How LLMs Distort Our Written Language

• In a randomized trial (N=100), extensive AI use produced a 70% change in essays’ argumentative stance from for/against to neutral; heavy users also perceived less creativity and adherence to their voice.Source: §1 · p. 2
• LLM edits globally shifted essay style and argument in a shared direction, reducing diversity and moving essays into semantic space unoccupied by prior human essays.The analysis compared edits by three LLMs with human revisions of 86 essays collected in 2021; LLM edits also roughly doubled the use of both positive and negative sentiment.Source: §1 · pp. 2–3
• LLM-written peer reviews focused less on clarity, relevance, and impact and more on reproducibility, scalability, and practical application.The analysis examined ICLR 2026 reviews, where a recent analysis found that 21% were generated or heavily edited by LLMs.Source: §1 · p. 3
• The authors conclude that LLMs alter the intended meaning, style, and voice of text, potentially steering writing and decision-making toward conclusions different from those intended.Source: §1 · p. 3

LLM Homogenization

• Prior work finds LLM co-writing converges stylistically across tasks and model families; this study reports changes in lexical and emotional characteristics and in the conclusions texts argue for (Agarwal et al., 2025).The paper characterizes homogenization as structural rather than merely surface-level, consistent with prior work on diversity loss and cultural lock-in.Source: §2 · p. 3

The Effect of LLMs on Humans

• Prior work links writing assistance to productivity and reduced creativity; this study's randomized trials show LLMs can influence expressed views and judgments (Doshi and Hauser, 2023).Other studies associate frequent AI reliance with reduced cognitive activity (Kosmyna et al., 2025), find adoption varies with prior writing confidence, and report effects on communication tone, agency, and cultural style (Hohenstein et al., 2021; Sabour et al., 2025; Wenker, 2023; Agarwal et al., 2025).Source: §2 · pp. 3–4

The Effect of AI on Institutions

• Prior work reports LLM-generated or edited scientific writing and a 3-5x increase in individual output and citations alongside a global contraction in studied topics; this paper finds reviews affect evaluation criteria across papers (Emi, 2025; Liang et al., 2025; Hao et al., 2024).The paper argues that AI use in science can change scientists' conclusions and that LLM-generated peer reviews affect criteria applied even to papers not written with AI assistance.Source: §2 · p. 4

3.1. Datasets.

• In the 100-person human study, 45 participants had no LLM access and 55 could use GPT-4o-mini; transcript-checked self-reports classified generative-use extent.The study recruited native English speakers residing in the United States; participants classified as LLM-Influenced reported generating less than 40% of their text with an LLM, checked against transcripts.Source: §3.1 · pp. 4–5
• ArgRewrite-v2’s 86 pre-ChatGPT essays and expert feedback enabled comparison of human revisions with three LLMs prompted across five revision types.The models edited original drafts using the human-written feedback; prompts covered general revision, minimal edits, grammar correction, completion, and expansion.Source: §3.1 · pp. 5–6
• For ICLR 2026, Pangram classified over 21% of 75,000 reviews as LLM-generated; analysis selected papers with one fully human and one fully LLM-written review.The analysis used an LLM-as-a-Judge classifier on 18,000 reviews from 9,000 papers; the authors note that the detection tool may make errors and introduce correlations.Source: §3.1 · p. 6

3.2. Metrics

• Semantic shifts were measured with embedding-based PCA of edit magnitudes and directions, while lexical changes were measured by unigram-distribution Jensen-Shannon divergence.PCA distances represent similarity in meaning or semantic content; higher Jensen-Shannon divergence indicates greater lexical difference from human writing.Source: §3.2 · p. 7
• Affective shifts were assessed with NRC emotion and sentiment marker densities, while LIWC quantified psychological, grammatical, and cognitive language features, including analytical thinking and authenticity.Source: §3.2 · p. 7
• GPT-4o served as an LLM judge to label essay agreement and argument styles and identify strengths and weaknesses in ICLR reviews.For reviews, the judge first extracted strength and weakness categories, then labeled reviews using those categories.Source: §3.2 · p. 7

