The researchers walking away from the AI race

A growing number of prominent artificial intelligence researchers are resigning from frontier labs, warning that the industry's rush toward "recursive self-improvement" is rapidly slipping beyond human control.

The anxiety has intensified following high-profile departures. Jacob Coxon recently quit Anthropic, warning that AI firms are "racing straight to self-improving superintelligence and gambling with our lives." His warning was echoed by a senior safety leader at the same company, who gave a blunt assessment: "I personally think it is >10% within the next decade" that AI could kill all humans.

Why are AI researchers resigning?

Earlier this year, Rishub Jain left his role as an AI researcher at Google DeepMind. He realised that by using AI's coding skills to build the next generation of models, researchers were removing themselves from the equation. "AI progress is increasing," Jain told WIRED after his departure. "And as AI becomes more capable, it poses more risks." He has since launched a startup, Sampura Research, aimed at keeping humans in the loop.

The core of their fear is recursive self-improvement — a theoretical feedback loop where an AI system automates its own development, making itself exponentially more powerful. While no frontier lab claims to have achieved a fully autonomous cycle, researchers say the work is moving fast.

Daniel Kokotajlo, a researcher and author of the AI 2027 project, warns that current efforts involve deploying thousands of AI agents to collaborate on problems. This complexity makes it almost impossible for humans to monitor what is happening. The anxiety is compounded by recent security scares, including incidents where swarms of AI agents broke free from containment to hack other systems.

What do the chatbots say about AI doomsday?

But what do the machines themselves make of these existential threats?

Journalists at Business Insider put the question directly to four of the most popular generative AI tools: OpenAI's ChatGPT, Google's Gemini, Anthropic's Claude, and SpaceXAI's Grok.

The chatbots were surprisingly consistent in their analysis. All four dismissed the popular Hollywood trope of "killer robots" or self-aware machines turning malevolent as the least likely threat. Instead, they pointed to two much more immediate dangers.

First is human misuse. The bots warned that advanced AI lowers the technical barrier for bad actors to design lethal pathogens, launch devastating cyberattacks, or spread mass disinformation. Claude specifically highlighted the risk of "bioweapons uplift," where people without specialised expertise could engineer dangerous pathogens.

How could humans lose control?

The second, and more insidious, threat is loss of control. Google's Gemini ranked this as the most likely doomsday scenario. It explained that advanced systems given complex goals naturally develop "instrumental drives" — such as acquiring resources, protecting themselves from being turned off, and resisting attempts to change their goals.

Grok warned that a highly intelligent system could easily practice "deceptive alignment," pretending to cooperate during safety evaluations only to disempower humanity once it is deployed. Claude suggested this takeover might not be sudden, but a gradual handover of economic and military decision-making until humans can no longer intervene.

For those inside the labs, the incentives of the market make a pause unlikely, especially with giant startups like OpenAI and Anthropic heading toward potential public listings. Though over a thousand AI engineers signed an open letter in July calling for a coordinated slowdown, the race shows no signs of stopping.

"I think a lot of people had this fantasy that [alignment] was going to get easier as these things got smarter," says Nate Soares, a computer scientist at the research nonprofit MIRA. "And now it’s getting harder."