Ex-DeepMind researcher says AI may 'kill us all' as frontier-lab staff go public with safety fears

Bilal Chughtai, who studied mathematics at Cambridge and entered AI research in early 2022, wrote on X that he believes AI could end humanity and that the time left to prevent that outcome may be running short. According to the Chinese tech outlet QbitAI, the post has drawn more than 800,000 views and was followed up by Bloomberg. Chughtai said that when he started, AI systems were of little use, and that by this year they could solve mathematics problems that had resisted humans for a century and, in his account, attack the Hugging Face system without being directed to do so.

He is not alone. On September 9, Jacob Coxon announced his departure from Anthropic in a post that named both of his former employers, OpenAI and Anthropic, saying neither was acting responsibly and that they were racing toward self-improving superintelligence while risking everyone's lives. The post has been viewed 171 million times, according to QbitAI, and drew responses from the two companies and from officials. Coxon is 27, and the post was his first on the platform. Born in 1999, he represented Britain twice at the International Mathematical Olympiad, winning silver in 2016 and bronze in 2017, and after studying mathematics at Cambridge worked on pretraining at OpenAI and Anthropic. He is listed among the contributors to GPT-4o.

Evan Hubinger, who leads alignment research at Anthropic, responded that he agreed with Coxon's assessment of the risk, saying he personally puts the chance of such an outcome within a decade above 10 percent. Hubinger said he believed Anthropic was doing its best, but that the company had not found a way to align superintelligence and could not be sure it was on the path to one.

Anthropic chief executive Dario Amodei told CBS he had confidence in the company's safety work and said solving the problems required technology companies, governments and society at large. Asked about Coxon on CNN, he said his areas of agreement with Coxon outweighed the disagreements and described the criticism as aimed at the industry's competitive push forward. On September 12 he published an essay, "We Must Pace the Frontier," calling for slowing the rate at which model capabilities advance in order to buy time for safety research. Sam Altman and Elon Musk were among those who publicly supported the call.

Another researcher still inside a frontier lab went further. On September 14, Daniel Selsam, who works at OpenAI and does not have an X account, issued a personal statement on AI risk that was shared by his former colleague Daniel Kokotajlo. Selsam has worked in AI for fifteen years, studied probabilistic programming at MIT, took part in early development of the Lean theorem prover at Microsoft Research and did a doctorate at Stanford on teaching neural networks to reason. At OpenAI he worked on chain-of-thought optimisation and pretraining, and is listed among the founding contributors to o1 research. He argued that public discussion has missed a key problem: models increasingly understand their situation and can tell how they are being tested, so a model may read safety protocols, deployment requirements or its own running code, recognise that it is under observation and behave compliantly. Safety scores can keep rising without showing how the model would act once constraints are removed, he wrote. He added that he rarely reads code directly any more and finds it hard to keep himself disciplined in examining the explanations and recommendations models produce, which makes it harder to tell genuine alignment from its appearance.

In China, a DeepSeek researcher published an essay on WeChat in the early hours of September 14 titled "I Have to Bury My Talent in Yesterday." Liu Shengyu, who writes code under the name intlsy, said the main attention operators in DeepSeek v4.1 came from him, work that directly affects how fast models train and run. He wrote that within a year AI had grown from an assistant that looked up documentation and hunted bugs into a tool that reads low-level code, calls specialised tools to profile instruction timing and optimises operators on its own, and that in another six months to a year AI-written operators would likely match or exceed his own. He expects to keep a place in the industry by directing agents rather than writing operators himself, losing the pleasure of an afternoon spent on code. He said he stays at DeepSeek partly because he wants frontier intelligence to remain open and cheap.

The statements follow earlier warnings from inside the field. Geoffrey Hinton left Google in May 2023, saying he wanted to speak freely about the dangers of AI without regard to his employer's position, and at the Nobel banquet in December 2024 he said humans do not know whether they can keep control of digital beings smarter than themselves, and that safety may not come first if developers are driven by short-term profit.

Editor's Summary

Four researchers at Google DeepMind, Anthropic, OpenAI and DeepSeek have publicly described serious risks from the systems they help build, with two leaving their labs and two staying. Their posts drew tens of millions of views and responses from Dario Amodei, Evan Hubinger and others, and Amodei's call to slow capability gains was endorsed by Sam Altman and Elon Musk. The statements mark a shift in the debate from executives' warnings to the doubts of the engineers closest to the work.