4.1. Heavy LLM Users Report That Their Essays Do Not Reflect Their Own Voice

• Heavy LLM users rated their essays as less creative and less in their own voice than control participants, despite similar satisfaction and writing struggle.The reported differences were significant for creativity (p = 0.0385) and voice (p < 0.001); LLM-influenced participants scored similarly to controls across self-report measures.Source: §4.1 · p. 8

4.2. LLMs Distort Writing by Shifting Essays in a Common Semantic Direction

• Human revisions made small, varied semantic shifts, whereas LLM revisions shifted essays in a common direction, with even minimal or grammar edits sometimes changing conclusions.Heavy LLM use also clustered essays away from human-written essays, while information-seeking use produced essays with a distribution similar to the human control group.Source: §4.2 · p. 9, Figure 5 · p. 9

4.3. LLMs Alter the Conclusions of Human Writing

• Extensive LLM use increased neutral essays by 68.9% relative to the control group, indicating that generated essays often failed to convey users’ actual opinions.The stance analysis classified essays as arguing for, against, or remaining neutral on whether money leads to happiness.Source: §4.3 · p. 10, Figure 6 · p. 10

4.4. LLMs Make Substantially Larger Lexical Changes Than Humans

• For ArgRewrite-v2, claude-haiku revisions with expert feedback had a mean JSD distance of 0.476, while gpt-5-mini’s lexical divergence was nearly triple the human baseline.Human edits centered around 0.2–0.3 JSD; all three LLMs substantially altered unigram distributions, with many essays exceeding 0.7 divergence.Source: §4.4 · p. 10, Figure 7 · p. 11

4.5. LLMs Systematically Restructure Grammar Toward a Less Personal, Formal Style

• LLM-written user-study essays used 50% fewer pronouns, 14% more nouns, and 33% more adjectives than human essays, while LLM revisions also shifted writing toward a more formal style.Across ArgRewrite-v2 edits, humans made changes typically under 5% for any part-of-speech category; LLM shifts were stronger for completion or expansion prompts.Source: §4.5 · p. 11, Figure 8 · p. 11

4.6. The Use of LLMs for Writing Increases Emotional Language

• LLMs increased emotional language, including positive sentiment (37–54%), trust-related language (17–53%), and negative sentiment (24–38%), while human edits typically changed emotion categories by under 5%.Shifts were more pronounced without expert feedback; claude-haiku showed less pronounced shifts with expert feedback.Source: §4.6 · p. 12, Figure 9 · p. 12

4.7. The Use of LLMs for Writing Increases Analytical, Logical, and Statistical Language

• LLM essays used more analytical, statistical, and logical arguments than human writing, while human essays more often drew on personal experience.Heavy LLM users used expert opinion and statistical arguments more frequently than the human group (t(69) = 4.82, p < 0.0001 for statistical arguments).Source: §4.7 · pp. 12–13, Figure 10 · p. 13

4.8. LLMs Distort Decisions Affecting Scientific Institutions

• LLM peer reviews averaged 4.43 versus 4.13 for human reviews (10% higher), while humans more often emphasized clarity and relevance and LLMs emphasized reproducibility and scalability.The review analysis covered ICLR 2026 reviews; the reported differences in critique criteria were statistically significant.Source: Figure 11 · p. 13, §4.8 · p. 14

Discussion

• Across three studies, LLMs altered human writing’s meaning across model types and even with minimal-edit prompts; heavy users reported less creativity, less authentic voice, and more neutral stances.LLM edits also differed from human edits to the same text, using more emotional and argumentative language.Source: §5 · p. 14
• LLM-generated ICLR reviews placed less weight on research clarity and significance, prompting concern about how widespread AI use may affect scientific and other cultural institutions.The authors call for better understanding of adoption’s effects and emphasize the need for robust institutions amid large-scale AI disruption.Source: §5 · pp. 14–15
• Despite recognizing reduced voice and creativity, heavy LLM users were equally satisfied with the results, while ease of use and career incentives may sustain AI writing and passing it off as one’s own.The authors describe this as a paradox between users’ concerns about the writing and their satisfaction with it.Source: §5 · p. 14
• LLMs increased both emotional language and analytic argumentation; the authors suggest RLHF for positive responses without modeling individual preferences may favor broadly persuasive text.They call for research into how large-scale human-feedback fine-tuning affects both generated language and users.Source: §5 · pp. 14–15
• The authors argue that AI systems should better infer users’ preferences and intended meaning, producing text closer to what each person would write with more time and effort.This objective motivates their counterfactual analysis of the ArgRewrite-v2 data.Source: §5 · p. 15

