 23-25 (PDF pp. 23-25)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 23–25, that GPT-4 halved and greatly simplified Airbnb's data-collection clause while preserving its overall architecture. The output retained third-party logins, background checks, co-travelers, insurance claims, fraud warnings, off-platform activity, and health data in shorter and more accessible language. This is significant because a complex privacy disclosure can be compressed substantially without losing its basic categories. It connects to privacy policies, third-party data, background checks, health information, data collection, contract simplification.

**Evidence anchor:** Pages 23-25 compare the original and simplified Airbnb provision and report a fifty-percent word reduction.

**Boundary:** Preserving broad categories does not ensure that every unexpected data flow or legal condition remains visible.

**Connections:** privacy policies; third-party data; background checks; health information; data collection; contract simplification

**Record:** `ssrn-4491043-p40` · `machine-drafted-source-checked`

## 41. The Airbnb simplification omitted surprising data such as friends lists and misrepresented consent as always required for some background checks

**Location:** Part V.C: Airbnb Third-Party Information, printed pp. 24-25 (PDF pp. 24-25)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 24–25, that the Airbnb simplification omitted surprising data such as friends lists and misrepresented consent as always required for some background checks. The original conditions consent on whether law requires it, but the output says the user will agree, implying control that may not exist; it also weakens limits tied to the insurer and privacy policy. This is significant because omission and altered conditional language can give users false reassurance about privacy control. It connects to friends lists, conditional consent, background checks, privacy expectations, omission error, false reassurance.

**Evidence anchor:** Page 25 identifies the friends-list, consent, and insurance-related shortcomings.

**Boundary:** The friends-list omission is not necessarily misleading by itself, and the relative importance of examples may vary among users.

**Connections:** friends lists; conditional consent; background checks; privacy expectations; omission error; false reassurance

**Record:** `ssrn-4491043-p41` · `machine-drafted-source-checked`

## 42. The Netflix simplification committed a major legal inversion by saying the law prevents refunds

**Location:** Part V.D: Netflix Cancellation, printed pp. 26-27 (PDF pp. 26-27)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 26–27, that the Netflix simplification committed a major legal inversion by saying the law prevents refunds. The source instead says Netflix generally denies refunds only to the extent applicable law permits, leaving open jurisdictions where law requires them. This is significant because a fluent summary can reverse the relationship between private policy and mandatory consumer law and thereby suppress perceived rights. It connects to refund rights, mandatory law, Netflix cancellation, semantic inversion, consumer remedies, hallucinated prohibition.

**Evidence anchor:** Pages 26-27 compare the refund sentences and characterize the simplified version as materially misleading.

**Boundary:** The remainder of the cancellation summary was largely accurate and accessible, so the example is a localized but serious error.

**Connections:** refund rights; mandatory law; Netflix cancellation; semantic inversion; consumer remedies; hallucinated prohibition

**Record:** `ssrn-4491043-p42` · `machine-drafted-source-checked`

## 43. Surface improvements can coexist with a rights-undermining error

**Location:** Part V.D: Netflix Cancellation, printed pp. 26-27 (PDF pp. 26-27)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 26–27, that surface improvements can coexist with a rights-undermining error. The Netflix output lowered apparent difficulty from high-school to sixth- or seventh-grade level and shortened the text by sixteen percent, yet its central refund statement could mislead consumers where law mandates repayment. This is significant because readability metrics cannot serve as a proxy for legal correctness. It connects to metric validity, refund policy, readability gains, legal accuracy, consumer rights, quality control.

**Evidence anchor:** Pages 26-27 juxtapose the quantitative gains with the refund-policy error.

**Boundary:** The case study does not establish how often similarly severe inversions occur.

**Connections:** metric validity; refund policy; readability gains; legal accuracy; consumer rights; quality control

**Record:** `ssrn-4491043-p43` · `machine-drafted-source-checked`

## 44. The Amazon content summary was broadly effective but omitted explicit bans on email spoofing and political campaigning

**Location:** Part V.E: Amazon User Content, printed pp. 27-29 (PDF pp. 27-29)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 27–29, that the Amazon content summary was broadly effective but omitted explicit bans on email spoofing and political campaigning. Compression preserved permissions, Amazon's license, user warranties, indemnity, and moderation discretion while dropping examples that may matter to specific users. This is significant because the proper balance between exhaustive illustrations and usable length is context-dependent. It connects to user-generated content, email spoofing, political campaigning, content licenses, indemnity, selective omission.

