AI's top startups are barely publishing their research

AI's Top Startups Are Barely Publishing Their Research

Source: Science Magazine | 2026-07-30, Vol 393, Issue 6810

Authors: Science staff coverage; meta-analysis by John Ioannidis et al. (bioRxiv preprint, 16 July 2026)

Category: Industry / Open Science | Quality Score: 4/5

English Summary

A bioRxiv preprint by John Ioannidis and colleagues analyzed all 317 AI unicorn companies (private firms valued over $1 billion) that have existed from 1998 to 2025. They searched for publications—journal articles, conference papers, reviews, and preprints—where a company researcher played a leading role as first or last author.

Key Findings

1. More than half of AI unicorns have never produced a single qualifying paper. The final dataset included only 2,077 publications (1,389 peer-reviewed + 688 preprints) across all 317 companies.

2. Collectively, AI unicorns accounted for just ~1 in every 1,000 AI papers published in 2025.

3. Scientific influence is hyper-concentrated: The top 5% of firms accounted for >90% of all citations. OpenAI alone was responsible for nearly 40% of all citations, followed by Megvii and Hugging Face.

4. Prolific output comes from a tiny group: Despite OpenAI employing ~4,500 people, only eight researchers authored five or more qualifying papers.

5. China vs. U.S. divergence: Chinese AI firms consistently published more papers than U.S. counterparts. Leading U.S. frontier labs have increasingly kept their most capable models closed-source, while leading Chinese companies (e.g., Moonshot AI / Kimi K3) have embraced open-source model releases.

Why So Few Publications?

Commercial incentives: Unlike pharma (where patents protect discoveries), AI companies gain little from public disclosure. Google's Transformer paper (2017) is cited as a cautionary example—despite patenting, competitors freely built on it.
Speed mismatch: AI startups operate on much faster timelines than academic peer review (months to years).
"Blogification" of research: Companies increasingly announce models via blog posts, technical reports, and code/dataset releases rather than peer-reviewed journals (Avijit Ghosh, Hugging Face).

Broader Concerns

Validation paradox: "For a field that is supposedly reshaping science... not having any scientific documentation seems like a very weird paradox," says Ioannidis. "How can you judge that what they say is real, validated, and reproducible?"
Safety assessment gap: The scarcity of publications makes it harder to assess AI's social impacts, including energy use and safety risks.
Acceleration risk: Emma Pierson (UC Berkeley) argues that whether published or secret, rapid progress toward powerful generalist AI risks accelerating models that pose serious societal concerns, including cyberattack amplification.

Stakeholder Perspectives

| Voice | Position | |-------|----------| | John Ioannidis (Stanford) | Publishing scarcity is a paradox for a field claiming to reshape science | | Mohamed Abdalla (U. Alberta) | Reflects commercial incentives—"The company's job is to advance money" | | Nur Ahmed (U. Arkansas) | AI companies have learned they gain little from publishing | | Avijit Ghosh (Hugging Face) | The real question is code/weights release, not journal vs. blog | | Emma Pierson (UC Berkeley) | Whether open or closed, racing toward powerful AI carries safety risks |

\u4e2d\u6587\u6982\u8981

Science \u62a5\u9053\u4e86 bioRxiv \u4e0a\u7684\u4e00\u7bc7\u5143\u7814\u7a76\u9884\u5370\u672c\uff08Ioannidis \u7b49\uff0c2026\u5e747\u670816\u65e5\uff09\uff0c\u5206\u6790\u4e86 1998\u20132025 \u5e74\u95f4\u5b58\u5728\u7684\u5168\u90e8 317 \u5bb6 AI \u72ec\u89d2\u517d\u516c\u53f8\uff08\u4f30\u503c\u8d85 10 \u4ebf\u7f8e\u5143\u7684\u79c1\u8425\u4f01\u4e1a\uff09\u7684\u79d1\u5b66\u53d1\u8868\u60c5\u51b5\u3002