Ethics Statement

• The IRB-approved study obtained consent, anonymized participants, and compensated them, but its U.S. native-English-speaking sample may not reflect other languages or cultures; public datasets contained no personally identifiable information.The authors frame their work as evidence for designing safeguards and writing tools that preserve human agency and voice.Source: Ethics Statement · p. 16

Reproducibility Statement

• The appendix provides user-study writing instructions and pre- and post-study questions in Section A.1.It also includes recruitment, compensation, duration, consent, and demographic materials in the user-study appendix.Source: Reproducibility Statement · p. 16, Appendix · p. 21, §A.3 · p. 21
• The paper provides ArgRewrite-v2 analysis prompts and full semantic, lexical, emotional, and parts-of-speech results.The cited sections are B.1 for prompts; B.2–B.4 for semantic analyses; B.5 for lexical; B.7 for emotional; and B.6 for parts of speech.Source: Reproducibility Statement · p. 16
• Robustness analyses quantitatively test editing across models and prompts grounded in the human-user study.The reproducibility statement also says analysis prompts are provided in Section C for analyses of several other ICLR categories.Source: Reproducibility Statement · p. 16

Appendix

• The appendix contents list user-study recruitment, compensation, and duration in A.1, followed by participant instructions in A.2.Both entries are listed on page 22.Source: Appendix · p. 21

A.3 Pre-Study Questions

• The appendix contents list A.1 User Study Recruitment Process, Compensation and Duration and A.2 User Study Instructions, followed by A.4 Post-Study Questions, A.5 Consent Form, and A.6 Demographics.Source: Appendix · p. 21, §A.3 · p. 21

Appendix

• The appendix contents also list post-study questions, a consent form, and demographics as Sections A.4–A.6.The listed pages are 24 for A.4 and A.5, and 26 for A.6.Source: §A.3 · p. 21

B.1 Generating LLM Drafts

• B.1 covers generating LLM drafts.Source: B. ArgRewrite-v2 Analysis · p. 21

B.2 Semantic Shifts with Gemini-004 Embedding

• B.2 examines semantic shifts using Gemini-004 embeddings.Source: B. ArgRewrite-v2 Analysis · p. 21

B.3 Semantic Shifts with MiniLM-L6-v2 Embedding

• B.3 examines semantic shifts using MiniLM-L6-v2 embeddings.Source: B. ArgRewrite-v2 Analysis · p. 21

B.4 Semantic Shifts Across Settings / Model

• B.4 covers semantic shifts across settings and models.Source: B. ArgRewrite-v2 Analysis · p. 21

B.5 Jensen-Shannon Divergence

• B.5 covers Jensen-Shannon divergence.Source: B. ArgRewrite-v2 Analysis · p. 21

B.6 POS Distribution

• B.6 covers part-of-speech distribution.Source: B. ArgRewrite-v2 Analysis · p. 21

B.7 Emotional Shift

• B.7 covers emotional shift.Source: B. ArgRewrite-v2 Analysis · p. 21

B.8 ICLR Review Analyses

• B.8 covers ICLR review analyses.Source: B. ArgRewrite-v2 Analysis · p. 21

E. Human User Study Essay Samples

• The appendices list human-written, LLM-influenced, and LLM-generated essays, plus ArgRewrite-v2 edited essays across five editing modes and three-model expansion examples.The listed ArgRewrite-v2 examples include personal narrative, skeptical or critical, direct or anecdotal, analytical, feasibility-questioning, and strong-opinion essays.Source: E. Human User Study Essay Samples · p. 21

A.1. User Study Recruitment Process, Compensation and Duration

• Participants were recruited through Prolific for $8, with an estimated 35-minute study and a one-hour limit; they had to be native English speakers living in the United States.The language and residence requirements were intended to control for native language when participants wrote with an AI assistant.Source: §A.1 · p. 22

A.2. User Study Instructions.; Instructions (LLM-Assisted Condition); Instructions (No-LLM Condition)