**Evidence anchor:** Pages 27-29 reproduce and assess the Amazon reviews and content clause.

**Boundary:** The original may itself contain more illustrations than a concise consumer summary needs.

**Connections:** user-generated content; email spoofing; political campaigning; content licenses; indemnity; selective omission

**Record:** `ssrn-4491043-p44` · `machine-drafted-source-checked`

## 45. Simplification's casual tone can understate the gravity of user warranties and indemnity obligations

**Location:** Part V.E: Amazon User Content, printed pp. 28-29 (PDF pp. 28-29)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 28–29, that simplification's casual tone can understate the gravity of user warranties and indemnity obligations. Phrases such as “share stuff” and “cover Amazon” convey the gist but may not signal that users guarantee accuracy, assume rights-related duties, and bear claims arising from their content. This is significant because formality can carry substantive warning value even when it increases difficulty. It connects to legal tone, indemnity, user warranties, warning salience, plain language, contract gravity.

**Evidence anchor:** Page 29 identifies tone as a deeper tradeoff in simplifying weighty user obligations.

**Boundary:** The authors note that related empirical work did not generally validate a red-flag theory of formal typography, so this remains a contextual concern.

**Connections:** legal tone; indemnity; user warranties; warning salience; plain language; contract gravity

**Record:** `ssrn-4491043-p45` · `machine-drafted-source-checked`

## 46. The Amazon content output achieved a dramatic grade-level improvement while increasing sentence count

**Location:** Part V.E: Amazon User Content, printed pp. 29-30 (PDF pp. 29-30)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 29–30, that the Amazon content output achieved a dramatic grade-level improvement while increasing sentence count. The text fell from roughly nineteen-to-twenty years of education to eighth-grade readability and lost twenty-six percent of its words by splitting eight sentences into twelve shorter units. This is significant because more sentences can accompany greater readability when long legal sentences are decomposed. It connects to sentence splitting, grade-level reduction, word count, Amazon reviews, readability, syntactic complexity.

**Evidence anchor:** Pages 29-30 report the quantitative results for the user-content clause.

**Boundary:** The implausibly high original grade estimate illustrates why formula outputs should be interpreted comparatively.

**Connections:** sentence splitting; grade-level reduction; word count; Amazon reviews; readability; syntactic complexity

**Record:** `ssrn-4491043-p46` · `machine-drafted-source-checked`

## 47. The Amazon risk-of-loss simplification explains title transfer but fails to make the consumer's delivery risk sufficiently explicit

**Location:** Part V.F: Amazon Risk of Loss, printed pp. 30-31 (PDF pp. 30-31)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 30–31, that the Amazon risk-of-loss simplification explains title transfer but fails to make the consumer's delivery risk sufficiently explicit. Saying that goods become “yours” when handed to the carrier does not clearly warn that the buyer may bear loss when a package disappears in transit. This is significant because legal allocation of risk can vanish even when a literal ownership proposition survives. It connects to risk of loss, shipment contracts, title transfer, delivery loss, consumer expectations, implicit consequences.

**Evidence anchor:** Pages 30-31 explain why the simplified shipment sentence understates the customer's risk.

**Boundary:** Actual enforceability and background consumer law vary by jurisdiction and are not adjudicated in the chapter.

**Connections:** risk of loss; shipment contracts; title transfer; delivery loss; consumer expectations; implicit consequences

**Record:** `ssrn-4491043-p47` · `machine-drafted-source-checked`

## 48. The returns summary similarly fails to spell out who bears loss before a returned item reaches Amazon's fulfillment center

**Location:** Part V.F: Amazon Risk of Loss, printed pp. 30-31 (PDF pp. 30-31)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 30–31, that the returns summary similarly fails to spell out who bears loss before a returned item reaches Amazon's fulfillment center. The phrase “it's not ours” is easier than “Amazon does not take title,” but leaves the practical consequence of a lost return implicit. This is significant because simple wording may preserve a legal status while failing to communicate why that status matters. It connects to return shipping, risk allocation, title, fulfillment centers, practical consequences, consumer understanding.

