A bibliometric count across 317 AI unicorns
Bibliometric analysis measures research activity by counting publications and how they get cited. This study pointed that method at the private firms driving AI development. AI development is concentrated within private firms, yet how much those firms participate in scientific publishing is poorly understood — that is the question it starts from.
How the study was framed
Every AI unicorn since 1998
A unicorn is a private startup valued above one billion dollars. The study took the 317 in the AI field as a whole rather than sampling. Only publications with leading contributions from the firm were counted: 1,389 peer-reviewed and 688 preprints.
A preprint is a manuscript posted without going through peer review. This study is itself a preprint on bioRxiv and has not been certified by peer review. That is the frame for reading its numbers.
Artificial intelligence (AI) development is concentrated within private firms, yet their participation in scientific publishing remains poorly understood. We conducted a bibliometric analysis of all 317 AI unicorn startups (1998–2025). Only 1,389 eligible peer-reviewed publications and 688 preprints involving some leading startup contributions were identified. / This article is a preprint and has not been certified by peer review / Posted July 16, 2026. — From the research question, the scope and counts, the preprint notice, and the posting date
Most publish nothing, and influence sits with a handful
More than half of the 317 never appear in the academic record at all, and among those that do, the distribution is extreme.
Academic contribution across 317 AI unicorns
| Measure | Value |
|---|---|
| Firms with no qualifying scientific output | 52.4% |
| Firms with any highly cited paper (≥200 citations) | 24 firms (7.6%) |
| Share of citations held by the top 10% of firms | 96.8% |
| Of 134 firm-attributed highly cited papers, the share from three firms | 92 papers |
| Share of the overall AI literature in 2025 | 0.1% |
52.4% produced nothing that qualified
Over half the column is blank. Of the 317 firms, 52.4% produced no qualifying scientific output. Narrow it to firms with at least one highly cited paper — 200 citations or more — and you are down to 24 firms, or 7.6%.
More than nine in ten crossed a billion-dollar valuation without producing widely referenced research. For most of these companies, there is simply no way to check the technology from the outside through published work.
The top 10% hold 96.8% of citations
The skew goes further. The top 10% of firms account for 96.8% of citations. Of 134 firm-attributed highly cited papers, 92 came from just three startups.
Not only do few firms do research; within that few, influence concentrates further still. "Research from AI companies" is not a category that holds together.
Valuation and publication output are unrelated
The study also looked at how investor valuation tracks academic productivity. Valuation was not associated with publication productivity or with highly cited output. Funding raised showed weak associations.
The intuition that the most highly valued firms publish the most does not hold. Valuation is being set on a different axis from academic output.
More than half of startups (52.4%) produced no qualifying scientific output, and only 24 firms (7.6%) produced any highly cited papers (≥200 citations). Scientific influence was highly concentrated: the top 10% of firms accounted for 96.8% of citations, while three startups accounted for 92 of 134 firm-attributed highly cited papers. Firm valuation was not associated with publication productivity or highly cited output, whereas funding raised showed weak associations. — All three points in this section (the share producing nothing, the citation concentration, and the link to valuation) come from one continuous passage of the abstract, so the quote is consolidated into a single block at the end of the section
What the absence of papers actually means
The concerns the authors raise matter more in practice than the size of the numbers.
The concerns the authors raise
0.1% of the AI literature puts it in scale
The share is worth holding onto. Participation in formal scientific communication among AI unicorn startups is negligible, comprising only 0.1% of the overall AI literature in 2025. AI research papers are being published in volume. The unicorns' contribution to that volume is one part in a thousand.
The authors call it negligible and conclude that most leading developers of AI technology do not engage with the scientific literature — which is where the concern for transparency, reproducibility, and accountability at this frontier comes from.
How to verify what the technology does
What do you do with this as a reader? When you go looking for a company's papers and find none, that is not a search problem. The implication of this study is that there was probably nothing there to begin with.
What you can check is limited to the technical documentation and specifications the company publishes itself, plus benchmarks measured by third parties. Read them knowing peer-reviewed verification does not exist. This paper is a preprint too, so the same stance applies to it.
Overall, participation in formal scientific communication among AI unicorn startups is negligible, comprising only 0.1% of the overall AI literature in 2025. Most leading developers of AI technology do not engage with the scientific literature, raising concerns for the transparency, reproducibility, and accountability of this rapidly moving innovation frontier. — Both points in this section (the share of the AI literature and the concerns for transparency, reproducibility, and accountability) come from the same closing passage of the abstract, so the quote is consolidated into a single block at the end of the section
The preprint itself is published as a PDF. If you want the methods and exclusion criteria rather than just the abstract, converting the PDF to Markdown first keeps the tables and headings intact.
Conclusion: missing papers are not a search problem
Of 317 AI unicorns, 52.4% produced no qualifying academic output. The top 10% of firms hold 96.8% of citations, and 92 of 134 highly cited papers come from three companies. Valuation shows no relationship to publication count, and the group as a whole makes up 0.1% of the AI literature in 2025. If you try to verify an AI company's technology through published research, most of the time the material was never there. What you have to work with is the company's own documentation and third-party measurements, both of which have to be read as unreviewed. Applying the same yardstick to this study — itself a preprint — is enough.