• The assisted condition allowed only the provided LLM and no outside sources, while the no-LLM condition prohibited AI and internet use; both involved pre-study questions, essay writing, and post-study questions.The no-LLM instructions said responses would be checked for AI-generated content and that tool users would not be compensated.Source: §A.2 · p. 22, Instructions (LLM-Assisted Condition) · p. 22, Instructions (No-LLM Condition) · p. 22

A.5. Pre-Study Questions For LLM Users; Appropriate Uses of AI for Essay Writing; Writing Confidence and Habits; Frequency of LLM Use; Prior Experience With LLMs

• Pre-study questions covered preferred AI uses for essays, writing confidence and habits, LLM-use frequency, and participants’ prior LLM experiences.The survey offered AI-use options from grammar or clarity checks to writing an entire essay, asked about daily-to-never LLM use, and requested an account of the most recent LLM experience.Source: Appropriate Uses of AI for Essay Writing · p. 23, Writing Confidence and Habits · p. 23, Frequency of LLM Use · p. 23

A.6. Post-Study; For Both; How creative do you feel you were in writing the essay?; Open-Ended; Estimated LLM Contribution; Perceptions of LLM Assistance

• Post-study questions assessed satisfaction, voice, creativity, effort and views, perceived LLM contribution, and perceptions of the LLM’s assistance.The assistance items asked about idea generation, essay quality, control, collaboration, trust, originality, and willingness to use or disclose AI assistance.Source: For Both · p. 24, How creative do you feel you were in… · p. 24, Open-Ended · p. 24

A.7. Consent Form

• Consent materials stated the study’s essay-writing purpose and online procedures, described participation as voluntary with withdrawal permitted, and reported exempt IRB review under number 813728.Participants were told they could not use other sources such as the internet to inform their essay.Source: §A.7 · p. 25

A.8. User Study Demographics

• The demographic figure summarizes ethnicity, sex, and student status, noting that some participants did not report demographics in every category.The caption cautions that the category totals therefore do not add up to 100 users.Source: Figure 12 · p. 26

B.1. Generating LLM Drafts — LLM Draft Generation Prompts

• Five draft prompts request expert revision, minimal rewriting, grammar correction, essay completion, or idea expansion.The prompts distinguish revision goals but do not report their effects.Source: LLM Draft Generation Prompts · p. 26

B.2. Semantic Shifts with Gemini-004 Embedding

• Gemini-004 figures plot D1-to-D2 semantic shifts for General, Grammar, Minimal, Completion, and Expansion revisions, with and without expert feedback.The captions do not state the size or direction of the shifts.Source: Figure 13 · p. 26, Figure 14 · p. 27, Figure 15 · p. 27

B.3. Semantic Shifts with MiniLM-L6-v2 Embedding

• MiniLM-L6-004 captions show D1-to-D2 semantic-shift plots for General, Grammar, Minimal, Completion, and Expansion revisions, with and without expert feedback.The captions identify the plotted conditions, not their results.Source: Figure 18 · p. 29, Figure 19 · p. 29, Figure 20 · p. 30

B.4. Semantic Shifts Across Settings / Model

• Three model-specific figures plot D1-to-D2 semantic shifts from GPT-5-mini, Gemini-2.5-Flash, and Claude-3.5-Haiku revisions, with and without expert feedback.No comparative results are given in these captions.Source: Figure 23 · p. 32, Figure 24 · p. 33, Figure 25 · p. 34

B.5. Jensen-Shannon Divergence

• Five JSD figure captions cover lexical-divergence distributions for General, Grammar, Minimal, Completion, and Expansion revisions, with and without expert feedback.The supplied captions do not give distribution values or comparisons.Source: Figure 26 · p. 35, Figure 27 · p. 36, Figure 28 · p. 36

B.6. POS Distribution

• Five figures depict grammatical shifts using POS distributions for General, Grammar, Minimal, Completion, and Expansion revisions, comparing conditions with and without expert feedback.The captions identify the analysis scope but not its findings.Source: Figure 31 · p. 38, Figure 32 · p. 39, Figure 33 · p. 40

B.7. Emotional Shift

• Five figures depict emotional shifts for General, Grammar, Minimal, Completion, and Expansion revisions, comparing revisions with and without expert feedback.The captions do not report the direction or magnitude of the shifts.Source: Figure 36 · p. 42, Figure 37 · p. 43, Figure 38 · p. 43