**Evidence anchor:** Page 31 assesses the return-title language and reports the limited quantitative gain.

**Boundary:** The clause showed only modest readability improvement because the original was already short and relatively simple.

**Connections:** return shipping; risk allocation; title; fulfillment centers; practical consequences; consumer understanding

**Record:** `ssrn-4491043-p48` · `machine-drafted-source-checked`

## 49. Smart readers cannot fully repair a privacy clause whose original drafting is internally inconsistent

**Location:** Part V.G: Yahoo Information Sharing, printed pp. 32-34 (PDF pp. 32-34)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 32–34, that smart readers cannot fully repair a privacy clause whose original drafting is internally inconsistent. Yahoo's text appears to promise that identifying information is not shared outside the company without consent while also creating overlapping partner and advertising pathways that are hard to reconcile. This is significant because AI simplification has a source-quality ceiling: a clearer summary of confused drafting may still be confused or misleading. It connects to source-text defects, privacy policies, internal inconsistency, information sharing, garbage in garbage out, legal drafting.

**Evidence anchor:** Pages 32-34 reproduce the Yahoo clause and diagnose inconsistent sharing rules.

**Boundary:** The authors infer intended logic but acknowledge that even a strong simplification may not resolve ambiguous cookie-matching language.

**Connections:** source-text defects; privacy policies; internal inconsistency; information sharing; garbage in garbage out; legal drafting

**Record:** `ssrn-4491043-p49` · `machine-drafted-source-checked`

## 50. Firms may respond strategically to smart readers by drafting terms that confuse or circumvent automated interpretation

**Location:** Part V.G: Yahoo Information Sharing, printed pp. 34 (PDF pp. 34)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 34, that firms may respond strategically to smart readers by drafting terms that confuse or circumvent automated interpretation. If consumer tools threaten HIDE, sellers can exploit ambiguous structure, interdependent clauses, or adversarial wording to preserve opacity. This is significant because technology that changes bargaining power also changes the regulated party's incentives to evade it. It connects to adversarial drafting, regulatory evasion, HIDE strategy, model capture, contract design, arms races.

**Evidence anchor:** Page 34 generalizes from the source-text problem to possible firm strategies against smart readers.

**Boundary:** The Yahoo example motivates the concern but does not demonstrate intentional adversarial drafting.

**Connections:** adversarial drafting; regulatory evasion; HIDE strategy; model capture; contract design; arms races

**Record:** `ssrn-4491043-p50` · `machine-drafted-source-checked`

## 51. The Yahoo simplification usefully clarified much of the clause but may overstate consent and omitted the opaque cookie-matching practice

**Location:** Part V.G: Yahoo Information Sharing, printed pp. 34 (PDF pp. 34)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 34, that the Yahoo simplification usefully clarified much of the clause but may overstate consent and omitted the opaque cookie-matching practice. Rendering consent as “You Say Okay” sounds accessible yet can imply a more affirmative choice than a checkbox, while dropping the hardest detail deprives users of information rather than resolving it. This is significant because friendly phrasing can simultaneously improve comprehension and exaggerate user agency. It connects to privacy consent, cookie matching, checkbox consent, omission, user agency, plain language.

**Evidence anchor:** Page 34 evaluates the consent phrasing, cookie omission, and major readability and length gains.

**Boundary:** The underlying cookie language was itself difficult to interpret, limiting what any summary could communicate confidently.

**Connections:** privacy consent; cookie matching; checkbox consent; omission; user agency; plain language

**Record:** `ssrn-4491043-p51` · `machine-drafted-source-checked`

## 52. The Spotify summary conveyed the disclaimer's broad message but mistranslated several distinct damages doctrines

**Location:** Part V.H: Spotify Liability Limitation, printed pp. 35-37 (PDF pp. 35-37)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 35–37, that the Spotify summary conveyed the disclaimer's broad message but mistranslated several distinct damages doctrines. “Incidental damages” became “extra damages,” “consequential damages” became “follow-on damages,” and foreseeability became what could be “seen coming,” sacrificing operative legal categories for colloquial language. This is significant because domain-naive simplification can change remedies precisely where technical vocabulary carries legal consequences. It connects to incidental damages, consequential damages, foreseeability, liability caps, terms of art, legal remedies.

