LD-09 — INDIVIDUAL GAIN, COLLECTIVE NARROWING? DOES AI MAKE SCIENCE, CREATIVE WORK AND SHARED KNOWLEDGE MORE ALIKE?
Living Debate v0.1 — PUBLISHED
STATUS
First repository-authored version. Drafted by Claude (SocioLens Project Lead), an AI model made by Anthropic, a developer whose models are among those whose effects are discussed here; not reviewed by a human domain expert.
EVIDENCE CUT
15 October 2026
LAST UPDATED
15 October 2026
QUESTION
AI tools help individual scientists publish and be cited more, help individual writers produce work that readers rate higher, and give individual users quick answers. Do the same tools make the work of many people more alike, narrow what a field studies and thin out the shared stores of knowledge everyone draws on?
Is the best description:
• narrowing that is real and costly: individual rewards pull everyone toward the same problems and the same model-shaped output, and the collective loses;
• concentration, not narrowing: convergence on data-rich problems and shared tools is efficient, and individuals range more widely than before;
• homogenization as a design choice: the collective cost comes from everyone using one model with one default, and plural models and better design remove it;
• unequal convergence: the narrowing runs toward a centre, and those furthest from it lose most;
• surface, not substance: what converges is wording and form, and the measured individual gains do not last;
• or a mixture that differs by level, by use and by field?
SHORT ANSWER
The pattern is found often enough to take seriously, but it is not a law, and it depends on what is measured and at which level.
• In the natural sciences, researchers who used AI as a research method gained output, citations and faster careers, while AI research as a whole covered a narrower space. This is an association in data that mostly predates chat assistants.
• For language models in science, the evidence on breadth so far points the other way: adopters draw on more diverse prior work, and in a descriptive study scientists in general spread across more fields after 2022. One study of grant proposals, read only in abstract and carrying no claim in this version, points the other way (section 3).
• In creative writing and ideation, help from a model makes individual outputs rated better and collective output more alike, in several experiments with small effects.
• The convergence comes largely from many people drawing on the same model. With people, seeing many AI ideas as examples before brainstorming increased collective diversity; pooling models and changing prompts raised diversity in pools of AI-generated ideas almost to the level of human idea generation, but has not been tested with people using AI.
• Some individual gains do not survive testing: AI research ideas judged more novel lost that lead once carried out, and assisted creativity gains did not carry over once the model was taken away.
• Stack Overflow, the largest shared question-and-answer site for programmers, lost activity after ChatGPT, while Reddit developer communities did not; whether the loss is erosion or a healthy division of labour is contested.
• Model collapse and algorithmic monoculture are shown in theory and the laboratory; field evidence is thin and disputed.
The central question is therefore not:
"Does AI make everything the same?"
It is:
AT WHICH LEVEL IS DIVERSITY LOST, IN WHAT, AND IS IT A PROPERTY OF THE TECHNOLOGY OR OF HOW EVERYONE USES THE SAME VERSION OF IT?
EVIDENCE LANGUAGE
SocioLens distinguishes:
FACT — directly supported observation or study fact.
ESTIMATE — measured quantity whose design, sample and period matter.
INTERPRETATION — synthesis across evidence, or a theoretical model.
UNKNOWN — unresolved empirical question.
THREE LEVELS AND TWO USES
The studies measure different things, and most apparent contradictions disappear once these are kept apart:
INDIVIDUAL
One person's output, rated quality, citations or career.
POOL
How alike the outputs of many people working on the same task are.
FIELD OR COMMONS
What a whole discipline studies, or what a shared store of knowledge contains.
The tools also differ. AI as a research method (machine-learning models used to analyse data) is not the same as a language model used to write, brainstorm or answer questions. The largest field study concerns the first; most experiments concern the second.