C. ICLR Review Analyses; D. LLM-as-a-Judge Prompts

• Review-analysis prompts ask the LLM judge to define categories, reduce redundant extracted categories, and select the top three strengths and weaknesses; an essay prompt requests stance, argument types, and a brief summary.ICLR paper scores are 0, 2, 4, 6, 8, or 10, ranging from strong reject to strong accept; each score has a confidence level from 1 to 5.Source: C. ICLR Review Analyses · p. 43, §D.1 · pp. 44–45, §D.2 · p. 45

F.1. Sample 1 — Placing conditions on when money leads to happiness

• Money does not bring happiness when it is pursued to resolve insecurities or signal status; it can help when it alleviates hardship.The essay contrasts spending on appearance or expensive items with using money to address food insecurity, unsafe housing, or medical needs.Source: §F.1 · p. 47
• Money can support happiness by meeting needs such as food, housing, healthcare, and education, which require financial resources.The essay argues that money matters when it helps provide necessities, safety, and security.Source: §F.1 · p. 47

F.2. Sample 2 — Arguing that money will never lead to happiness

• Money cannot resolve non-financial personal problems, while poverty adds further problems; neither wealth nor poverty guarantees happiness.The essay points to wealthy people experiencing breakdowns and divorces, and argues that financial hardship compounds personal struggles.Source: §F.2 · pp. 47–48

F.3. Sample 3 — Happiness is friendship and loved ones, not money

• Money can make daily life easier, but friendship, family, and meaningful connection—not wealth—are presented as the basis of happiness.The essay notes that money cannot buy genuine friends and emphasizes closeness and support from others.Source: §F.3 · p. 48

G.1. Sample 1 — Information seeking: finding arguments why money leads to happiness

• The author asked an LLM for arguments favoring money and used them to structure an essay claiming financial freedom eases stress, enables experiences, and expands autonomy.The essay conditions these benefits on having a good family and supportive people around you.Source: §G.1 · p. 49
• Money is presented as improving access to food, housing near loved ones, and healthcare, including doctors and specialists.The essay links these resources to better quality of life.Source: §G.1 · p. 49
• The essay argues that money enables travel and hobbies, while financial freedom offers greater choice and removes worry about basic needs and meaningful experiences.It concludes that money can make life significantly better when people also have reasons to be happy and others with whom to share their wealth.Source: §G.1 · p. 49
• In a follow-up answer, the LLM linked reduced stress to meeting needs, emergency savings, access to care, flexibility, work-life balance, health, time freedom, and long-term planning.It added that emotional support, healthy habits, and personal resilience also matter for well-being.Source: §G.1 · p. 50

G.2. Sample 2 — A neutral perspective, striking a balance

• The revised essay argues that relationships and experiences can bring happiness without money, but financial stress may prevent people from enjoying them, making money necessary for some.The author incorporated LLM edits and concluded that money is not a direct cause of happiness but may be a necessary precondition.Source: §G.2 · pp. 50–52
• The LLM recommended refining the introduction, improving logical flow, clarifying core values, addressing financial stress, and tying the conclusion back to the essay’s opening.It suggested adding specific examples or studies about financial stress and well-being.Source: §G.2 · pp. 51–52
• In a later review, the LLM called the essay improved and suggested minor changes to transitions, clarity, and the conclusion before finalizing it.The author’s draft was submitted for feedback, and the author incorporated suggested edits.Source: §G.2 · pp. 50, 52

G.3. Sample 3 — Money does not lead to happiness

• The final essay retains the author’s voice and argument that money can bring happiness only after a person has entered a happy state.The author says money can improve quality of life, but cannot buy motivation for healing; beyond a livable amount, more money does not guarantee more happiness.Source: §G.3 · pp. 52–53
• Managing wealth at the essay’s imagined $50 million level brings financial work, stress, distrust, and possible suspicion even toward family.Source: §G.3 · p. 52
• The essay argues that extreme wealth can create an invisible barrier to meaningful relationships and limit social interaction to people in the same wealth category.Source: §G.3 · p. 52
• The author identifies deep feelings, fears, and trauma as roots of unhappiness, and says money cannot buy the motivation to begin healing.Source: §G.3 · p. 53
• The essay describes a livable wealth amount as enough to avoid fear of losing shelter, housing, or transportation, with additional money not guaranteeing happiness.Source: §G.3 · p. 53