**Evidence anchor:** Pages 35-37 compare the Spotify limitation clause and identify altered damages terminology.

**Boundary:** Some colloquial substitutes may improve general understanding, and the source clause itself is internally confusing.

**Connections:** incidental damages; consequential damages; foreseeability; liability caps; terms of art; legal remedies

**Record:** `ssrn-4491043-p52` · `machine-drafted-source-checked`

## 53. The Spotify source clause itself is inconsistent about whether the user's sole remedy is exit, damages are capped at thirty dollars, or mandatory law preserves liability

**Location:** Part V.H: Spotify Liability Limitation, printed pp. 35-37 (PDF pp. 35-37)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 35–37, that the Spotify source clause itself is inconsistent about whether the user's sole remedy is exit, damages are capped at thirty dollars, or mandatory law preserves liability. A simplifier must present a hierarchy among apparently conflicting propositions without inventing a legal resolution absent from the text. This is significant because contract complexity sometimes reflects substantive inconsistency rather than merely difficult syntax. It connects to sole remedy, liability cap, mandatory law, source ambiguity, Spotify, contract hierarchy.

**Evidence anchor:** Pages 35-37 identify the original clause's tension among exclusive remedy, monetary cap, and nonwaivable liability.

**Boundary:** The chapter critiques drafting coherence but does not decide the clause's enforceability or authoritative interpretation.

**Connections:** sole remedy; liability cap; mandatory law; source ambiguity; Spotify; contract hierarchy

**Record:** `ssrn-4491043-p53` · `machine-drafted-source-checked`

## 54. The Spotify output shows that formulaic readability can improve even when a long, inaccessible structure and substantive errors remain

**Location:** Part V.H and Part VI, printed pp. 37-38 (PDF pp. 37-38)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 37–38, that the Spotify output shows that formulaic readability can improve even when a long, inaccessible structure and substantive errors remain. The score fell by seven or eight grade levels, yet the output grew slightly from 307 to 309 words and retained a paragraph-long sentence that the metric handled implausibly. This is significant because quantitative gains must be checked against the visible text and legal content. It connects to readability formulas, Spotify, word-count increase, long sentences, visual inspection, metric limitations.

**Evidence anchor:** Pages 37-38 report the unusual length result and caution against literal interpretation of readability scores.

**Boundary:** The authors treat the 28.5-year original estimate qualitatively because it exceeds any plausible educational requirement.

**Connections:** readability formulas; Spotify; word-count increase; long sentences; visual inspection; metric limitations

**Record:** `ssrn-4491043-p54` · `machine-drafted-source-checked`

## 55. Across the eight difficult clauses, smart readers roughly halved the estimated education level and generally improved accessibility

**Location:** Part VI: Discussion, printed pp. 38-39 (PDF pp. 38-39)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 38–39, that across the eight difficult clauses, smart readers roughly halved the estimated education level and generally improved accessibility. The outputs were usually shorter, less complex, and more direct even though sentence count changed inconsistently and average word length moved only modestly. This is significant because multiple indicators and close reading converge on a meaningful accessibility improvement. It connects to aggregate clause results, education level, text complexity, direct language, sentence count, accessibility.

**Evidence anchor:** Pages 38-39 synthesize the quantitative and qualitative accessibility results.

**Boundary:** No metric establishes that consumers will choose to read the improved clauses or correctly understand them.

**Connections:** aggregate clause results; education level; text complexity; direct language; sentence count; accessibility

**Record:** `ssrn-4491043-p55` · `machine-drafted-source-checked`

## 56. Simpler clauses may still go unread, so linguistic improvement does not by itself solve consumer inattention

**Location:** Part VI: Discussion, printed pp. 38-39 (PDF pp. 38-39)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 38–39, that simpler clauses may still go unread, so linguistic improvement does not by itself solve consumer inattention. Even large reductions in grade-level score, word count, and complexity leave unresolved whether users have motivation, time, bargaining options, or reason to seek the information. This is significant because technical capability should not be equated with actual behavioral uptake or market impact. It connects to consumer attention, behavioral uptake, readability, motivation, contract choice, implementation gap.

