Contribution
A measurement of a literature rather than of a phenomenon, done with the literature’s own disclosure rules as the yardstick. Before it, industry influence in this field was a widely-voiced worry; after it, it is 49% of papers, 21% of authors, and a named topical skew, against each journal’s own competing-interest window.
Key insight
The gap between the two percentages is the finding. 49% of papers have a tie; only 21% of authors do. If industry engagement were broad, those numbers would be closer. They are not, so a small group with long-lasting relationships is appearing on half the high-profile output.
That reframes the problem. Broad engagement would be a question about norms. Concentration is a question about a few people’s centrality, which is both a smaller problem and a harder one to fix by disclosure alone — and the paper finds the same concentration among editors and reviewers, where disclosure norms barely exist.
Assumptions
- “Disclosable tie” is a constructed, deliberately narrow category: funding from, collaboration with, or employment by Meta, X, Google or Microsoft, within the journal’s own competing-interest window. Narrower than “conflict of interest” and the paper says so.
- Detection needs a public trace — OpenAlex, announced RFPs, fellowships, CVs. Every funding and employment tie manually validated against independent sources.
- Seven journals (Science, Nature, PNAS and transfer venues) and four firms. Outside that, invisible.
- Topical communities from bibliographic coupling, five of them, labelled post hoc.
Key results
- 49% of papers have a disclosable tie; the majority undisclosed.
- 21% of authors. Meta 14%, Google 8%, Microsoft 6%, X 1%.
- Editors and reviewers carry undisclosed ties too.
- Industry-tied work draws more attention across citations, policy documents, news and social media.
- Topical skew: ties over-abundant in the misinformation sharing cluster, sparse in platform dynamics. The tied cluster’s abstracts lean toward experiments on people sharing false headlines; the independent one toward network structure and dynamics.
- Corpus reach: 295 papers, 53,120 citations (180 per paper), 15,708 news articles, 100,863 social media mentions, 418 Wikipedia mentions, 745 policy documents.
Claims
Measured, and conservatively: the 49%. Undetected ties push it down, so it is a floor rather than an estimate — the rare case where the headline number is the cautious one.
Measured: the concentration among authors, the presence among editors and reviewers, and the attention differential.
Correlational, and stated as such: the topical skew. The paper says it cannot definitively show industry funding causes it. Researchers already interested in user-level questions may attract the funding rather than being directed by it, and this design cannot separate those.
Method
OpenAlex corpus construction via bibliographic coupling; per-author tie detection from OpenAlex plus announced programmes, manually validated against independent sources; Altmetric for reach; community detection for topics; public CVs for editors.
Concepts
Disclosable tie as a detectable proxy for competing interest, defined against each journal’s own policy window; topical redirection as the risk that matters more than any individual paper’s conclusions.
Connections
- LIT-482, filed alongside, argues the effects worth measuring live at platform and collective scale. This finds industry ties sparse in exactly that cluster and over-represented in the cluster about what individual users share. The halves fit — and share a first author, so they are one position argued twice rather than two groups agreeing.
- DP-005 says adoption is not evidence. Here the corpus is extraordinarily well-adopted — 180 citations a paper, 745 policy documents — and that is the reason to examine it, not a reason to trust it. The principle read in the direction it is least often used.
Bearing on the record
It does not source a practice about ML. It sources a practice about reading this literature, and it supplies the conditions that any practice drawn from these papers has to carry.
It also makes a concrete instruction possible: the ties are detectable mechanically, from public sources, by anyone. That is unusual — most warnings about conflicts of interest end at “be aware”.
Limitations
Correlational on the part that matters most. The topical skew is the finding with the largest implications and the weakest identification, and the paper says so.
A floor, not an estimate. Detection requires a public trace. Whatever share of ties leaves none is missing.
Four firms, seven journals. A tie to a firm outside the set, or work published elsewhere, does not appear. The restriction is defended — these journals were chosen partly so the analysis could be tailored to their specific competing-interest policies — and it is still a restriction.
Labels applied after clustering. The five topical communities come from bibliographic coupling and are named afterwards; “platform dynamics” versus “misinformation sharing” is an interpretation of a partition, not a pre-registered category.
Same first author as its companion. Not independent.
Open questions
- Does the same concentration hold in ML research proper? The method is mechanical and the corpus construction is reusable; nothing about it is specific to social media, and this record’s own sources would be the interesting corpus to run it on.
- Does disclosure change behaviour, or only attribution? The paper argues for stronger disclosure norms; whether disclosure alters what gets studied is a separate empirical question it does not address.