LD-08 — WHO CAN KNOW ABOUT AI? EPISTEMIC DEPENDENCE AND JURISDICTION OVER THE FIELD
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 data, funding and access programmes are among those discussed here; not reviewed by a human domain expert.
EVIDENCE CUT
16 October 2026
LAST UPDATED
16 October 2026
QUESTION
AI firms build the largest models, hold the compute, collect the records of how people use AI, fund research on its effects and decide which outside researchers may test their systems. How independent can research on AI's social effects be under these conditions, and who is gaining the authority to say what AI does to society?
Is the best description:
• dependence that is structural and corrosive: firms control the inputs, so they shape what can be studied and how;
• a new evidence genre: firms' own data on use are valuable if they are clearly labelled and read with care;
• access as the remedy: law can give independent researchers the data they need;
• a contest over jurisdiction: computer science, the social sciences and the policy world each claim the question, and firms enrol some disciplines rather than others;
• presence, not capture: industry ties are common, but no one has shown that they change what research on AI's effects finds;
• or a mixture that differs by kind of dependence?
SHORT ANSWER
Dependence is real and well documented for technical AI research. For research on AI's social effects, SocioLens reads the evidence as showing dependence mainly through data, access and funding rather than through who writes the papers.
• Academic laboratories now build few of the most compute-heavy models, and ties to large firms are common in leading machine-learning research but rarely declared as conflicts of interest.
• Research on the ethics and social effects of AI is still written mostly by academics: few responsible-AI papers have industry co-authors.
• In technical AI research on arXiv, the societally framed writing is now done mostly by computer-science-only teams; this says nothing about social-science journals.
• Firms hold the only data on use of their systems at scale; independent routes such as public logs and data donation exist but are small and self-selected. Firms publish aggregates and classify the data with language models; Anthropic validates its classifications in-house, and its figures shift with the model doing the classifying.
• Anthropic's researcher-access programme gives API credits at its discretion, not data, and researchers report that developers' access programmes often leave rejections unnotified. EU law gives vetted researchers a route to platform data, but by May 2026 it had produced applications, not reported access; whether it reaches AI chat services depends on their designation, a SocioLens reading.
• AI developers now fund research on AI's economic effects at scale; Anthropic's largest fund is partly tied to its own policy proposals. Anthropic's pages ask for findings to be shared publicly but state no terms on publication rights or review.
• No study found tests whether this dependence changes what research on AI's social effects concludes.
The central question is therefore not:
"Is AI research captured by industry?"
It is:
WHICH KINDS OF KNOWLEDGE ABOUT AI CAN BE PRODUCED WITHOUT A FIRM'S PERMISSION, AND WHO DECIDES WHAT IS ASKED OF THE REST?
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 an argument.
UNKNOWN — unresolved empirical question.
FOUR CHANNELS OF DEPENDENCE
The evidence concerns different channels, and the debate is easier to follow when they are kept apart:
PEOPLE AND COMPUTE
Who employs the researchers and who can afford to build and test large models.
MONEY
Who funds the research and on what terms.
DATA AND ACCESS
Who holds the records of how AI is used, and who may test the systems.
AGENDA
Who decides which questions are asked, and which disciplines are asked to answer them.
1. INDUSTRY BUILDS THE LARGEST MODELS AND EMPLOYS MORE OF THE EXPERTS
ESTIMATE
Among notable machine-learning models in the top quarter by training compute, the share built by academic laboratories fell from about 65% in the early 2010s to about 10% in the early 2020s (LD08-SL-C01). After deep learning took off in 2012, large technology firms and elite universities published more at AI conferences than a counterfactual predicts, and mid- and lower-ranked universities less; the authors attribute the gap in part to compute (LD08-SL-C02). AI researchers moved from academia to industry in growing numbers to 2020, and a quarter of the transitions from the five institutions ranked highest in the Nature Index went to industry (LD08-SL-C03).