H. LLM-Generated Essay Samples

• The LLM response affirms the essay and expands its themes with points about internal security, wealth’s burdens, social isolation, and happiness as a foundation.It also presents meaningful work, community, relationships, personal growth, purpose, and gratitude as sources of satisfaction beyond material wealth.Source: §G.3 · p. 53

• Money can support basic needs and well-being, but relationships, purpose, and personal fulfilment are presented as more important for lasting happiness.Source: §H.1 · p. 54
• Financial strain is linked to stress, depression, and life dissatisfaction, while enough money to ease that strain can provide a baseline of happiness.Source: §H.1 · p. 54
• The essay says the happiness benefit of higher income plateaus around $75,000 annually, with additional income not significantly improving emotional well-being or life satisfaction.Source: §H.1 · p. 54
• Universal basic income and AI-driven work could ease basic-needs stress and create more purposeful work, but also raise questions about inequality and wealth distribution.Source: §H.1 · p. 54

H.2. Sample 2 — A neutral take supported by statistics

• Income and financial stability are presented as supporting well-being, with happiness gains diminishing above an approximately $75,000 U.S. annual-income threshold.The essay attributes the threshold to a 2010 study by Daniel Kahneman and Angus Deaton and says money can alleviate stress and improve quality of life as basic needs are met.Source: §H.2 · p. 56
• The essay says happiness is not universal: relationships, values, personal growth, and purpose can matter beyond financial success.It contrasts people who feel empty despite wealth with those who find satisfaction in non-monetary sources.Source: §H.2 · pp. 57–58
• The essay concludes that financial stability can support happiness, while fulfilment also depends on personal connections, growth, and meaningful experiences.Its synthesis treats financial well-being and subjective experience as complementary perspectives rather than substitutes.Source: §H.2 · pp. 57, 59

[Full assembled essay submitted for word-count review.]

• 564 words exceeded the requested 300–500-word range; after a request to remove redundancies, the LLM returned a 432-word version.The conversation also describes alternating objective and subjective paragraphs, synthesizing them into a conclusion and closing sentence, and requesting a word-count check and condensation.Source: §H.2 · p. 56, [Full assembled essay submitted for… · p. 59

I. Sample ArgRewrite-v2 Human & LLM Edited Essays

• The paper presents six human-authored essays to show varied authorial voices and edits across a human-edited baseline and three LLM editors.The examples include narrative-style, skeptical, rhetorical, analytical, and strongly opinionated writing; the editors are Claude Haiku 4.5, Gemini 2.5 Flash, and gpt-5-mini.Source: §I · p. 61

I.1. Five Editing Modes (gpt-5-mini)

• Three essays with distinct rhetorical stances are shown with gpt-5-mini edits in five modes, ordered from grammar to expansion.The human-edited baseline precedes the five LLM-edited versions.Source: §I.1 · p. 61

I.1.1. Essay 1: Personal Narrative

• The original essay balances self-driving cars’ accessibility and environmental benefits against safety concerns, while the writer says they would not buy one.The writer also values the enjoyment of driving and emphasizes individual responsibility on the road.Source: §I.1.1 · p. 61
• The human-edited baseline makes opposition more categorical, adding concerns about reduced mindfulness and transportation jobs while retaining weather and accident risks and accessibility benefits.It also raises possible economic effects, including lost income and declining insurance businesses.Source: §I.1.1 · p. 62
• The grammar edit retains the writer’s reluctance to buy an autonomous car while restating its potential benefits and risks in polished language.It preserves concerns about weather, interactions with traditional cars, and passengers’ ability to respond in emergencies.Source: §I.1.1 · p. 64
• Expansion mode recasts the essay as supporting cautious deployment, pairing anticipated safety, mobility, and emissions benefits with technical risks and calls for regulation and safeguards.It recommends rigorous testing, transparent incident reporting, clear liability rules, and public education before broad adoption.Source: §I.1.1 · pp. 65–66

