Most billion

The biggest names in artificial intelligence promise to transform drug discovery, software development and scientific research. But more than half of AI’s billion-dollar startups have never produced a single qualifying scientific paper or preprint, a new analysis reveals.

The finding, laid out in a preprint from Stanford professor John Ioannidis, is drawing sharp attention online and in the research community. It raises a basic question: if these companies say they are building world-changing technology rooted in science, how can anyone outside the company evaluate whether the work is real or reproducible?

The researcher who flagged Theranos is asking the same question about AI

Ioannidis is not a newcomer to this kind of scrutiny. In 2015, he was among the first to publicly question the lack of peer-reviewed studies behind Theranos, the blood testing startup that turned out to be based on fraudulent data. Now he is turning that same lens on the AI industry.

The analysis looked at AI unicorns, private companies valued at more than $1 billion. It found that scientific influence proved even more challenging for startups than simple paper counts might suggest. Many firms that do publish play only a minor role, contributing to papers rather than leading them.

Celina T. Zhao reported the findings for Science Magazine, noting the gap between the bold claims these companies make and their near-total absence from the scientific literature.

Why AI companies stopped sharing what they know

The retreat from publishing is not random. As one commenter on Hacker News put it, publishing is most valuable to people who have no other way to get the attention of smart strangers. Once a company can hire nearly anyone and everyone already returns its calls, the main remaining effect of publishing is to tell competitors which things worked.

That dynamic has played out before. When dyes became commercially valuable in the 19th century, interesting chemistry moved from open journals into company labs and stopped coming out. AI appears to be following the same path, shifting from an open science into a guarded industry.

There is an added wrinkle. The business model of many AI companies depends heavily on model training based on available published data, academic and otherwise. Critics on Hacker News called this ironic and selfish: companies consume the open research of others while contributing little of their own. Public papers give an advantage to competitors who may give nothing back, and rivals can build under the radar when research is freely available. That fear, commenters noted, is why companies are so cautious about publishing.

Peer review, reproducibility and the line between science and hype

The concern goes deeper than competitive strategy. Science relies on peer review and reproducibility. When AI companies shift from publishing open research to guarding trade secrets in black boxes, some observers argue it marks a transition from genuine scientific discovery to commercial hype and valuation games.

Without published, reviewable work, outsiders have no way to independently verify the claims these startups make to investors, regulators and the public. The Theranos comparison is pointed: that company’s lack of published, peer-reviewed evidence was an early warning sign that something was wrong.

Could licensing rules force openness?

The preprint has also sparked a lively debate about whether legal tools could push companies to share their findings. Some commenters floated the idea of a copyleft requirement for research, a kind of viral knowledge license that would compel anyone who builds on a published paper to publish their own related work.

The concept drew skepticism. A copyleft paper does not compel someone who makes a product based on it to publish more papers. Copyright law does not deal with abstract ideas passing through people’s minds. The GPL, the most famous viral software license, has never had its definition of “derivative work” fully tested in court.

Supporters countered that if a person can agree not to talk about something under a nondisclosure agreement, or not to work in a field under a non-compete clause, then a contract requiring publication of future work in a given area should be possible. Others pointed out that any such agreement would need tightly defined boundaries to survive legal challenge. The reach of “related research” is inherently fuzzy, and overly broad terms would likely be struck down.

An existing license, the Reciprocal Public License, already attempts something similar for software, though it has seen limited adoption.

A physics student’s case for publishing anyway

Not everyone sees the retreat from openness as inevitable. One undergraduate physics student described publishing a paper on recursive self-improvement mapped to Epoch AI data, then reaching out to a UK professor working on similar research. The two spoke for about an hour, comparing overlapping results and different methods. The student’s takeaway: for young people getting into any field, research is a great way to meet new people and build connections that would otherwise be out of reach.

That argument cuts against the grain of the industry’s direction. For established companies, secrecy may be rational. For the broader scientific community, the cost of that secrecy is harder to measure but potentially enormous.

The preprint is still under review. Whether it prompts any concrete change in how AI startups operate remains to be seen, but the conversation it has started is already forcing a question the industry would rather avoid: if the science is real, why not show the work?

I cover what the internet is actually talking about: the new release, the viral moment, the creator beef, the product launch. My job is to explain it fast and clearly, without the breathless hype. I read the timeline…