Why are AI's top startups publishing less research than ever?

The AI field's early years, roughly from the early 2010s into the beginning of this decade, were known for a culture of openness. Major labs published their methods as papers, shared code, and built on each other's work. That openness was widely credited as one of the reasons the field advanced so quickly.

Today the picture has shifted. The field's leading startups are sharing less and less of what they consider their most groundbreaking findings. Research that once appeared at academic conferences and in peer-reviewed journals now mostly stays in-house — either never published at all, or shared only as a marketing-oriented blog post stripped of technical detail.

The clearest driver behind the shift is commercial competition. A lab that spends months of research and a massive compute investment to gain an edge has little appetite for handing that advantage to competitors in a public paper. Publishing, once a path to scientific prestige, is now increasingly viewed as giving away a free roadmap.

Adding to that, some companies justify their decision not to publish on safety grounds: they argue that making details of how the most powerful models are trained publicly available could raise the risk of misuse. That argument is contested — critics say safety concerns are sometimes used as a convenient cover for protecting trade secrets.

National security dynamics further complicate the picture. AI is now viewed not just as a commercial technology but as a field of geopolitical competition, pushing some companies — and even governments — toward greater caution about sharing research considered strategically valuable.

This closing-off has concrete consequences for the scientific community. Academic researchers, unable to access leading labs' latest findings, lose the opportunity to independently verify or build on them. That slows the cumulative progress of knowledge — the basic mechanism by which science works.

The problem is even sharper for AI safety researchers. Being able to independently examine how a model behaves and what risks might emerge depends on transparency. When companies conduct their own safety evaluations in isolation, no outside party can confirm those evaluations are accurate.

Still, not every company is moving in the same direction. Some nonprofit research organizations and academic partnerships continue to publish safety-focused findings in particular, and some companies maintain openness in specific research areas that aren't commercially sensitive.

For policymakers, the trend creates a new challenge: regulation generally depends on transparency — assessing a technology's risks requires first understanding how it works. As companies increasingly keep their core research confidential, it becomes harder for regulators to make informed decisions.

Researchers don't expect the field to fully return to its open-science culture, but some are advocating for at least industry-wide norms around sharing safety-relevant findings. That debate will continue to shape how AI develops as both a science and a business.