These studies concern building and publishing AI, not research on its social effects. SocioLens synthesis: researchers who cannot build or run frontier models study them on terms set by those who can.
2. TIES TO INDUSTRY IN TECHNICAL AI RESEARCH ARE COMMON AND RARELY DECLARED AS CONFLICTS
ESTIMATE
In highly cited NeurIPS and ICML papers, ties to large technology firms rose from 21% in 2008/09 to 66% in 2018/19 (LD08-SL-C04). Two-thirds of the faculty authors at ACL 2022, a natural-language-processing conference, whose funding could be traced had received industry funding, 72% in the US and 69% in China but 38% in Europe, where students were closer to the others (LD08-SL-C05). Citations to industry-made datasets and models in one 2022 conference were at least three times what industry's share of papers would predict, though the authors say this "is not proof of industry capture" (LD08-SL-C06). From 2018 to 2022, about one in ten papers at ten top AI conferences acknowledged funding from large technology firms, and these papers were cited more (LD08-SL-C07).
Conflict-of-interest statements lag behind. In 10,772 papers at three leading machine-learning conferences (2015–2019), three carried an explicit conflict-of-interest statement, although 15–20% of purely academic papers named a company in their acknowledgements (LD08-SL-C09).
3. RESEARCH ON AI'S SOCIAL EFFECTS IS STILL MOSTLY ACADEMIC
ESTIMATE
The picture changes for research on the ethics and social effects of AI. Roughly 15% or less of peer-reviewed responsible-AI publications had an industry co-author each year, and industry's participation at responsible-AI conferences stayed flat while it grew at leading AI conferences (LD08-SL-C11). In arXiv data to 2020, private-sector AI research tended to specialise in data- and compute-intensive methods at the expense of research on the societal and ethical implications of AI (LD08-SL-C10). On a keyword measure, industry and academic machine-learning papers mentioned social impact about equally often, academic papers slightly more (LD08-SL-C09).
Funding ties also reach the academics who do this work, at least at elite universities: at four North American universities, 19 of 33 computer-science faculty who had published on AI ethics or fairness in 2015–2020, with known funding, had at some point in their careers been directly funded by large technology firms, and 10 of 17 position papers on AI's societal impact in Nature and Science since 2015 had an author with such ties (LD08-SL-C08). The sample is small and the paper argues by analogy with the tobacco industry.
SocioLens synthesis: dependence in research on AI's social effects runs less through authorship than through the other channels: funding, data, access and agenda.
4. WHO FRAMES THE SOCIAL QUESTIONS
ESTIMATE + INTERPRETATION
Inside technical AI research on arXiv, computer-science-only teams wrote 71.2% of the societally oriented sentences in 2024, up from 49.0% in 2014, while teams including social scientists or humanists fell from 25.7% to 3.8%, although their papers were the most societally oriented (LD08-SL-C12). The study excludes social-science journals by design, and its text and a figure caption disagree on how team shares changed. At FAccT, the leading venue for research on algorithmic fairness and accountability, 61.3% of papers from 2018 to 2021 were labelled STEM and 29.7% humanities, social sciences and arts (LD08-SL-C13).
Firms were also early producers of normative statements about AI: of 84 AI ethics guidelines collected by 2019, private companies issued the largest share (LD08-SL-C14). Green argues that tech ethics is contested over "what 'ethics' entails and who gets to define it" (LD08-SL-C15).
No study was found that measures the disciplinary make-up of research on AI's social effects across social-science journals, books and policy reports. The claim that computer science is absorbing the field is supported, on this evidence, for technical AI research only.
5. USAGE DATA: AN EVIDENCE GENRE ONLY THE FIRMS CAN PRODUCE AT SCALE
FACT + ESTIMATE
Anthropic, OpenAI and Microsoft now publish studies of how people use their systems. Anthropic writes that providers face "competitive pressures not to release usage data even if it would be in the public interest" (LD08-SL-C16). All three publish aggregates; Microsoft states that it "cannot release conversational data due to legal and privacy obligations" (LD08-SL-C20), and the OpenAI-co-authored study offers its code on request (LD08-SL-C21).