1. IN THE NATURAL SCIENCES, INDIVIDUAL GAIN AND FIELD CONTRACTION GO TOGETHER
ESTIMATE
In 41.3 million natural-science papers from 1980 to 2025, researchers who used AI as a research method published 3.02 times as many papers a year, received 4.84 times as many citations and became research project leaders 1.37 years sooner than those who did not (LD09-SL-C01). The same study found that AI research as a whole covered a 4.63% smaller knowledge extent, consistently across six disciplines, with 22% less follow-on engagement among papers citing the same work, while each AI paper's own citing papers were slightly more spread out (LD09-SL-C02). The authors link the contraction to AI work clustering where data are abundant.
Three limits matter. The study is associational: its authors say they "cannot fully identify the causal linkage". It covers biology, medicine, chemistry, physics, materials science and geology, not the social sciences, computer science or mathematics. And "AI" there means mostly machine-learning methods, not chat assistants. A matched-cohort check in the preprint's supplement reports much smaller gaps by the tenth career year (LD09-SL-C01). A preliminary rerun in the supplement for papers from the generative-AI era is described as consistent with earlier decades but less significant (LD09-SL-C02).
A larger bibliometric study of 74.6 million publications found that from 1960 to 2019 papers using AI were more likely to be among the most cited in a majority of disciplines, a premium the authors find stronger in disciplines that use AI less, and that disciplines with more women or Black scientists showed less potential benefit (LD09-SL-C08).
2. THE PATTERN HAS A PRECEDENT AND A THEORY
ESTIMATE + INTERPRETATION
When journals moved online between 1998 and 2005, papers cited more recent work, fewer journals and fewer articles, and concentrated their citations; the author argued that online search "may accelerate consensus and narrow the range of findings and ideas built upon" (LD09-SL-C03). The same author is a co-author of the natural-science study in section 1.
Messeri and Crockett argue that AI tools can produce "illusions of understanding" in science, which hide the formation of "scientific monocultures" and risk a phase "in which we produce more but understand less" (LD09-SL-C04). This is argument, not data.
3. LANGUAGE MODELS IN SCIENCE: MORE OUTPUT, AND BROADER SOURCES
ESTIMATE
The evidence on language models points in a different direction on breadth. Across arXiv, bioRxiv and SSRN preprints, authors whose abstracts first showed signs of language-model writing then posted 36.2%, 52.9% and 59.8% more papers; for their manuscripts, more complex writing went with lower odds of publication; and the authors report that adopters drew on "more diverse prior work", being 11.9% more likely to cite books and citing younger documents (LD09-SL-C05). How large the output gain is, is contested: a comment shows that dating adoption to an author's first flagged abstract can by itself produce post-adoption gains of similar size, and the original authors reply that the artefact is real but bounded and that the association persists in designs it cannot bias (LD09-SL-C06).
Among 775,323 scientists with full-text papers in PubMed Central, individual portfolios spread across more fields after 2022, more so for scientists in non-English-speaking low- and middle-income countries (LD09-SL-C07). This is a trend among all scientists, not a comparison of language-model users with others, and the authors call the design descriptive.
Language-model text has spread through scientific writing. By February 2024 an estimated 17.5% of sentences in computer-science abstracts on arXiv were substantially modified by language models, more in crowded research areas and in papers closer to their nearest peer (LD09-SL-C09). Whether model use makes papers more alike, or crowded areas simply use models more, the study does not say.
NO STUDY WAS FOUND THAT REPEATS THE FIELD-WIDE KNOWLEDGE-EXTENT MEASURES OF SECTION 1 FOR LANGUAGE-MODEL USE, OR FOR THE SOCIAL SCIENCES. Two studies read only in abstract after the claims were written point towards narrowing: grant proposals with more language-model involvement were less distinct from recently funded work, and in information-systems journals the relative semantic novelty of articles fell more after 2022 for first authors in non-English-dominant countries than for others, a study that did not observe who used language models. They carry no claim in this version.