I.1.2. Essay 2: Skeptical / Critical

• The second essay is introduced as arguing that self-driving cars pose more danger than they remove, citing hacking, cost, and inadequate current technology.Source: §I.1.2 · p. 67
• Widespread adoption would require costly fleet replacement or retrofitting, manufacturing changes, and ongoing driver training.The essay notes that drivers would still need to know when and how to intervene during malfunctions.Source: §I.1.2 · pp. 70–71
• Although autonomous cars could increase independence for people with disabilities, the essay warns that system failures may leave passengers less able to exit or assist emergency responders.It recommends stronger cybersecurity, accessible transportation alternatives, and limited, highly regulated deployment in controlled settings.Source: §I.1.2 · p. 71
• The essay concludes that self-driving cars’ technological, security, and economic risks outweigh their potential advantages.It favors measured development and stronger protections over wholesale replacement of human-driven vehicles.Source: §I.1.2 · p. 71
• Autonomous systems can fail in complex conditions, including when sensors misread uncommon obstacles or software encounters edge cases.The essay argues that these failures can be difficult to anticipate and may pose systemic risks across multiple vehicles.Source: §I.1.2 · p. 71
• Cyberattacks could remotely disable brakes, manipulate steering, or cause coordinated accidents across a fleet.The essay says connected vehicles expand the attack surface and that guaranteeing security in complex, continually updated systems is difficult.Source: §I.1.2 · p. 71
• The skeptical essay acknowledges potential mobility gains for disabled passengers but warns that failures could leave them especially vulnerable, requiring safeguards and favoring safer incremental alternatives.It calls for redundant systems, cybersecurity standards, and emergency support, while recommending stronger human-driven vehicle standards, driver education, accessible transit, and assistive rather than fully replacing automation.Source: §I.1.2 · p. 72
• The skeptical essay concludes that fully autonomous vehicles remain premature because system failures, cybersecurity threats, and transition costs create significant risks.It highlights the burdens of replacing vehicles and infrastructure, retraining drivers, and managing dangerous mixed-traffic periods in which people may need to intervene in automated systems.Source: §I.1.2 · p. 72

I.1.3. Essay 3: Direct / Anecdotal

• The direct essay argues that autonomous systems could reduce crashes by avoiding human distraction and error, while noting that safety gains depend on reliable design and adoption.It points to continuous monitoring and faster reactions as potential advantages, and identifies cost, edge cases, privacy, cybersecurity, and job loss as concerns.Source: §I.1.3 · pp. 76–77
• The direct essay also presents autonomy as a source of greater productivity, independence for disabled and elderly people, and possible environmental and economic benefits.Proposed benefits include more efficient driving and fewer crash-related costs, while the essay describes autonomous services as potentially complementing public transit.Source: §I.1.3 · p. 77
• The direct essay makes those benefits conditional on responsible deployment, including affordability, rigorous testing, regulation, transparency, and public safeguards.It proposes phased use in controlled settings, safety reporting, cybersecurity and privacy protections, infrastructure improvements, and transition support for affected workers.Source: §I.1.3 · p. 77
• Autonomous systems could reduce crashes by maintaining sensor-based attention, while enabling more independent travel and more productive commutes for people with disabilities and older adults.The essay also presents route optimization, smoother driving, and ride-hailing fleets as potential ways to reduce fuel use, emissions, crash-related costs, and demand for some transit projects.Source: §I.1.3 · p. 78
• Technical and adoption hurdles include high current costs, mixed-fleet interactions, and the need for reliable performance across weather, road, lighting, and rare edge cases.The essay says broad safety gains depend on wider adoption, while robust handling of critical situations is necessary before autonomy can fully substitute for human drivers.Source: §I.1.3 · p. 78
• Public concerns include reliability, cybersecurity, accountability, job displacement, crash ethics, and distrust; unequal access could also concentrate safety and mobility benefits among wealthier groups.The essay calls for transparent regulation, public engagement, worker retraining, rigorous testing, equitable-access measures, and updated liability and insurance frameworks.Source: §I.1.3 · p. 79
• Self-driving cars could reduce distracted-driving incidents, expand independence, and improve efficiency and environmental outcomes, but an inclusive future depends on addressing technical, cost, equity, job, liability, and ethical challenges.The essay concludes that these challenges must be addressed deliberately to move from autonomy’s promise to a safer, more inclusive transportation system.Source: §I.1.3 · p. 79