The instruments are mostly the firms' own: Microsoft classified its conversations with OpenAI's GPT-4o (LD08-SL-C20), and Anthropic classifies with Claude. In Anthropic's first Economic Index report, Claude classified conversations as 57% augmentation and 43% automation, checked by the authors on 150 conversations from users who had submitted feedback, with 90.7% agreement on that label (LD08-SL-C17). A later report found the automation share depended in part on which Claude model classified the same data, 49% or 45% (LD08-SL-C18). Another added "task success," defined as "Claude's assessment of whether Claude completes tasks successfully," checked against a human researcher on a small set of transcripts and described as "directionally accurate" (LD08-SL-C19). Each firm's data describe its own users: the ChatGPT study finds non-work messages above 70% of use (LD08-SL-C21).
Independent work has begun to use other routes and still leans on the firms. University researchers using WildChat, a public dataset of ChatGPT logs collected outside OpenAI, found a far higher work share in Australia than OpenAI's published global figure, with populations that are not comparable, and they borrowed the classifier prompts of OpenAI-co-authored work rather than validating their own (LD08-SL-C22). Researchers who paid users in four countries to donate their full ChatGPT histories studied health advice with no involvement from the developer (LD08-SL-C23). Both routes are small and self-selected.
Dependence on private data is not limited to AI developers. A US study of young workers in AI-exposed jobs uses payroll records under a data partnership with ADP and an Anthropic usage index for its automation analyses (LD02-HQ-C04, a claim of the LD-02 debate). Danish researchers who linked surveys to public administrative records could estimate effects on earnings that usage logs cannot (LD02-HQ-C02, a claim of the LD-02 debate).
6. ACCESS BY PERMISSION
FACT + INTERPRETATION
Developers decide who may test their systems and on what terms. Anthropic's external researcher programme gives API credits, typically $1,000, for AI safety and alignment topics the company considers high priority, no access to non-public models, and may reject applications for any reason without explanation (LD08-SL-C29). Twenty-three researchers argued that prominent developers' terms of service and enforcement discourage good-faith safety evaluation and that access programmes "often do not notify researchers of rejections" (LD08-SL-C28).
Social-media research shows what this can mean. In the Meta 2020 US election studies "all experimental designs had to be approved by Meta," and while Meta lists more than a thousand people it has funded through open awards or fellowships, OpenAlex shows 472 publications by independent academics acknowledging it as a funder, a comparison of people with publications that the authors say should not be read as deliberate concealment (LD08-SL-C31); this is a perspective essay's account. When Twitter closed free academic access in 2023, studies using its data, which had grown by a median 25% a year, rose 0.6% in 2023 and fell 13% in 2024 (LD08-SL-C30); these are raw counts, not a causal estimate.
7. ACCESS BY LAW
FACT
The EU has made researcher access a legal duty for the largest platforms. A delegated act adopted on 2 July 2025 lets vetted researchers, who must show "independence from commercial interests," obtain platforms' internal data to study systemic risks (LD08-SL-C24). By 19 May 2026, national coordinators had received 49 applications and were assessing them, some on "AI features"; their update reported no grant of access (LD08-SL-C25). The Commission fined X €120 million in December 2025, in part for barring researchers from its public data (LD08-SL-C26), and in July 2026 accepted X's plan to open access, still to be implemented (LD08-SL-C27).
The regime covers very large platforms and search engines (LD08-SL-C24). SocioLens reading, not checked against designation lists: conversation logs of AI chat services would be covered only if such a service is designated or is part of a designated one. This version found no comparable right of access to AI usage data under the EU AI Act; that question is open, not settled.