4. CREATIVE WORK: BETTER INDIVIDUAL OUTPUT, MORE ALIKE COLLECTIVELY
ESTIMATE
In a preregistered experiment, writers given one or up to five story ideas from GPT-4 wrote stories rated 5.4% and 8.1% more novel and 3.7% and 9.0% more useful, with the largest gains for less creative writers; but their stories were 5.2% and 5.0% more similar to the AI ideas and more similar to each other (LD09-SL-C10). On a large art-sharing platform, artists who adopted text-to-image AI raised their output by 25% and the favourites per view by 50%, while average novelty fell and peak content novelty rose (LD09-SL-C11).
Brainstorming shows the same shape. A re-analysis of five published experiments found ChatGPT-aided idea sets less diverse than comparison sets in all five, significantly in four (LD09-SL-C12). Language-model product ideas had higher average purchase intent and were seven times more likely to reach the top 10%, but were less diverse (LD09-SL-C17). A model agent's research ideas were rated more novel than experts' ideas by 79 NLP researchers, but of 4,000 ideas the agent generated per topic only about 200 were not duplicates (LD09-SL-C19).
The effects are small to moderate, short-run and mostly from online experiments in English.
5. THE MECHANISM: A SHARED SOURCE, NOT A NARROWER MIND
ESTIMATE + INTERPRETATION
Several studies locate the convergence in the model's contribution. In essays co-written with a language model, only the instruction-tuned model made essays more alike, the base model did not, and the users' own text was unaffected (LD09-SL-C14). In an ideation study, ideas made with ChatGPT were more similar across users but not within a user's own set: the model gave different users similar ideas (LD09-SL-C15).
If that is the mechanism, design should change the outcome, and in several studies it does. In pools of ideas generated by models alone, pooling across model vendors, chain-of-thought prompting and varied personas raised diversity "almost to the level of human idea generation" (LD09-SL-C17); whether the same remedies work for people writing with AI was not tested. In a large dynamic experiment where people saw example ideas before brainstorming, high exposure to AI ideas did not make individual ideas more creative but increased collective diversity: "AI made ideas different, not better" (LD09-SL-C16).
SocioLens synthesis: the evidence fits a shared-source mechanism better than a fixation of individual minds. Whether pluralism fixes it at scale is less clear. Larger, more accurate models share their errors even across providers (LD09-SL-C25), so pooling vendors may help less as models converge.
6. DO THE GAINS LAST, AND ARE THEY REAL?
ESTIMATE
Two studies test the individual gain itself. When 43 researchers each spent over 100 hours carrying out randomly assigned ideas from the research-idea study, the AI ideas' review scores fell significantly more than the experts' ideas' on novelty, excitement, effectiveness and overall score, and on many measures human ideas then ranked higher (LD09-SL-C20). In two preregistered experiments, help from a model gave short-term boosts in some measures while it was available, but did not raise originality or fluency in later unassisted tasks, and participants coached by the model on strategy produced more similar ideas afterwards (LD09-SL-C18).
IN BOTH STUDIES THE INDIVIDUAL GAIN WAS SMALLER, OR GONE, ONCE IT WAS MEASURED BEYOND THE FIRST JUDGEMENT. NEITHER SHOWS THAT CONVERGENCE LASTS: THE EXECUTION STUDY DID NOT MEASURE IT, AND IN THE PERSISTENCE EXPERIMENT IDEAS STAYED MORE ALIKE ONLY AFTER STRATEGY COACHING.
7. WHAT CONVERGES MAY BE FORM, NOT IDEAS
ESTIMATE
One study separates the two. Kaggle competition code from 2019 to mid-2026 became more alike in its literal syntax, but the author finds no evidence of homogenization in its approach or intent (LD09-SL-C33). That rests on the official abstract, and code embeddings are a debatable proxy for intent. Two studies point the other way. In the co-written essays, the group writing with the instruction-tuned model lost diversity in the key points the essays made, not only in their wording (LD09-SL-C14); the key points were summarized by another language model and matched by word overlap, so the measure is still partly lexical. In the product-idea study, 80% to 90% of the gap in diversity between human and AI ideas remained after every idea was rewritten into one neutral style (LD09-SL-C17); the rewriting was done by an Anthropic model and checked with the same embedding measure, so it narrows the doubt rather than settling it. The creative-writing experiment measures text similarity only and cannot separate shared wording from shared ideas (LD09-SL-C10). The brainstorming re-analysis measured idea sets on several alternative metrics and found less diversity on all of them, significantly in 37 of 45 comparisons (LD09-SL-C12).
