Background & Context§
The artificial intelligence industry was built on a foundation of open research. The seminal 2017 paper "Attention Is All You Need" by Google researchers introduced the transformer architecture that powers today's generative AI boom. Without that publication, there would be no GPT, no Claude, no Llama. Yet as the field has commercialized, a growing number of startups are choosing to keep their work secret. A recent study published in Science quantifies this shift for the first time, finding that approximately half of AI unicorn startups do not publish any research at all. This marks a fundamental break from the academic ethos that seeded the field and raises urgent questions about reproducibility, accountability, and the trajectory of AI progress.
The News: What Happened Exactly§
The Study's Findings§
Researchers analyzed the publication and citation records of the top 50 AI startup unicorns—private companies valued at over $1 billion—and found that half of them had not published a single research paper in the prior year. Even among those that did publish, output was concentrated in a handful of firms. OpenAI led all startups in cumulative citations, followed by MEGVII, Hugging Face, Waymo, Momenta, Preferred Networks, Anthropic, Owkin, Databricks, and Aibee. Notably, the study used citations as a proxy for research significance, acknowledging the metric's imperfection but arguing it captures impact better than raw publication counts.
The Irony of Closed Science§
The finding strikes at a peculiar irony: the entire AI revolution was launched by Google's decision to publish its transformer research openly. As one Hacker News commenter put it, "Yet none of them would have been here if Google hadn't published 'Attention is all you need', the irony." The comment reflects a broader sentiment that the industry is now free-riding on the open research of the past while contributing little back. Another commenter noted, "All startups barely publish their research, because doing that would be incredibly stupid for the most part, because they want to sell the stuff they invent not give it away for free." This tension between commercial secrecy and communal knowledge is the crux of the controversy.
Voices of Concern§
The study's authors, led by economists and science policy researchers, express alarm beyond the numbers. One researcher quoted in the article says, "If we were racing forward on cancer-curing AI, I would be like, 'Fantastic, full steam ahead,' but that's not what we're racing toward, right?" This echoes the broader worry that without publication, there is no way to validate claims or ensure that progress is genuine. The article further notes that many AI startups are not research labs but product and marketing layers over existing models, and asking them to publish papers may be unrealistic. However, the lack of transparency has real consequences: it enables hype cycles, makes reproducibility impossible, and lets problematic practices go unchecked.
The Shift Toward Blog-Driven Research§
A parallel trend is the "blogification" of AI research. Increasingly, companies publish findings as informal blog posts or model cards rather than peer-reviewed papers. This allows claims to spread through social media dynamics, with minimal verification. One commenter described the result as "not dissimilar to setting termites loose in a library." The combination of closed labs and unvetted blog posts creates an environment where any claim can be backed by cherry-picked numbers from gamified benchmarks, degrading the quality of the collective knowledge base.
What the Study Does Not Say§
It's important to specify what the study does not claim. The paper does not name OpenAI and Anthropic as offenders; in fact, both are highlighted as companies that do publish. As one Hacker News participant clarified: "The paper never actually mentions the companies who aren't publishing papers. OpenAI, Anthropic, and Hugging Face are all specifically mentioned as companies that do publish papers." The analysis focuses on the set of 50 unicorns, and the 50% publication rate is actually higher than many commenters expected. One noted, "50% of startups contributing to public research is actually a crazy good outcome. That's far more than I had expected." This nuance complicates the narrative of a wholesale retreat from openness.
Historical Parallels & Similar Incidents§
The Clash Between Bell Labs and Xerox PARC§
The current situation echoes a historical tension in industrial research. Bell Labs, the iconic research arm of AT&T, produced world-changing inventions like the transistor, the laser, and the UNIX operating system—all while publishing openly and allowing researchers to pursue fundamental science. In contrast, Xerox PARC invented the graphical user interface, the mouse, and Ethernet, but Xerox failed to commercialize these breakthroughs, in part because of a culture of secrecy. The inventions eventually escaped to Apple and Microsoft, generating billions in value for others. The lesson: closed research can stifle both the commons and the originating company's ability to capture value. Today, AI startups face the same choice: contribute to the open ecosystem and risk losing competitive advantage, or keep secrets and risk slowing the entire field.
The Software Industry's Shift from Open Source§
A more recent parallel is the software industry's retreat from open source. In the 2000s and early 2010s, open-source software was the norm, with companies like Red Hat building billion-dollar businesses on open models. More recently, companies like MongoDB, Elastic, and HashiCorp have switched to non-open licenses (e.g., SSPL, BSL) to prevent cloud providers from exploiting their work without compensation. This shift was driven by the same commercial logic: investors demand proprietary moats. AI startups' reluctance to publish papers mirrors this trend. Both cases demonstrate that when capital enters a field, the incentive structure flips from reputation-building (academic) to rent-seeking (commercial). The counterargument is that without publishing, the field may suffer from fragmentation and duplication of effort, just as the software world now deals with license incompatibility and vendor lock-in.
The Lesson: Trade Secrets Are Back, Baby§
One Hacker News commenter encapsulated the shift: "Trade secrets are back, baby!" This is not necessarily a condemnation. In many industries—pharmaceuticals, aerospace, defense—trade secrets and proprietary know-how are the norm. The unique aspect of AI was that its early advances came from open research, and many practitioners feel a moral obligation to continue that tradition. However, as the field matures, it may simply be converging to standard industrial practice. The danger lies not in secrecy itself but in the erosion of trust and verifiability. Without published research, the community loses the ability to scrutinize methods, reproduce results, and build upon each other's work. The outcome could be a slower, less innovative future for AI.