Tyler Cowen Warns of AI “Unreleased Models” Deepening Secrecy Gap Between Labs and the Public

Tyler Cowen, filling in as a guest host on a New York Times podcast, used the episode to probe a growing concern in artificial intelligence: the widening divide between what leading labs know about their most capable systems and what the public is allowed to see.

Cowen framed the issue around “unreleased models”—systems that exist inside companies but are not made available to outside researchers or the general public. He noted that firms often withhold models because releasing them could be expensive, could require significant safety work, or could pose risks the companies are not prepared to manage. The result, he suggested, is that outsiders are left trying to understand AI progress with only partial information.

In conversation with AI commentator Dwarkesh Patel, Cowen pressed for an “insider” perspective: if many powerful systems remain behind closed doors, how should the public update its view of how fast AI capabilities are advancing?

Patel argued that the present situation is uniquely difficult because some people inside AI labs say they have concrete reasons to be more alarmed than the public is—yet they are constrained from sharing details. That dynamic, he said, leaves policymakers, journalists, researchers and the public attempting to gauge risks without access to the evidence that is allegedly driving private concern.

As an example, Patel described a recent discussion with someone at Anthropic who predicted that AI systems could surpass humans at AI research within nine months. Patel said the person presented the forecast as grounded in internal information rather than mere intuition. But when Patel asked what specifically supported the claim—whether it was a new model, a particular internal evaluation, or another form of testing—he could not get a public explanation.

Patel called this a “terrible” state of affairs, arguing that if insiders truly believe the world is less than a year away from automating AI research—potentially accelerating progress dramatically—then the underlying evidence should be released so external experts can assess it. Without that transparency, he suggested, outsiders have no reliable way to calibrate how concerned they should be.

He also outlined a potential feedback loop that could make the information gap worse over time. If AI systems increasingly help automate the process of building better AI, Patel said, labs may become even more reluctant to release models that approach their highest capability levels. Releasing advanced systems publicly could, in that view, make it easier for others to develop similarly powerful models. And, Patel added, shipping models to the public is itself burdensome, requiring safeguards and other infrastructure that companies may be disinclined to build for anything close to their top-performing systems.

Cowen summed up the implication: under that scenario, the distance between internal knowledge and external understanding would continue to grow. That would mean more uncertainty for regulators, lawmakers and other decision-makers—and greater chances of being caught off guard by sudden capability jumps.

The episode also nodded to a wider range of topics, including “AI psychosis,” and Cowen closed by crediting Patel for engaging from the interviewee’s seat rather than the host’s.

Why it matters

If the most advanced AI models remain hidden, outside experts and governments may be forced to make high-stakes policy decisions with incomplete data. A growing secrecy gap could increase the risk of surprise breakthroughs—especially if AI begins to meaningfully automate AI research itself. Watch for whether labs provide more public evaluations and evidence to support their strongest private claims about near-term timelines.