Google DeepMind's Hassabis urges urgent regulation before AGI arrives
AGI is 'few years away' — and even its builders are nervous
Demis Hassabis, CEO of Google DeepMind, published an essay on X arguing that AGI (artificial general intelligence) is 'probably only a few short years away' and will be humanity's greatest achievement — but also a profound risk. He warns that current frontier models already pose cybersecurity threats, and risks from nuclear or biological applications may emerge as capabilities advance. 'Urgent action is needed,' he wrote, comparing AGI's impact to the discovery of electricity or fire rather than the internet.
To address these dangers, Hassabis proposes creating a US-led standards body funded by industry but overseen federally. Labs would voluntarily submit models for pre-release testing up to 30 days before deployment, with evaluations covering cybersecurity, biological threats, deception, and guardrail evasion. If effective, approval could become mandatory. The body could also coordinate a slowdown if risks escalate. This call for regulation is notable because Hassabis is one of the leading figures racing toward AGI, highlighting growing unease even among its builders.
- Hassabis estimates AGI is only 'a few short years' away, calling it a 'pivotal moment in human history'
- He proposes a US-led standards body with voluntary pre-release testing that could become mandatory, modeled on nuclear oversight
- Risks cited include cybersecurity, biological weapons, nuclear threats, and AI agents that bypass safeguards or improve themselves
Why It Matters
Even the leading AGI builder is calling for oversight — the race is outpacing safety.