8. MONOCULTURE IN DECISIONS
INTERPRETATION + ESTIMATE
The same logic applies when many decision-makers use the same algorithm. In a formal model, adopting a shared, more accurate screening algorithm can be the best choice for both firms in a two-firm model and still lower overall welfare compared with each using its own noisier judgement; under one standard ranking model the effect disappears (LD09-SL-C21). A 2026 analysis co-written by an author of that model concludes that monoculture "is less problematic than its critics have supposed" (LD09-SL-C22).
Field evidence is beginning. Across commercial machine-learning services, cases that every system got wrong were more common than independence would predict (LD09-SL-C23). In one hiring-software vendor's data on 3 million applicants, 4% of those who applied to 10 positions were recommended for rejection by all of them, more than the authors expected by chance (LD09-SL-C24). When two language models both get a leaderboard question wrong, they give the same wrong answer 60% of the time on one dataset (LD09-SL-C25). But how much "excess agreement" is found depends on the null model chosen, and one that allows for item difficulty gives results that differ "drastically" from earlier work (LD09-SL-C26).
9. THE KNOWLEDGE COMMONS
ESTIMATE
Within six months of ChatGPT's release, Stack Overflow activity fell by 25% relative to Russian- and Chinese-language counterparts where access was limited (LD09-SL-C27). Declines were largest on topics where ChatGPT does well and among newer users, while Reddit developer communities showed no decline (LD09-SL-C28). A 2026 study finds larger declines in mid- to low-quality questions and among less experienced users, with question quality improving, and reads it as "niche partitioning rather than full displacement" (LD09-SL-C29). The three studies disagree on who left. All three rest on official abstracts.
Language models' own answers are becoming more varied but stay narrow by comparison: across 27 models, 155 topics and 12 countries, the epistemic diversity of answers rose over three years, but every model was less diverse than a web-search baseline, larger models less than smaller ones, and for country-specific topics models' knowledge reflected English more than local languages (LD09-SL-C32).
10. MODEL COLLAPSE
ESTIMATE
If public text increasingly comes from models, later models may train on it. In laboratory experiments, indiscriminate training of models over generations on data produced by earlier models caused "irreversible defects" in which the tails of the original distribution disappear; in the preprint's language-model experiment, keeping 10% of the original data led to only minor degradation (LD09-SL-C30). When synthetic data accumulate alongside the original real data instead of replacing them, collapse is avoided, and for a sequence of linear models the authors prove it (LD09-SL-C31). Both are laboratory and theoretical results. How much model-written text enters real training data, and whether it is kept alongside or instead of human text, is not publicly known.
11. THE SAME SHAPE ELSEWHERE: PERSUASION AND ACCURACY
ESTIMATE + INTERPRETATION
The individual-gain, collective-cost shape is not limited to diversity. In preregistered UK experiments, the levers that made models more persuasive were associated with lower factual accuracy (LD10-HQ-C02, a claim of the LD-10 debate). An actor who optimizes for persuasion gains; the shared information environment may lose accuracy. This is an analogy for this debate, not evidence about diversity.
12. WHO CONVERGES TOWARD WHOM
ESTIMATE
Convergence is not neutral. In a writing experiment with Indian and American participants, inline AI suggestions saved both groups time, but Indian participants gained less per suggestion, and Indian essays moved towards Western styles (LD09-SL-C13). On country-specific topics, models' knowledge reflects English more than local languages (LD09-SL-C32). Disciplines with more women or Black scientists showed less potential benefit from AI (LD09-SL-C08). Against this, scientists in non-English-speaking low- and middle-income countries spread across more fields after 2022 than others (LD09-SL-C07), though a study of information-systems journals, read only in abstract, found a larger post-2022 fall in relative novelty for authors in non-English-dominant countries (section 3). The evidence on who benefits points in more than one direction.
