Source
Bak-Coleman, West, O’Connor and Bergstrom (2026), LIT-481 — read as NOTE-231. 295 papers in Science, Nature, PNAS and their transfer journals, 1210 authors.
The claim
49% of high-profile social media papers have a disclosable industry tie and most go undisclosed. The published competing-interest statement is therefore not a reliable signal, and a reader who treats its absence as the absence of a tie will be wrong about half the time.
The reason this is a practice rather than a complaint is that the ties are detectable: funding, collaboration and employment records in OpenAlex, plus publicly announced RFPs and fellowships, plus CVs for editors. The source validated every funding and employment tie against sources independent of OpenAlex. Anyone can do this, and it takes a query rather than an investigation.
What to actually look at
The topical distribution, more than any individual paper. The skew is the finding with the largest implications: ties are over-abundant in the cluster about what users share and sparse in the cluster about platform dynamics. If you are reading to find out whether a platform-scale design choice has an effect, that is the cluster with the least industry money in it and the fewest papers.
Authors, not just papers. 49% of papers carry a tie and only 21% of authors do. That gap says a small group with durable relationships appears on half the output — a different situation from a field broadly engaged with industry, and one disclosure norms alone will not fix.
Editors and reviewers. The same detection finds undisclosed ties there too, where disclosure norms are weakest and where the effect on what gets published is most direct.
Attention, as a reason to look rather than a reason to trust. Industry-tied work draws more citations, more news coverage and more policy references. The corpus as a whole reaches 745 policy documents. DP-005 in the direction it is least often read: being well-adopted is what makes a body of work worth auditing.
Conditions
The 49% is a floor. Detection requires a public trace, so undetected ties push the figure down rather than up. That is the conservative direction and it means the number should not be quoted as an estimate.
“Disclosable tie” is narrower than “conflict of interest”, deliberately — it is defined to be detectable and to match each journal’s own policy window (three years for Nature, four for PNAS, five for Science). Do not read it as capturing everything a reader might reasonably want disclosed.
The topical skew is correlational and the source says so. Researchers already interested in user-level questions may attract industry funding rather than being directed toward it. The design cannot separate those, and the practice’s instruction — look at the distribution — survives either way.
Four firms, seven journals, one field. Meta, X, Google, Microsoft; Science, Nature, PNAS and transfer venues. Ties outside that set are invisible, and nothing establishes the same rates elsewhere.
And this record has not run the check on itself. The method is mechanical and its corpus construction is reusable; whether the anthology’s own sources carry comparable rates is unknown, and the promotion condition asks for exactly that measurement somewhere other than social media.