AI Unicorns Rarely Publish; 5% of Firms Own 90% of Citations

A new analysis puts a number on something a lot of us have felt for a while, that the loudest labs in AI have basically stopped showing their work. Stanford metascientist John Ioannidis and his co-authors trawled every AI startup that has hit unicorn status between 1998 and 2025, all 317 of them, and looked at how often their researchers led a paper as first or last author. The dataset covers 2,077 publications, 1,389 peer-reviewed and 688 preprints. More than half of the unicorns have never produced a single qualifying paper, and the group collectively accounts for roughly one in every thousand AI papers published in 2025, as Science reported from a new bioRxiv preprint.

The influence numbers are the more striking part. The top 5% of firms account for more than 90% of citations in the dataset. OpenAI alone is responsible for nearly 40%, followed by the Chinese computer vision company Megvii and the platform Hugging Face. So the shape is not just that unicorns don't publish, it is that the field's scientific record is a handful of firms and everyone else is dark.

Why this matters if you are not a researcher: when there is no literature, there is nothing outside a company to check its claims against. Ioannidis makes the direct comparison to his 2015 scrutiny of Theranos, the blood testing startup that turned out to be built on fraudulent data, where the tell was the same absence of peer-reviewed work. Nobody is saying today's unicorns are Theranos. The point is that when energy footprints, safety behaviour, and capability claims all live in blog posts and marketing rather than in papers other people can pull apart, the outside world loses its main tool for pushback, and the study notes this scarcity makes it harder to assess AI's social impacts including energy use and safety.

The honest caveat is that the study measures leading-author output, so a company could be doing serious science internally and simply keeping it as trade secret; the reporting available so far does not disentangle those two motives, and does not break the pattern down by country or year. Ioannidis himself calls the situation a 'very weird paradox' for a field 'supposedly reshaping science.'

The forward-looking read is that whoever does publish, at any scale, will look increasingly unusual, and that gives the research-oriented labs, the academic spinouts, and any regulator wanting disclosure standards a fairly clean lever to pull.