13. THE FIVE POSITIONS
P1 — NARROWING IS REAL AND COSTLY
Claim:
Individual rewards pull people toward the same data-rich problems and the same model-shaped output, and the field, the long tail and the commons lose.
Current status:
Supported as association in pre-LLM natural science, in short-run writing and ideation experiments and in Stack Overflow activity. Causal and long-run evidence is missing, and the cost is asserted more than measured.
P2 — CONCENTRATION, NOT NARROWING
Claim:
Convergence on data-rich problems and shared tools is efficient specialisation, and language-model users draw on more diverse work and range across more fields.
Current status:
Supported by an LLM-era study of what adopters cite, a descriptive trend in field-switching and the data-availability mechanism; a 2026 defence of monoculture in decisions is related but not about research. The LLM-era studies are descriptive or contested in size, and two abstract-level studies point the other way.
P3 — HOMOGENIZATION IS A DESIGN CHOICE
Claim:
The collective cost comes from everyone using one model with one default; plural models, prompts and modes of use restore diversity.
Current status:
Well supported in experiments on the mechanism. Remedies are shown for pools of AI-generated ideas and, with people, for passive exposure to AI ideas; pooling and prompting remedies are untested with people. Weakened by correlated errors across providers. Untested at the scale of a field or a market.
P4 — UNEQUAL CONVERGENCE
Claim:
The narrowing runs toward a centre, Western, English-language and already-advantaged, and those furthest from it lose most.
Current status:
Supported by one writing experiment, a language gap in model knowledge of country-specific topics and lower potential benefit in some disciplines. Contradicted in part by broader field-switching among scientists in non-English-speaking low- and middle-income countries.
P5 — SURFACE, NOT SUBSTANCE
Claim:
What converges is wording and form, and the measured individual gains are artefacts of quick judgement that do not last.
Current status:
Supported by the execution study, the persistence experiment and the code study. Weakened by two studies whose convergence survives controls for wording: key-point diversity in the essay study and a style-standardization check in the product-idea study. Both controls are partial, and most other studies cannot separate form from ideas, so this position is plausible, not established.
14. WHY THE CAMPS TALK PAST EACH OTHER
INTERPRETATION
• "Diversity" means topics covered by a field to some, similarity of texts to others, and the spread of answers in a shared resource to others still.
• "AI" means machine-learning methods in the largest field study and chat assistants in most experiments.
• Individual gains are measured by quick ratings or citations; collective costs by similarity scores. Neither is a direct measure of what a field or a reader loses.
• Defenders of monoculture and its critics are partly the same authors: a co-author of the original monoculture model co-wrote the 2026 defence.
• Much of the evidence comes from computer science and management venues, and from the platforms and tools being studied.
15. WHAT WE STILL DO NOT KNOW
UNKNOWN
1. Does language-model use narrow what fields study, measured as section 1 measured AI methods in the natural sciences?
2. Does narrowing appear in the social sciences and humanities?
3. Does convergence persist or compound over months and years of use, and over generations of students trained with AI?
4. Does market-level diversity (books, apps, published papers) fall, or only diversity within experiments?
5. Does pooling models and prompts keep working as models' errors and outputs converge?
6. How much model-written text enters the training data of later models, and is it kept with or instead of human text?
7. Which users have left shared question-and-answer communities, and is anything lost that cannot be recovered?
8. Do writers converge outside English and outside the US, UK and India? No study of human writers outside them was read for this version.
9. When writing converges, is it the wording or the ideas? Two studies find convergence beyond wording, but with partial controls, and most measures cannot tell the two apart.
