Former DeepMind Researcher Warns AI Race Could Put Humanity at Risk
A former Google DeepMind researcher has warned that advanced artificial intelligence could threaten humanity’s survival, adding to public concern among researchers about the pace of development.
Bilal Chughtai, who worked on artificial general intelligence safety and alignment, made the warning on September 14. According to Reuters, his LinkedIn profile shows he left DeepMind in July 2026.
Chughtai said safe development remains possible but requires coordination between competing companies and a pace that allows society to address emerging dangers. Google had not responded to Reuters’ request for comment outside regular business hours.
His comments follow concerns raised by researchers at Anthropic. Earlier this month, Jacob Coxon announced his resignation, criticising the race to develop systems capable of improving themselves. Coxon said he had worked on pretraining research at both OpenAI and Anthropic.
Anthropic alignment science lead Evan Hubinger subsequently expressed his own concern, estimating a greater than 10% chance of an AI-driven human extinction outcome within the next decade. That figure represents his personal assessment, rather than a measured probability or an agreed scientific forecast.
Hubinger also distinguished between present systems and the more advanced technology he fears could emerge. He described the risk from current models as low, while warning about future superintelligence arising through recursive self-improvement. He said Anthropic was making an effort but had not yet solved the associated alignment challenge.
The technical debate centres partly on whether increasingly capable AI systems can remain reliably under human control. Google DeepMind’s own published safety work describes alignment as training AI to behave safely and helpfully. Its approach also recognises that training alone may provide incomplete protection.
In June, DeepMind outlined an AI Control Roadmap that adds layers of security around advanced agents. These include restricting access, monitoring behaviour and intervening when systems attempt potentially harmful actions.
The framework treats an agent as a possible internal security risk even when it has been trained to follow instructions. Permissions are intended to expand gradually, based on verified behaviour, while safeguards become stronger as capabilities increase.
DeepMind also reported analysing one million coding-agent tasks to improve monitoring. It said most flagged events reflected misunderstanding or excessive eagerness to complete a task, rather than hostile intent. That distinction matters when interpreting unusual AI behaviour: an operational failure does not, by itself, demonstrate a deliberate attempt to evade human control.
The roadmap calls for collaboration between industry, policymakers and researchers on security practices and standards.
Taken together, these statements reveal two connected questions: how severe future AI risks could become, and what evidence companies should provide before giving systems greater autonomy. The warnings describe possible future outcomes; they do not establish that human extinction is inevitable or that today’s tools can cause it.
For readers, the distinction between a researcher’s forecast, an observed incident, and a company’s proposed safeguard is essential. Public statements can draw attention to unresolved problems, but assessing progress requires scrutiny of how protections are tested and whether they work under demanding conditions.