\u6838\u5fc3\u53d1\u73b0

1. \u8d85\u8fc7\u534a\u6570 AI \u72ec\u89d2\u517d\u4ece\u672a\u4ea7\u51fa\u4e00\u7bc7\u8fbe\u6807\u8bba\u6587\u3002 \u5168\u90e8 317 \u5bb6\u516c\u53f8\u4ec5\u6709 2,077 \u7bc7\u51fa\u7248\u7269\uff080,389 \u7bc7\u540c\u884c\u8bc4\u5ba1 + 688 \u7bc7\u9884\u5370\u672c\uff09\u3002 2. AI \u72ec\u89d2\u517d\u5408\u8ba1\u4ec5\u5360 2025 \u5e74 AI \u8bba\u6587\u7684\u7ea6\u5343\u5206\u4e4b\u4e00\u3002 3. \u5b66\u672f\u5f71\u54cd\u9ad8\u5ea6\u96c6\u4e2d\uff1a \u524d 5% \u7684\u516c\u53f8\u5360\u636e 90% \u4ee5\u4e0a\u7684\u5f15\u7528\u91cf\u3002OpenAI \u5355\u72ec\u5360\u8fd1 40%\uff0c\u5176\u6b21\u662f\u6fb3\u9e44\u5185\u578b\u548c Hugging Face\u3002 4. \u4e2d\u56fd vs. \u7f8e\u56fd\u5206\u5316\uff1a \u4e2d\u56fd AI \u4f01\u4e1a\u53d1\u8868\u663e\u8457\u591a\u4e8e\u7f8e\u56fd\u540c\u884c\u3002\u7f8e\u56fd\u524d\u6cbf\u5b9e\u9a8c\u5ba4\u65e5\u8d8b\u5c01\u95ed\uff0c\u4e2d\u56fd\u516c\u53f8\uff08\u5982 Moonshot AI / Kimi K3\uff09\u5219\u62e5\u62b1\u5f00\u6e90\u3002

\u539f\u56e0\u4e0e\u4e89\u8bae

\u5546\u4e1a\u6fc0\u52b1\uff1a \u4e0e\u5236\u836f\u4e1a\uff08\u53d7\u4e13\u5229\u4fdd\u62a4\uff09\u4e0d\u540c\uff0cAI \u516c\u53f8\u516c\u5f00\u62ab\u9732\u6280\u672f\u83b7\u76ca\u751a\u5c11\u3002Google \u7684 Transformer \u8bba\u6587\uff082017\uff09\u88ab\u89c6\u4e3a\u524d\u8f66\u4e4b\u9274\u3002
\u901f\u5ea6\u4e0d\u5339\u914d\uff1a \u521b\u4e1a\u516c\u53f8\u8282\u594f\u8fdc\u5feb\u4e8e\u540c\u884c\u8bc4\u5ba1\uff08\u6570\u6708\u81f3\u6570\u5e74\uff09\u3002
\u201c\u535a\u5ba2\u5316\u201d\u8d8b\u52bf\uff1a \u516c\u53f8\u8f6c\u5411\u901a\u8fc7\u535a\u5ba2\u3001\u6280\u672f\u62a5\u544a\u548c\u4ee3\u7801/\u6570\u636e\u96c6\u53d1\u5e03\u6a21\u578b\uff0c\u800c\u975e\u540c\u884c\u8bc4\u5ba1\u671f\u520a\u3002

\u66f4\u5e7f\u6cdb\u7684\u62c5\u5fe7

\u9a8c\u8bc1\u6096\u8bba\uff1a Ioannidis \u8d28\u7591\uff0c\u4e00\u4e2a\u58f0\u79f0\u8981\u91cd\u5851\u79d1\u5b66\u7684\u9886\u57df\u5374\u6ca1\u6709\u79d1\u5b66\u6587\u6863\uff0c\u5982\u4f55\u9a8c\u8bc1\u5176\u58f0\u660e\u662f\u771f\u5b9e\u3001\u7ecf\u8fc7\u9a8c\u8bc1\u4e14\u53ef\u590d\u73b0\u7684\uff1f
\u5b89\u5168\u8bc4\u4f30\u7f3a\u53e3\uff1a \u51fa\u7248\u7269\u7a00\u5c11\u4f7f\u5f97\u8bc4\u4f30 AI \u7684\u793e\u4f1a\u5f71\u54cd\u3001\u80fd\u8017\u548c\u5b89\u5168\u98ce\u9669\u66f4\u52a0\u56f0\u96be\u3002