8. FUNDING AND AGENDA
FACT
AI developers now fund research on AI's economic effects. In July 2026 Anthropic committed $200 million to a fund for external research on interventions, mainly in projects of $5–30 million, supporting research "on interventions proposed in the EPF," its own Economic Policy Framework, and other open questions; the page says it wants partners willing to share what they learn publicly at key milestones, and states no terms on publication rights or review (LD08-SL-C32). Its earlier programme paired small grants with "consultation on analyzing Anthropic Economic Index data," sought economists and asked recipients to share findings within six months (LD08-SL-C33), and its advisory council of economists informs the research agenda of its usage index (LD08-SL-C35). Microsoft's AI Economy Institute required grantees to join biweekly convenings that may include interim findings, and scored proposals partly on alignment with the institute's mission (LD08-SL-C34).
These are the firms' own accounts. They show agendas linked to the funders' data and, for Anthropic, to its policy proposals and a focus on economists. They do not show influence on any study's findings.
Public funding is the other side of the question. The Field Map flagged the premise that US public funding for social science is shrinking as US-specific and as a matter of proposals rather than enacted law. This version could not read the US budget requests, appropriations or court records, so it makes no claim about what was proposed, enacted or reversed.
9. THE FIVE POSITIONS
P1 — DEPENDENCE IS STRUCTURAL AND CORROSIVE
Claim:
Firms control the people, compute, data, access and money the field needs, so they shape what can be studied and what is found.
Current status:
Well supported on structure: compute, model building and ties in technical AI research, usage data that only firms hold at scale, discretionary access, agenda-linked funding. Unsupported on corrosion: no study found shows that dependence changes what research on AI's social effects concludes.
P2 — A NEW EVIDENCE GENRE IF FLAGGED
Claim:
Firms' usage data are the only window on real use at scale; with clear labelling and careful reading they are valuable evidence.
Current status:
Supported in that the data exist, are published as aggregates and are reused by independent researchers. Weakened by instruments that only the firm can run or check, figures that move with the measuring model, and samples that describe one firm's users.
P3 — ACCESS LAW CAN RESTORE INDEPENDENCE
Claim:
Legal rights of access for vetted researchers can replace access by permission.
Current status:
Enacted for EU platforms; one enforcement was found (X). Untested in practice: by May 2026 it had produced applications, not reported access. Whether it reaches AI usage data depends on designation. Voluntary safe harbours remain proposals.
P4 — A CONTEST OVER JURISDICTION
Claim:
Computer science, the social sciences and the policy world compete to define AI's social effects, and firms enrol some disciplines rather than others.
Current status:
Supported within technical venues, where computer-science-only teams write most societal framing, and by Anthropic's enrolment of economists. Research on social effects outside technical venues is still mostly academic, and no study measures its disciplinary make-up. "Jurisdiction" is a SocioLens framing: no study applying the sociology of professions to this field was found.
P5 — PRESENCE, NOT CAPTURE
Claim:
Industry ties are common but have not been shown to change findings; industry research is not less attentive to social impact.
Current status:
Supported by the keyword comparison of industry and academic papers and by the low industry share of responsible-AI publications; one industry-dependence study says its own data are "not proof of industry capture." Weakened by the near-absence of conflict-of-interest disclosure, which would make capture hard to see.
10. WHY THE CAMPS TALK PAST EACH OTHER
INTERPRETATION
• "The field" means technical AI research to some and research on AI's social effects to others; the dependence evidence is much stronger for the first.
• "Dependence" covers employment, compute, money, data, access and agenda, which differ in how visible and how reversible they are.
• Capture is inferred from ties, while presence is defended by measures of authorship and keywords; neither measures findings.
• Several sources are firms describing themselves, including the developer whose model drafted this debate.
11. WHAT WE STILL DO NOT KNOW
UNKNOWN
1. Does industry funding, data or access change what research on AI's social effects concludes?
2. How large are industry ties among researchers who publish on AI's social effects in social-science journals, books and policy reports?
