A new analysis suggests many leading AI startups are not actively publishing their research, a practice fundamental to scientific validation and progress. Over half of AI unicorns—private companies valued at over $1 billion—have never played a leading role in publishing a scientific paper or preprint. Collectively, these companies accounted for only one in every 1000 AI papers published in 2025.

This scarcity of documentation raises concerns about the reproducibility and validity of AI claims. Experts like John Ioannidis from Stanford University highlight the "weird paradox" of a field reshaping science without scientific documentation. The lack of publications also hinders assessment of AI’s societal impacts, such as energy consumption and safety.

However, some argue this reflects commercial incentives. Mohamed Abdalla, an AI ethicist, notes that companies prioritize profit over advancing science. The study identified OpenAI as a significant contributor, responsible for nearly 40% of citations in the analyzed data, followed by Megvii and Hugging Face. Even at prolific companies, research output is often concentrated among a few authors.

Alternative forms of disclosure, such as blog posts and technical reports, are increasingly common, though not tracked in this analysis. The debate centers on whether companies release sufficient code, data, or model weights for independent verification. Notably, Chinese startups tend to publish more and embrace open-source models, contrasting with some U.S. firms.