A Stanford study of 317 AI unicorns reveals that more than half have never published a single research paper, exposing a quiet contradiction at the center of an industry that presents itself as the vanguard of scientific progress. The concentration of knowledge is severe — a handful of prolific researchers at a handful of dominant firms generate nearly all the citations that exist. This silence is not an oversight but a strategy, one that increasingly sequesters the trajectory of transformative technology from the public discourse that might otherwise shape it.
Stanford study: Over half of AI unicorns publish no research; China leads in openness
A field claiming to reshape science produces almost no scientific literature
Mark
Why does it matter if a company publishes research or not? Aren't they still advancing the field through their products?
Mimi
Products are one thing. But science is about evidence, replicability, and collective understanding. If a company builds something powerful but tells no one how it works, the broader research community can't learn from it, can't verify claims, can't build on it. You get a black box instead of knowledge.
Mark
So this is really about transparency and accountability?
Mimi
Partly. But it's also about the pace of progress. Science moves faster when people can see what others are doing. When half the unicorns publish nothing, we're essentially flying blind about where AI is actually heading.
Mark
The study shows Chinese companies publish more than American ones. Is that surprising?
Mimi
It reflects different business philosophies. U.S. frontier labs have decided that keeping research proprietary is a competitive advantage. Chinese companies seem to believe open-source models serve them better. Whether that's true long-term is still an open question.
Mark
What about the concentration among a few authors? Why does OpenAI's research come from so few people?
Mimi
That's the real puzzle. These are massive companies with thousands of employees. Yet the research output comes from a tiny fraction of them. It suggests either that most employees aren't doing research work, or that research isn't being published even when it happens.
Mark
And that matters because?
Mimi
Because it means the public understanding of AI development is shaped by an incredibly narrow slice of the actual work being done. We're seeing only what a few dozen people choose to show us.
The Pulse
- An industry claiming to reinvent science is producing almost no science the public can read — over half of the world's most valuable AI startups have never published a single paper.
- Citation power is pooled at the very top: five percent of companies command ninety percent of references, with OpenAI alone holding nearly forty percent of that share.
- A geographic fault line has opened — Chinese AI unicorns publish at significantly higher rates than their U.S. counterparts, reflecting a deepening divide between open-source and closed-source philosophies.
- The people doing the publishing are vanishingly few — among nearly two thousand affiliated authors, just twenty-seven prolific researchers account for close to forty percent of all highly cited authorship.
- Without published research, policymakers and the broader scientific community are left navigating the future of AI largely blind, dependent on what companies choose to reveal about themselves.
A Stanford study of 317 AI unicorns reveals that more than half have never published a single research paper, exposing a quiet contradiction at the center of an industry that presents itself as the vanguard of scientific progress. The concentration of knowledge is severe — a handful of prolific researchers at a handful of dominant firms generate nearly all the citations that exist. This silence is not an oversight but a strategy, one that increasingly sequesters the trajectory of transformative technology from the public discourse that might otherwise shape it.
A Stanford research team examined the publication records of 317 AI unicorns — private companies valued above a billion dollars — across nearly three decades, and published their findings in Science magazine. The result was a striking paradox: more than half of these companies, which collectively claim to be reshaping the nature of science, have never produced a single research paper.
The output that does exist is concentrated to an almost absurd degree. The top five percent of companies generate over ninety percent of all citations, with OpenAI responsible for nearly forty percent of that share. In 2025, AI unicorn papers represented just 950 of the more than 900,000 AI papers published globally — a fraction of a percent. Among the companies that do publish, the work flows from a remarkably small circle: just 27 researchers account for nearly forty percent of authorship across the 132 most-cited papers.
Corresponding author John Ioannidis framed the silence as a deliberate contradiction. Leading U.S. frontier labs have increasingly chosen to keep research proprietary, while Chinese AI unicorns — publishing at rates nearly twenty percentage points higher — have moved toward open-source models. The divergence is consequential. Published research enters a shared space where it can be examined, challenged, and built upon. Research that stays inside a company remains visible only to those within its walls, leaving the broader direction of AI development opaque to the researchers, policymakers, and public who will ultimately live with its consequences.
A Stanford University research team set out to answer a straightforward question: how much are the world's most valuable artificial intelligence startups actually publishing? What they found was striking enough to catch the attention of Science magazine. Between 1998 and 2025, researchers analyzed the publication records of 317 AI unicorns—private companies valued at over a billion dollars—and discovered a paradox at the heart of an industry claiming to reshape science itself. More than half of these companies have never published a single research paper.
The concentration of scientific output among AI startups is severe. The top 5 percent of companies account for more than 90 percent of all citations in the field. OpenAI alone is responsible for nearly 40 percent of that citation share, followed by Megvii and Hugging Face. Among the 317 companies studied, only 6.4 percent have produced work considered highly cited. The rest either publish sporadically or not at all. In 2025, the global AI research community published over 900,000 papers. Papers connected to these unicorn startups made up just 950 of them—a mere 0.1 percent of the total.
John Ioannidis, a meta-scientist at Stanford and the corresponding author of the study, framed the finding as a fundamental contradiction. Here is a field that claims to be reshaping the nature of science itself, yet produces almost no scientific literature to show for it. The disconnect is not accidental. It reflects a deliberate choice by many companies to keep their research behind closed doors, a strategy that has become increasingly common among leading U.S. frontier labs. When research stays inside a company, only a handful of people understand where the technology is actually heading.
Within the companies that do publish, the work comes from a remarkably small group of people. Among nearly 2,000 authors affiliated with these startups, more than half work exclusively for their companies. Of the 132 highly cited papers, just 27 prolific authors accounted for nearly 40 percent of all authorship. At OpenAI and Megvii, which each employ thousands of people, only eight researchers from each company have published more than five papers. The disparity between workforce size and research output is stark.
Company valuation, it turns out, has little bearing on how much research gets published. Data shows no meaningful correlation between how much money a startup has raised and how many papers it produces. While well-funded companies are somewhat more likely to publish than their less-capitalized peers, the relationship is weak. High-impact research does not scale proportionally with investment. The growth in startup-led papers has been real—from 18 in 2016 to 534 in 2025—but it remains marginal relative to the broader scientific landscape.
A striking geographic divide has emerged. Among 40 Chinese AI unicorns, nearly two-thirds have published research papers. Among 209 U.S. companies, more than half have published nothing. This reflects a fundamental difference in strategy. Leading U.S. frontier labs have increasingly adopted closed-source models, keeping their work proprietary. Leading Chinese companies, by contrast, are actively embracing open-source approaches. The choice has real consequences. When research is published, it enters the public domain where researchers, policymakers, and industry observers can examine it, debate it, and build on it. When it remains confined to a company, the direction of AI development becomes visible only to those inside the walls.
Notable Quotes
For a field claiming to be reshaping science and regarded as having tremendous scientific potential, there is almost no scientific literature.— John Ioannidis, Stanford University meta-scientist
Want the full story?
Read the original at KuCoin ↗
Technology