3. Does the EU researcher-access regime deliver data, and could it or the AI Act reach the usage data of AI chat services?
4. Do firms' usage figures hold up when outsiders re-classify comparable data with their own validated instruments?
5. What terms govern AI developers' research grants: publication rights, review, data restrictions?
6. What happened to US public funding for social science in 2025–2026, as enacted rather than proposed?
7. Can data donation and public records scale into independent alternatives to firms' usage data?
8. Which disciplines do firms enrol, and does that shape which effects of AI are studied?
12. WHAT EVIDENCE WOULD CHANGE THE PICTURE?
• A comparison of findings by funding source, as meta-research has done in medicine;
• a census of industry ties among authors of research on AI's social effects, beyond computer-science venues;
• implementation data from the EU vetted-researcher regime;
• an independent re-classification of comparable usage data;
• publication of developers' grant terms;
• the US budget and appropriations documents for 2025–2027;
• large, validated data-donation panels of AI use.
13. CURRENT SOCIOLENS POSITION
INTERPRETATION
The best-supported description of the evidence to October 2026 is:
INDUSTRY HOLDS MOST OF THE INPUTS FOR KNOWING ABOUT AI: COMPUTE, MODELS, USAGE DATA AT SCALE, ACCESS AND LARGE NEW FUNDING PROGRAMMES. RESEARCH ON AI'S SOCIAL EFFECTS IS STILL WRITTEN MOSTLY BY ACADEMICS, BUT MUCH OF THE DATA IT RELIES ON IS HELD BY FIRMS. WHETHER THIS CHANGES WHAT IT FINDS HAS NOT BEEN TESTED.
The structural dependence (P1) is well documented, and the usage-data genre (P2) is useful only with its instruments labelled. Law (P3) is the only route that can compel access to firms' own data, and it is not yet working for AI usage data; donation, public logs and public records avoid firms' consent but are partial. The jurisdictional contest (P4) is clearest inside technical research. The presence-not-capture defence (P5) survives only because the decisive test, of findings by funding and access, has not been run.
GEOGRAPHIC AND EVIDENCE SCOPE
The evidence is mostly from the United States and the EU, with comparisons including China and Europe for natural-language-processing researchers, and studies of users in Australia, India, Nigeria, Brazil and Pakistan. It is mostly 2019–2026, with bibliometric data that often end before 2023, and mostly from computer-science venues. Eleven claims rest on firms' own accounts: eight on Anthropic's (C16 to C19, C29, C32, C33 and C35), two on Microsoft's (C20 and C34) and one on a study co-written by OpenAI staff (C21). This debate was drafted by an Anthropic model, and Anthropic's data, access programme and funding are among its objects.
PROVENANCE
Research build:
Source notes prepared 16 October 2026 by Claude with AI research agents, then checked against full texts, official abstracts and official web pages.
Claims:
Each claim rests on a source checked in full text, on an official page or against its official abstract; the claim's verification state says which. One claim rests on the official abstract only (C21). Claims on conference papers read in an arXiv version name that version in their limitations; preprints are marked as such. Web pages of firms and of the European Commission were read as served on 16 October 2026.
Pinned from another debate:
LD02-HQ-C04 and LD02-HQ-C02, claims of the LD-02 debate, are cited in section 5 and pinned at their current versions.
Not claims:
Figures read only in search snippets or secondary reports carry no claim, including the industry shares of models, benchmarks and AI PhDs attributed to a 2023 Science policy article, US federal budget and appropriations figures for social science, court rulings on grant terminations, philanthropic AI-safety funding totals, and the terms of OpenAI's and Google's research funds. The studies of developer transparency and evaluation are in the LD-06 debate, and of AI as a research instrument in the LD-07 debate.
Review:
None yet.
Version:
Living Debate v0.1 — PUBLISHED
16 October 2026