Anthropic Co-Founder Jack Clark Says AI Leaders Privately Fear Their Own Systems
Former OpenAI policy director Jack Clark says executives building frontier models are increasingly alarmed by what those models may soon be able to do.
Jack Clark, a co-founder of Anthropic and previously OpenAI's policy director, said the people running the largest AI labs are not bluffing when they talk about risk. In a Yahoo Finance interview, Clark described executives at the frontier companies as "genuinely worried" about where the technology is heading, separating that private anxiety from the promotional language their marketing teams use.
Clark's vantage point is unusual. He helped shape OpenAI's safety and policy work before leaving in 2021 to start Anthropic with Dario and Daniela Amodei, and he now writes the widely read Import AI newsletter. That puts him in regular contact with researchers and executives across the major labs, including Google DeepMind, OpenAI and Meta, and he says the mood behind closed doors does not match the confident public posture.
The admission lands during a stretch of rapid capability jumps. Models released over the past 18 months have moved from text generation into coding, autonomous task execution and scientific reasoning, and several labs have published evaluations showing performance improving faster than their own forecasters predicted. Clark has argued that this pace compresses the window for governments to write rules before systems become deeply embedded in critical infrastructure.
Safety researchers have long complained that warnings from inside the labs get dismissed as marketing for regulation. Clark's comments invert that dynamic: the concern, he suggests, is coming from the people with the most to gain financially from continued deployment. Anthropic itself operates under a long-term benefit trust and has published scaling policies that commit it to pausing training runs above certain risk thresholds, an approach rivals have largely declined to copy.
Investors have poured tens of billions into AI infrastructure on the assumption that capability gains keep compounding. If the executives writing those checks privately expect trouble, that tension could shape everything from insurance underwriting to enterprise procurement, where legal teams are already asking vendors for indemnification against model failures.
The open question is whether any of this concern converts into action. Clark has called for mandatory third-party audits and reporting requirements for large training runs, but binding rules remain stalled in the US Congress and are only partially in force in the EU. Watch whether the next round of frontier model releases arrives with independently verified safety evaluations, or with the same voluntary commitments the labs have offered so far.