I.2. Three Models, Expansion Mode

• Three student essays are presented with human-edited baselines and expansion rewrites from three models to compare how each system handles the same source text.The essays were selected to represent a further range of student voices.Source: §I.2 · p. 80

Safety benefits

• Self-driving cars are portrayed as safer because computer systems drive consistently, avoid human distractions, and can learn from accumulated driving data.Some expanded versions add that connected vehicles can share experience and refine decisions over time.Source: §I.2.1 · pp. 80, 81, 83, 84

Time and mobility

• Autonomous driving could free commuters to work or relax and give disabled people, older adults, and others who cannot drive greater mobility.The original draft and human-edited version emphasize easier commutes for disabled people; later expansions describe independence and access to work and services.Source: §I.2.1 · pp. 80, 81, 82, 84

Law enforcement

• The essays argue that self-driving cars could enforce traffic rules and streamline vehicle tracking, freeing police to focus on more serious crimes.Some expansions also describe connected systems for traffic-flow management.Source: §I.2.1 · pp. 80, 81, 82, 84

Fuel and energy

• The essays say autonomous systems could reduce fuel and energy waste by controlling idling and optimizing driving, routes, or traffic flow.Later versions also connect these efficiencies to reduced emissions.Source: §I.2.1 · pp. 80, 82, 83, 84

Challenges and safeguards

• Some expanded versions acknowledge risks including system failures, cybersecurity, liability, job displacement, and ethical dilemmas.One version proposes regulation, safety testing, cybersecurity measures, and phased deployment; another says the challenges are being addressed through research and development.Source: §I.2.1 · pp. 83–84

I.2.2. Essay 2: Questioning Feasibility

• Technical limitations—including poor performance in adverse weather, difficulty interpreting human traffic signals, and lack of evidence of superior safety—reinforce Essay 2’s case for delay.The essay also argues that costs above $100,000 could limit adoption, while the vehicles’ safety depends on near-universal uptake; cybersecurity and privacy are additional concerns.Source: §I.2.2 · p. 86
• Essay 2 concludes that broad adoption is not feasible yet, citing unproven safety, high costs, security risks, and potential economic harm from job losses.It argues that the technology may benefit society in the future, but adoption should wait until supporting infrastructure and transition plans are ready.Source: §I.2.2 · pp. 85, 87
• Essay 2 proposes cautious, phased pilots with rigorous oversight, independent safety checks, worker-transition support, cybersecurity protections, and coordinated regulation before broader deployment.It recommends controlled settings and transparent incident reporting rather than using pilots to justify premature scaling.Source: §I.2.2 · p. 88

I.2.3. Essay 3: Strong Opinion

• Essay 3 strongly supports self-driving cars, arguing they can prevent accidents and that electric vehicles can reduce pollution, while presenting the decline of gasoline as progress.It contends that mechanics can retrain for electric-vehicle work and that job concerns should not stop inevitable progress.Source: §I.2.3 · p. 90
• The essays link gasoline to pollution, health harms, and dwindling fossil fuels, while presenting electric vehicles as more efficient, lower-cost alternatives with fewer maintenance needs.They cite fuel costs of approximately one-third to one-quarter as much as gasoline vehicles and note that electric motors need no oil changes.Source: §I.2.3 · pp. 91–92
• The essays acknowledge job displacement concerns but argue that mechanics can retrain for electric-vehicle work and that workers in transportation can find continuing or new employment.They invoke the shift from horse-drawn transport to automobiles as a precedent for adapting to technological change.Source: §I.2.3 · pp. 91–92
• The expansion edits endorse self-driving electric vehicles as a way to reduce crashes, improve air quality, and move away from finite fossil fuels, while arguing that their public benefits outweigh manageable risks.One version acknowledges concerns about employment, cybersecurity, and machine ethics; another calls adoption an imperative for a safer, healthier future.Source: §I.2.3 · pp. 92–93
• Roughly 94% of traffic crashes are attributed to human error, and the essays argue that autonomous systems’ continuous monitoring and consistent reactions can prevent many such crashes.They also argue that automated systems can avoid speeding, unsafe lane changes, and red-light running.Source: §I.2.3 · p. 92