More than half of the most highly valued AI startups have never published a single research paper, according to a report from Science published July 29, 2026. The finding appears in an article titled "AI's top startups are barely publishing their research" and confirms what many people working in AI have noticed informally: the companies at the center of the AI boom are producing very little public science (Science.org).

The Science article, dated July 27 in its byline and posted on the publication's website July 29, did not name specific companies. It described a broad pattern across what venture capitalists call unicorns — privately held startups valued at one billion dollars or more. The fact that most AI companies in this category have produced no public research at all is surprising, because AI has traditionally been a field where scientists share their work openly.

For comparison, the most prominent AI labs do publish, though what they publish has changed. OpenAI's research page lists several July 2026 entries, including "Scientific computing in the age of agentic AI" dated July 28, "GPT-Red: Unlocking Self-Improvement for Robustness" dated July 15, and "GPT-15.6 System Card" dated July 9 (OpenAI Research). Anthropic's research page lists a July 28 publication titled "Frontier Red Team: Discovering cryptographic weaknesses with Claude," a July 24 entry titled "Frontier Red Team," and a July 14 publication titled "Economic Research" (Anthropic Research).

The distinction matters. These are the two most closely watched AI labs, and both still publish regularly. But system cards, which describe what a model can and cannot do safely, and red-team reports, which summarize attempts to break or misuse a model, are not the same as the detailed research papers that AI labs used to release. In earlier years, even corporate labs routinely shared how their systems were built, how they were trained, and how different design choices affected performance. The Science findings suggest the gap between the top labs and everyone else is widening.

The broader context is that AI research now faces pressures that earlier technology waves did not. An AI model's capabilities are directly tied to what a company can sell. A startup that reveals how it trains its models or speeds up how quickly those models respond is effectively giving competitors a blueprint. In the early days of the web, companies like Google and Yahoo published research while competing because the research and the product were separate things. With AI, the research often is the product.

There are reasonable counterarguments. Companies may say that full disclosure creates safety risks, that keeping methods secret protects users, or that the enormous cost of building advanced models justifies keeping results private. None of these arguments are indefensible. But the Science finding is not about companies holding back some sensitive details. It is about companies publishing nothing at all. More than half of AI unicorns fall into that category.

One concern worth raising is the effect on how the field catches its own mistakes. Independent review, repetition of results, and outside scrutiny all depend on having public research to examine. When the best-funded companies produce none, the job of checking whether AI systems actually work as claimed falls to a small number of prominent labs and to university researchers who have far less computing power and data. That means the organizations most able to discover problems are the least likely to describe those problems in a way others can study.

The July 2026 publication titles from OpenAI and Anthropic show what does get released. "GPT-15.6 System Card" is a product document. "Frontier Red Team: Discovering cryptographic weaknesses with Claude" describes testing for vulnerabilities, not how the model was built. "GPT-Red: Unlocking Self-Improvement for Robustness" hints at methods but focuses on a specific use case rather than full disclosure. "Economic Research" from Anthropic is broader still. These publications serve transparency and safety purposes. They are not substitutes for the kind of open research that lets the wider community build on, test, or challenge the work.

In this author's view, the Science finding is less a criticism of individual companies than a sign of the incentives the field has built. The AI research community spent years establishing habits of openness that helped the whole field move faster. Those habits now compete with market forces that reward keeping things secret. Whether the top labs' publication records are enough to keep the field scientifically honest, or whether they become a thin cover over a mostly closed commercial landscape, depends on choices that companies have shown little willingness to make against their own financial interests.

The research published in July 2026 comes from a handful of names. The majority of the most highly valued AI startups have nothing to show. The field will need to decide whether that is acceptable.