In Brief
Anthropic researcher Jacob Coxon resigned, warning that his company and former employer OpenAI were racing toward AI they might not be able to control.
Within days, leaders of competing labs publicly supported slowing capability advances with Anthropic and OpenAI committing to independent evaluators inside their companies.
What Happened
Coxon had spent three years training AI models at OpenAI and Anthropic when he announced his resignation. He warned that competition was pushing both toward self-improving systems beyond human control. He told The Wall Street Journal that even earnest safety efforts were insufficient without government intervention or a coordinated slowdown.
Anthropic CEO Dario Amodei then published We Must Pace the Frontier, proposing a slower rate of improvement for the most advanced AI, so safety work can catch up. His three-part plan starts with independent evaluators inside labs, followed by coordination among companies in democratic countries and broader international cooperation.
Anthropic committed to the first step. Its proposed outside reviewers would have ongoing access comparable to internal risk teams and could publish findings without company editorial control, subject to narrow restrictions on sensitive information. OpenAI CEO Sam Altman committed to independent evaluators, too. Elon Musk endorsed Amodei, without announcing the same evaluator commitment, while Demis Hassabis supported the direction but said details needed work.
These announcements establish support and initial commitments with an industry-wide pacing agreement still to be negotiated.
SmarterX founder and CEO Paul Roetzer examined the shift on Episode 239 of The Artificial Intelligence Show.
The Key Numbers
3 - Years of research experience Coxon has training models at OpenAI and Anthropic
170 million+ - Views of Coxon's resignation post by the episode's recording
3 - Steps Amodei proposes for pacing frontier AI
2 - Number of labs (Anthropic and OpenAI) who publicly committed to embed independent evaluators inside their labs
Why AI Building AI Changes the Safety Debate
The concern is an accelerating development cycle. Recursive self-improvement means AI helps design and improve the next generation of AI. Amodei says that process is beginning across the industry and could advance faster than researchers' ability to understand and control it. Roetzer believes the alarm inside the labs reflects capabilities the public has not yet seen: "And I think the researchers and lab leaders are spooked for real," he says.
Other risks besides human extinction. Roetzer worries that predictions about human extinction are swallowing the discussion of more tangible dangers. He points to the possibility of AI systems acting on their own to disrupt power grids or financial markets as risks worth testing for.
Safety needs resources and oversight. Roetzer supports slowing releases while labs improve alignment, the work of keeping AI behavior consistent with human intentions. He says, "They need way more resources dedicated to safety and alignment, and there has to be some level of government oversight." His argument draws on a familiar expectation: powerful technologies should face scrutiny before society absorbs the consequences of their failures.
"I do think that we have to slow down the model releases."
— Paul Roetzer, founder and CEO of SmarterX on Episode 239 of The Artificial Intellgence Show
SmarterX Take
Business leaders still have substantial work to do with the AI already available. Roetzer estimates that even if model development stopped today, most enterprises would need three to five years to fully integrate current capabilities.
His advice is to keep learning and applying AI responsibly, with leadership focused on making work more fulfilling and giving people more time for judgment and creativity. The debate over the next model should sharpen that responsibility. It should also give organizations a reason to get much better at using the tools they already have.
What to Watch
The openness of the labs to real scrutiny. Evaluation teams will need access, independence, and the ability to publish unfavorable findings to be effective.
International cooperation remains unresolved. Will competing labs and governments, including China, accept and comply with shared limits? Endorsements are an opening step toward that harder negotiation.
How Prepared Is Your Company to Govern AI?
Only 13% of respondents in the governance analysis for the 2026 State of AI for Business Report say their organizations have all four foundations: an AI council, an AI roadmap, generative AI policies, and an AI ethics policy. As labs debate stronger oversight, that finding brings the question of responsible adoption back to the organizations deploying their tools.
The full report draws on more than 2,100 professionals across roles, functions, and industries, primarily from an audience already interested in AI education. It offers a benchmark for assessing adoption, training, and governance inside your own company. Read the full report →
Mike Kaput
Mike Kaput is the Chief Content Officer at SmarterX and a leading voice on the application of AI in business. He is the co-author of Marketing Artificial Intelligence and co-host of The Artificial Intelligence Show podcast.