AI’s Most Valuable Startups Are Barely Participating in Science

Science reports that more than half of the world’s billion-dollar AI startups have produced no qualifying scientific paper or preprint.

Researchers examined 317 privately held AI companies valued at $1 billion or more. They found that 52.4 percent had no qualifying scientific output, while only 24 companies had produced a paper receiving at least 200 citations. The top 10 percent of companies accounted for 96.8 percent of all citations, and startup publications represented only 0.1 percent of the wider AI literature in 2025.

That is disappointing.

The current AI industry was built on decades of openly published research, shared datasets and ideas that other researchers could challenge, reproduce and improve. Many of the companies benefiting from that work are now placing their own discoveries behind product announcements, selective benchmarks and carefully controlled demonstrations.

Companies do not need to publish every trade secret, and publication counts are not a perfect measure of useful research. But extraordinary technical claims should come with enough evidence for independent researchers to examine them.

You cannot ask society to trust your models, valuations and predictions about the future while withholding most of the work needed to determine whether your claims are real.

The underlying study is currently a preprint and has not yet completed peer review, so its conclusions should also be examined critically. Even with that qualification, it captures a troubling change. AI is moving away from a shared scientific discipline and toward a collection of privately held capabilities that outsiders are expected to accept on faith.

Read “AI’s top startups are barely publishing their research.”