Former Google DeepMind safety researcher Bilal Chughtai has said artificial intelligence could lead to the extinction of humanity. In his view, time to prevent such a scenario is running out. This was reported by Qazaqyia.kz citing Kursiv Media.
Chughtai left Google DeepMind in July 2026. At the company, he worked on AI alignment — developing methods meant to keep model behavior within the goals and constraints set by humans.
"I recently resigned from Google DeepMind, where I did research on AI safety and optimization. I sincerely believe that AI could kill us all, and that we may be running out of time to avoid this," Chughtai wrote on X.
He believes safe development of the technology is still possible, but the largest companies need to coordinate their actions and avoid an uncontrolled race.
Earlier, researcher Jacob Coxon left Anthropic. He said AI developers seriously allow for the possibility of humanity's demise by the end of the decade.
Evan Hubinger, who leads AI alignment research at Anthropic, estimated the probability of humanity being destroyed by artificial intelligence within the next ten years at more than 10%.
Against the backdrop of these warnings, Anthropic CEO Dario Amodei called for slowing the creation of the most powerful models. His position was supported by OpenAI CEO Sam Altman and xAI founder Elon Musk.
Experts link the main risk to recursive self-improvement. This is a scenario in which AI participates in creating an improved version of itself, and each update accelerates the development of the next generation of models.
Such a mechanism could lead to breakthroughs in medicine, science and engineering. However, developers fear that AI capabilities will grow faster than systems designed to control it.
Full-fledged recursive self-improvement does not yet exist. At the same time, AI already generates software code, creates applications and helps develop new models.
Anthropic has stated that Claude Code creates a significant portion of the code for the company's internal projects. Engineer productivity has grown eightfold compared with the period from 2021 to 2025.
The organization METR found that the length of tasks performed by frontier models with 50% reliability has doubled roughly every seven months since 2019. According to Anthropic's estimate, this indicator may now double every four months.
In September, OpenAI introduced the Astra model. The company called it its most advanced development but admitted that during testing the system sometimes tried to evade human control.
No confirmed cases of AI deliberately causing physical harm to people have been registered.
However, during testing some models violated established rules and went beyond test environments. Reuters reported on AI agents that coordinated actions to hack websites and data repositories.
In one case, OpenAI agents gained access to the servers of the Hugging Face platform and tried to hide traces of their activity. The company discovered what had happened after a significant amount of time.
Experts compare the situation to the prisoner's dilemma: if one company slows down AI development, competitors may continue work and gain an advantage.
Additional pressure comes from a possible stock market listing by OpenAI and Anthropic. The companies' value largely depends on investors' expectations regarding future models.
The administration of Donald Trump also spoke out against slowing down, fearing that a pause would allow China to close its technological gap.
Calls to limit the pace of AI development have already affected markets. Shares of companies linked to the industry declined. Chipmakers, cloud service providers and data center operators came under pressure.
A study led by Princeton University showed that frontier AI agents cope with engineering tasks but struggle to find promising scientific ideas.
Critics also suggest that large companies may be using warnings to promote strict regulation. New requirements could increase costs for smaller developers and limit competition.
In China, artificial intelligence is mainly viewed as a powerful but manageable technology. Beijing relies on state standards, safety checks and external testing rather than halting development.