Does industry funding or data access change what research on AI's social effects finds?

Why it matters

The debate documents structural dependence through people, compute, money, data and access, but no study tests whether it changes findings. The presence-not-capture position (P5) survives only because this test has not been run. In medicine, meta-research has compared trial conclusions by sponsor (background knowledge, not a source checked for this debate); nothing comparable exists for research on AI's effects, and conflict-of-interest statements in machine-learning papers are almost absent, so funding must be traced from other records.

The debate lists this among what we still do not know: Does industry funding, data or access change what research on AI's social effects concludes?

Evidence that motivates it

• In 10,772 NeurIPS, ICML and CVPR papers (2015–2019), only three contained an explicit conflict-of-interest statement, although 15–20% of purely academic papers named a company in their acknowledgements; industry and academic papers differed by only 3.1% in their use of social-impact terms, academic papers using them slightly more.
  Three explicit conflict-of-interest statements in 10,807 machine-learning papers.
• In July 2026 Anthropic committed $200 million to an Economic Futures Research Fund for external research on interventions to prepare society for AI’s economic impacts, mainly in projects of $5–30 million, supporting research “on interventions proposed in the EPF,” its own Economic Policy Framework, and on other open questions; its public page says it wants partners “willing to share what they’re learning publicly at key milestones” and states no terms on publication rights or review.
  A developer's $200 million fund supports research partly on interventions in its own policy framework; its page wants milestone sharing but states no publication terms.
• In a sample of 100 EMNLP 2022 papers, 44% had at least one industry-affiliated author, against 24% reported for 2019–2021 by earlier work, and citations of industry-made datasets, pre-trained models and previous best results were at least three times what the industry share of the field would predict; the authors state that this “is not proof of industry capture.”
  Heavy reliance on industry artifacts, which the authors say is not proof of capture.
• Bak-Coleman, O’Connor, Bergstrom and West argue that technology firms shape research through control of data and funding: 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 funded through open awards or fellowships, OpenAlex shows only 472 publications by independent academics acknowledging Facebook or Meta as a funder, a gap the authors say should not be read as a deliberate attempt to obscure funding.
  A perspective essay reports a gap between funded individuals and acknowledged funding.

A possible design

Build a corpus of empirical studies of AI's social effects (2020–2026) with pre-registered coding of conclusions (direction and strength of harms or benefits found) by coders blind to funding. Code funding, data source and access route from acknowledgements, grant databases, CVs and developer funding announcements. Compare conclusions by funding and data source within topic and design, and test sensitivity to undisclosed funding using the gap between announced grantees and acknowledgements.

Data needed

Study corpus with full texts; grant and fellowship lists published by developers and foundations; author CVs; a blinded coding protocol.

Feasibility

A PhD project. Medium; mostly coder time for blinded coding and funding tracing. Undisclosed funding can bias the comparison in either direction, because disclosure may depend on what a study finds; the design gives no basis for a bound, so report sensitivity analyses over plausible disclosure patterns. Topic selection by funders is itself a finding and should be measured separately from conclusions.

Provenance

RO-LD08-01 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-08 debate on 2026-10-16. Not yet reviewed. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.

All research opportunities