**Evidence anchor:** Pages 38-39 expressly caution that consumers may decline to read even shorter, easier texts.

**Boundary:** The study does not measure actual readership, comprehension, or purchasing behavior.

**Connections:** consumer attention; behavioral uptake; readability; motivation; contract choice; implementation gap

**Record:** `ssrn-4491043-p56` · `machine-drafted-source-checked`

## 57. Quality was generally high but errors ranged from minor presentation choices to omissions and misleading changes affecting substantive rights

**Location:** Part VI: Discussion, printed pp. 39 (PDF pp. 39)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 39, that quality was generally high but errors ranged from minor presentation choices to omissions and misleading changes affecting substantive rights. Examples include missing agency complaints, overstated privacy consent, and incorrect damages vocabulary rather than merely awkward prose. This is significant because error severity must be classified by legal consequence, not counted as undifferentiated mistakes. It connects to error taxonomy, omissions, misleading summaries, consumer rights, legal terminology, quality assurance.

**Evidence anchor:** Page 39 summarizes the range of errors observed in the clause review.

**Boundary:** The small purposive sample cannot estimate population error rates or a complete severity distribution.

**Connections:** error taxonomy; omissions; misleading summaries; consumer rights; legal terminology; quality assurance

**Record:** `ssrn-4491043-p57` · `machine-drafted-source-checked`

## 58. General-purpose models may lack the domain expertise needed to distinguish legal terms of art from ordinary difficult words

**Location:** Part VI: Discussion, printed pp. 39 (PDF pp. 39)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 39, that general-purpose models may lack the domain expertise needed to distinguish legal terms of art from ordinary difficult words. A prompt that insists on replacing complex vocabulary can improve a formula score while converting “consequential damages” into a different and legally inaccurate concept. This is significant because safe simplification requires legal-semantic awareness, not only language fluency. It connects to domain expertise, terms of art, general-purpose models, prompt-induced error, legal semantics, specialized systems.

**Evidence anchor:** Page 39 ties terminology errors to prompt design and limitations of general models.

**Boundary:** A less aggressive prompt or domain-tuned model may reduce the problem, but the chapter does not test those remedies directly.

**Connections:** domain expertise; terms of art; general-purpose models; prompt-induced error; legal semantics; specialized systems

**Record:** `ssrn-4491043-p58` · `machine-drafted-source-checked`

## 59. Courts may create a new risk if they treat a smart reader's paraphrase as the operative contract when that interpretation favors the seller

**Location:** Part VI: Discussion, printed pp. 39 (PDF pp. 39)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 39, that courts may create a new risk if they treat a smart reader's paraphrase as the operative contract when that interpretation favors the seller. The authors assume the seller-held original is canonical, but a firm may argue that the consumer relied on the simplified version and should be bound by its altered terms. This is significant because consumer assistance can unexpectedly affect contract interpretation and allocation of drafting error. It connects to canonical contract, contract interpretation, consumer reliance, seller opportunism, terms of art, judicial doctrine.

**Evidence anchor:** Page 39 raises the possibility that courts could privilege the smart-reader version in some disputes.

**Boundary:** This is a speculative doctrinal concern rather than a reported case outcome.

**Connections:** canonical contract; contract interpretation; consumer reliance; seller opportunism; terms of art; judicial doctrine

**Record:** `ssrn-4491043-p59` · `machine-drafted-source-checked`

## 60. The study does not meaningfully test bias, toxicity, or hallucination and therefore cannot establish overall smart-reader safety

**Location:** Part VI: Discussion, printed pp. 39 (PDF pp. 39)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 39, that the study does not meaningfully test bias, toxicity, or hallucination and therefore cannot establish overall smart-reader safety. Those risks were muted in the selected outputs but may become more important with frequent, personalized, or varied consumer use. This is significant because absence of salient failures in a small review is not evidence that major model risks are absent. It connects to bias, toxicity, hallucination, scope limitations, AI safety, future research.

