Over half of AI unicorns have published zero research papers, Stanford professor finds

More than half — 52.4 percent, or 166 — of the 317 AI companies valued at over $1 billion ($1.4 trillion won) have never published a single research paper, according to a new study.

Among those that have published, only 24 companies, or 7.6 percent of the total, had at least one paper cited more than 200 times. Of the 919,486 AI papers published worldwide last year, just 950 — roughly 0.1 percent — involved these companies in a leading role.

A research team led by John Ioannidis, a professor at Stanford University's Meta-Research Innovation Center (METRICS), published the findings on the preprint server bioRxiv on July 15. The paper has not yet undergone peer review.

Ioannidis has long challenged the reproducibility of published medical research. He gained prominence in 2016 when he criticized Theranos, the blood-testing startup, for making technology claims unsupported by any published research — coining the term "stealth research" to describe the practice.

Theranos had been valued at $9 billion after claiming it could diagnose hundreds of diseases from a single drop of blood. No peer-reviewed papers ever validated its technology, and the company was ultimately exposed as a fraud. Its founder was sentenced to prison.

The research team examined 317 AI companies valued at $1 billion or more from 1998 through last year, focusing on whether each company had led research rather than merely participated in it.

The results showed that the top 5 percent — 16 companies — accounted for 59.9 percent of all papers published by the group. The remaining 166 companies had no record of leading any published research.

By citation count, a measure of how widely other researchers reference a paper, those same 16 companies accounted for 92 percent of all citations. The research team said the concentration exceeded the skewed distributions typically seen across academia.

OpenAI was the most-cited company, accounting for 39.4 percent of all citations. Chinese facial-recognition firm Megvii followed at 26.6 percent, with Hugging Face at 6.9 percent.

Megvii led with 53 papers cited more than 200 times, followed by OpenAI with 30. The remaining 293 companies had none.

The research team also examined who was actually writing the papers.

Of the 1,988 company-affiliated authors identified, 38 percent held simultaneous appointments at universities or research institutions. About half were researchers employed exclusively by their companies.

Only eight researchers at OpenAI and eight at Megvii had authored five or more papers. OpenAI employs about 4,500 people as of 2026, while Megvii's workforce is estimated at between 2,300 and more than 3,000.

The research team noted that none of the companies qualified for inclusion in global research institution citation ranking databases, which require a minimum number of researchers with at least five published papers.

Perhaps the most striking finding was the absence of any relationship between a company's valuation and its research output.

The correlation coefficient between valuation and number of papers was 0.017 — statistically indistinguishable from zero.

The team said investors appear to pay virtually no attention to a company's academic track record, relying instead on non-public information or other criteria.

The research team pointed to the medical field as an example of the risks posed by insufficient scrutiny of AI technology, warning that unverified tools could spread widely even when they are ineffective or harmful.

The team said published papers serve a purpose beyond simply announcing results — they allow other researchers to replicate methods, test the validity of claims and expose limitations.

However, the team acknowledged that whether companies are obligated to publish research remains debatable, and noted that protecting proprietary technology is a legitimate business choice.

The team also flagged a limitation: research conducted under undisclosed affiliations or circulated only internally would not have been captured in their analysis.

DOI: 10.64898/2026.07.15.738744

Negligible participation in the scientific literature of AI unicorn startups, Quentin E.A. Loisel, Alejandro Sandoval-Lentisco, John P.A. Ioannidis, bioRxiv 2026.07.15.738744.