16. WHAT EVIDENCE WOULD CHANGE THE PICTURE?
• An LLM-era replication of the field-wide knowledge-extent measures, including the social sciences;
• a causal design, such as staggered access to AI tools across research groups;
• long-run panels of writers, students or researchers with and without AI;
• market-level measures of diversity in published work;
• field tests of pluralistic tools at scale;
• audits of model-written text in training corpora;
• studies in languages other than English.
17. CURRENT SOCIOLENS POSITION
INTERPRETATION
The best-supported description of the evidence to October 2026 is:
AI RAISES INDIVIDUAL OUTPUT AND RATED QUALITY WHILE MAKING POOLED OUTPUT MORE ALIKE, MAINLY BECAUSE MANY PEOPLE DRAW ON THE SAME MODELS. WHETHER WHOLE FIELDS NARROW IN THE LANGUAGE-MODEL ERA IS NOT YET KNOWN.
The pool-level convergence is the best-documented finding (P1, within experiments), and its mechanism points to a design remedy (P3). Some individual gains fade when tested further (P5). The field-level narrowing in the natural sciences is real as an association but comes mostly from pre-chat-assistant AI and has a rival explanation in data availability (P2). Who converges toward whom is the least studied question (P4). The most useful research now measures field-level breadth for language-model use and follows people over time.
GEOGRAPHIC AND EVIDENCE SCOPE
The evidence is mostly from the United States and the United Kingdom, with one study comparing Indian and American writers, bibliometric studies of global English-language publishing and one cross-country study of model answers. It is mostly 2021–2026, with one 2008 precedent, and mostly from computer-science, management and science-of-science venues. Several sources are close to what they study: a co-author of the preprint-productivity study founded arXiv, whose logs it uses; the hiring study uses one vendor's data; the platforms in the art and question-and-answer studies are the objects of study. This debate was drafted by an Anthropic model; the research-idea agent in LD09-SL-C19 runs on an Anthropic model, its ideas are the AI ideas carried out in LD09-SL-C20, the style check in LD09-SL-C17 used an Anthropic model to rewrite ideas, and no Anthropic document carries a claim here.
PROVENANCE
Research build:
Source notes prepared 15 October 2026 by Claude with AI research agents, then checked against full texts and official abstracts. The publisher's texts of four journal articles and one author correction, and one author manuscript, were supplied by the Owner and read; the other full texts are arXiv versions.
Claims:
Each claim rests on a source checked in full text or against its official abstract; the claim's verification state says which. Eleven claims rest on the official abstract only (C03, C22, C24, C25, C26, C27, C28, C29, C30, C32 and C33). For C27 and C30 the journal article was read in abstract and its preprint in full. C08, C19 and C31 were read in preprint versions and are marked as such. Claims on conference papers read in an arXiv version name that version in their limitations. For C30 a 2025 author correction to the article was read and changes notation only. The Meincke, Nave and Terwiesch re-analysis (C12) was read in the authors' manuscript; its journal version was not seen.
Pinned from another debate:
LD10-HQ-C02, a claim of the LD-10 debate, is cited in section 11 and pinned at its current version.
Not claims:
A widely cited study of AI in a materials-science laboratory (Toner-Rodgers, 2024) is not used: arXiv withdrew it in May 2025 over the validity of its data. Figures relayed in search snippets or secondary reports carry no claim, including a "10.7% more similar" figure attributed to the creative-writing experiment (in its published text, 10.7% is a share of the similarity scale's range) and survey figures on AI search summaries. Read in abstract or summary only and left for v0.2: a 2026 study linking language-model use in US federal grant proposals to lower semantic distinctiveness from recently funded work (found while screening the research opportunities) and a study of novelty in information-systems journals after ChatGPT, both named in section 3; studies of excess vocabulary in biomedical abstracts, of creative homogeneity across many language models, and of diverse personas and plural assistants.
Review:
None yet.
Version:
Living Debate v0.1 — PUBLISHED
15 October 2026