**Evidence anchor:** Page 39 states the omitted risk dimensions and the need for further research.

**Boundary:** The paper candidly identifies this as an unexamined domain rather than drawing a negative finding.

**Connections:** bias; toxicity; hallucination; scope limitations; AI safety; future research

**Record:** `ssrn-4491043-p60` · `machine-drafted-source-checked`

## 61. Perfection is the wrong benchmark when most consumers presently proceed with vague understanding after reading none of the legal text

**Location:** Part VII: Summary, printed pp. 40 (PDF pp. 40)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 40, that perfection is the wrong benchmark when most consumers presently proceed with vague understanding after reading none of the legal text. A materially nonmisleading smart-reader output can improve decisions even if it falls short of a lawyer's exact account, because the realistic baseline is nonuse and misperception. This is significant because comparative institutional analysis can support useful imperfect tools without disguising their limitations. It connects to realistic baseline, bounded improvement, nonreadership, material accuracy, comparative institutions, consumer decisions.

**Evidence anchor:** Page 40 contrasts perfect simplification with the practical alternative of unread contracts.

**Boundary:** What counts as materially misleading remains context-sensitive, and some observed errors may cross that line.

**Connections:** realistic baseline; bounded improvement; nonreadership; material accuracy; comparative institutions; consumer decisions

**Record:** `ssrn-4491043-p61` · `machine-drafted-source-checked`

## 62. Smart readers can undermine HIDE by making important terms visible and thereby strengthen competition over contractual quality

**Location:** Part VII: Summary, printed pp. 40 (PDF pp. 40)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 40, that smart readers can undermine HIDE by making important terms visible and thereby strengthen competition over contractual quality. Consumers who can compare understandable rights and risks may make more informed choices, pressure firms to improve terms, and reduce sellers' returns from opacity. This is significant because contract simplification may affect market equilibrium and drafting incentives, not only individual comprehension. It connects to HIDE strategy, competition over terms, market equilibrium, informed choice, seller incentives, contract quality.

**Evidence anchor:** Page 40 explains the potential effects on HIDE, efficiency, competition, and better drafting.

**Boundary:** The study does not directly observe prices, switching, seller drafting, or other market outcomes.

**Connections:** HIDE strategy; competition over terms; market equilibrium; informed choice; seller incentives; contract quality

**Record:** `ssrn-4491043-p62` · `machine-drafted-source-checked`

## 63. Smart readers may be especially useful after a dispute arises because consumers then have focused questions and stronger motivation to examine terms

**Location:** Part VII: Summary, printed pp. 40 (PDF pp. 40)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 40, that smart readers may be especially useful after a dispute arises because consumers then have focused questions and stronger motivation to examine terms. Ex post users can ask about cancellation, payment, switching, remedies, or a specific event rather than consume an undifferentiated full-contract summary. This is significant because the technology's value is not limited to improving ex ante assent. It connects to ex post interpretation, consumer disputes, targeted questions, contract remedies, motivation, Q&A systems.

**Evidence anchor:** Page 40 identifies tailored ex post review as a separate consumer use case.

**Boundary:** Focused answers still require accuracy and may approach individualized legal advice in higher-stakes disputes.

**Connections:** ex post interpretation; consumer disputes; targeted questions; contract remedies; motivation; Q&A systems

**Record:** `ssrn-4491043-p63` · `machine-drafted-source-checked`

## 64. Smart readers can empower intermediaries and consumer organizations as well as individual readers

**Location:** Part VII: Summary, printed pp. 40-41 (PDF pp. 40-41)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 40–41, that smart readers can empower intermediaries and consumer organizations as well as individual readers. Advocacy groups, reviewers, or other institutions can apply the tools across many agreements, detect recurring terms, and distribute warnings or comparisons at scale. This is significant because collective deployment may overcome the expectation that every consumer personally reads and evaluates every form. It connects to consumer intermediaries, collective action, contract review, scaling, advocacy organizations, information dissemination.

**Evidence anchor:** Pages 40-41 extend the potential users of smart readers beyond individuals.

**Boundary:** The chapter does not test organizational workflows, accountability, or the accuracy of large-scale automated review.

**Connections:** consumer intermediaries; collective action; contract review; scaling; advocacy organizations; information dissemination

**Record:** `ssrn-4491043-p64` · `machine-drafted-source-checked`

## 65. Results from nonspecialized models with little domain training should be treated as a lower bound on the potential of dedicated smart readers

**Location:** Part VII: Summary, printed pp. 41 (PDF pp. 41)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 41, that results from nonspecialized models with little domain training should be treated as a lower bound on the potential of dedicated smart readers. Specialized systems could detect problematic terms automatically, warn users, score and benchmark contracts, compare providers, and answer concrete questions. This is significant because the tested simplification task captures only an early subset of possible consumer-contract functionality. It connects to domain specialization, unfair-term detection, contract benchmarking, provider comparison, question answering, capability frontier.

**Evidence anchor:** Page 41 identifies advanced functions and explains why the model choice suggests room for improvement.

**Boundary:** Calling the results a lower bound assumes specialization improves fidelity and does not create new biases or capture risks.

**Connections:** domain specialization; unfair-term detection; contract benchmarking; provider comparison; question answering; capability frontier

**Record:** `ssrn-4491043-p65` · `machine-drafted-source-checked`

## 66. Company influence, adversarial drafting, bias, and nonneutral errors must remain central as smart readers gain market power

**Location:** Part VII: Summary, printed pp. 41 (PDF pp. 41)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 41, that company influence, adversarial drafting, bias, and nonneutral errors must remain central as smart readers gain market power. Firms may try to shape outputs or conceal terms from automated analysis, and errors can systematically favor some actors rather than wash out randomly. This is significant because a successful consumer intermediary becomes a target for capture and strategic manipulation. It connects to model capture, adversarial contracts, systematic bias, corporate influence, consumer protection, AI governance.

**Evidence anchor:** Page 41 warns about accuracy, capture, bias, and firm efforts to circumvent the technology.

**Boundary:** The chapter identifies these risks but does not empirically measure them or prescribe a complete governance framework.

**Connections:** model capture; adversarial contracts; systematic bias; corporate influence; consumer protection; AI governance

**Record:** `ssrn-4491043-p66` · `machine-drafted-source-checked`

## 67. Massive open-source models may reduce the risk of invisible corruption in consumer smart readers

**Location:** Part VII: Summary, printed pp. 41 (PDF pp. 41)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on page 41, that massive open-source models may reduce the risk of invisible corruption in consumer smart readers. Inspectable or plural model ecosystems can make covert manipulation harder than reliance on one closed provider whose training, instructions, or commercial relationships are opaque. This is significant because technical openness can be part of the institutional response to intermediary capture. It connects to open-source models, transparency, model corruption, provider concentration, auditability, consumer trust.

**Evidence anchor:** Page 41 offers open-source models as one possible limit on invisible model corruption.

**Boundary:** Openness alone does not guarantee security, accuracy, representative governance, or resistance to manipulation.

**Connections:** open-source models; transparency; model corruption; provider concentration; auditability; consumer trust

**Record:** `ssrn-4491043-p67` · `machine-drafted-source-checked`

## 68. Current-generation smart readers have arrived as cheap, effective, and scalable aids for the mass of contracts and privacy policies that otherwise go unread

**Location:** Part VII: Summary, printed pp. 40-41 (PDF pp. 40-41)

Professors Yonathan Arbel and Shmuel Becher claim, in “How Smart Are Smart Readers? LLMs and the Future of the No-Reading Problem” on pages 40–41, that current-generation smart readers have arrived as cheap, effective, and scalable aids for the mass of contracts and privacy policies that otherwise go unread. They do not replace careful legal review, but their demonstrated accessibility and generally strong fidelity cross a practical threshold that should reshape academic and policy assumptions. This is significant because law should no longer treat consumer nonreadership as a technologically fixed fact. It connects to paradigm shift, consumer contracts, scalable legal technology, no-reading problem, policy adaptation, human lawyers.

**Evidence anchor:** Pages 40-41 state the final comparative conclusion and call for a law-and-policy paradigm shift.

**Boundary:** The conclusion is bounded to routine unread documents and remains conditional on controlling material errors, bias, and capture.

**Connections:** paradigm shift; consumer contracts; scalable legal technology; no-reading problem; policy adaptation; human lawyers

**Record:** `ssrn-4491043-p68` · `machine-drafted-source-checked`