News Analysis
Guide
Is AI Making Your Organization More Productive or More Dependent?
Key Takeaways
- Demis Hassabis says AGI could arrive within a few years, raising the urgency around frontier AI oversight.
- His proposal calls for a US standards body to test advanced models before release.
- Evaluations would target cyber, biological, deceptive and agentic risks, not just performance.
Artificial general intelligence (AGI) could arrive within the next few years, but the systems may advance faster than we can safely evaluate them, according to Google DeepMind CEO Demis Hassabis.
In a new framework outlining his vision for governing frontier AI, Hassabis argued that the industry has entered a critical phase where rapid innovation must be matched with equally rigorous safety testing. His proposal centers on creating a new US-based Frontier AI Standards Body responsible for independently evaluating the capabilities and risks of the world's most advanced AI models before they are deployed.
The proposal comes as frontier AI developers race to build increasingly capable foundation models while governments around the world debate how those systems should be regulated.
AGI Could Arrive Within Years, Hassabis Warns
Hassabis described AGI as a system capable of matching the full range of human cognitive abilities, arguing that such technology would represent one of the most significant breakthroughs in human history.
“When we look back on this time in the decades to come, I think we will realise we were standing in the foothills of the singularity - nothing less than the dawning of a new age for humanity,” wrote Hassabis.
While acknowledging the uncertainty surrounding exactly when AGI will emerge, he suggested its arrival is likely only a few years away, and compares its long-term significance to foundational discoveries like electricity or fire.
If (big if) developed responsibly, according to Hassabis, AGI could dramatically accelerate scientific discovery, including drug development, advanced materials research and clean energy innovation, while delivering productivity gains on a scale far beyond previous industrial transformations.
At the same time, he said, the pace of frontier AI development has begun to outstrip society's ability to fully understand or evaluate its consequences.
“Nobody in the world knows for sure what is going to happen from here, and even the experts disagree,” he noted.
The Race for AI Leadership Is Leaving Safety Behind
According to Hassabis, the AI industry is currently driven by overlapping commercial and geopolitical competition that rewards rapid capability improvements.
While that competition has accelerated innovation, he argued that it’s also created an environment where frontier models are advancing faster than researchers can fully evaluate emerging risks.
Some of those risks are already becoming visible, particularly in cybersecurity. He warned that future frontier systems could also introduce challenges involving biological research, nuclear security and increasingly autonomous AI agents capable of pursuing complex objectives.
Rather than slowing innovation broadly, Hassabis advocated what he describes as "cautious optimism" — continuing AI development while establishing stronger technical safeguards and governance frameworks.
Proposed Oversight Body Would Test Frontier Models Before Launch
At the center of the proposal is a new Frontier AI Standards Body modeled after organizations such as the Financial Industry Regulatory Authority (FINRA), combining government oversight with substantial industry participation.
The organization would develop standardized evaluations for frontier AI models and work alongside federal agencies and US national laboratories to assess systems that could present national security or public safety risks.
Under the framework, models that exceed capability thresholds established by the standards body would be designated "Frontier Models," while the organizations developing them would become "Frontier Labs."
Those organizations would be encouraged to adopt a set of operational best practices, including:
- Publishing technical model documentation
- Maintaining robust internal cybersecurity protections
- Vetting personnel with access to frontier systems
- Investing dedicated resources in AI safety and security research
“Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release,” wrote Hassabis. “Once the assessment protocol is shown to be effective and robust, formalisation could quickly follow, meaning that Frontier Models would be required to pass it to be deployed in the US market.”
Evaluations Would Target Cyber, Bio & Agentic Risks
Rather than focusing solely on model performance, the proposed testing framework would examine how advanced AI systems behave under potentially dangerous conditions.
Assessments could include:
- Cybersecurity capabilities
- Biological knowledge
- Attempts to circumvent safety guardrails
- Signs of deceptive behavior
Hassabis also proposed evaluating technical safeguards, such as AI-generated image watermarking and methods that improve transparency into model reasoning.
Benchmark tests, he added, would require continual updates, potentially on a quarterly basis. Over time, the standards body would develop its own independent evaluations rather than relying primarily on benchmarks created in collaboration with AI developers.
The proposal also envisions a broader ecosystem of third-party auditors helping validate frontier systems and develop new testing methodologies.
“[This approach] is designed to keep up with the field’s acceleration and adapt to the biggest risks as they are identified, and could be ratcheted up if the seriousness of the situation demands…” wrote Hassabis.
The Future of Frontier AI Remains Uncertain
Time will tell whether policymakers will adopt Hassabis' proposal, but it adds to a growing list of calls for structured oversight as AI developers move closer to systems capable of performing increasingly autonomous work.
From Automation to Intelligence: How Leading Teams Are Rethinking Operations
Register
Ai4 2026
Register
How to Win the War for Agentic Citations: The AEO Playbook You Need Now
Register
Content Leaders Collective: Proving Content's Business Impact Starts With the Right CCMS
Watch Now
“There is both huge excitement and uncertainty around AI, and both are warranted,” wrote Hassabis. “But the future is not yet written, we must use this precious window before AGI arrives to shape this technology for the benefit of all humanity.”
Editor's Note: The political atmosphere around AI is only heating up…
- Trump’s AI Order Tests Big Tech’s Appetite for Oversight — Trump's scaled-back AI order keeps reviews voluntary, but frontier AI companies fear the next step may be mandatory approval.
- Anthropic CEO Accuses OpenAI of 'Safety Theater' in Pentagon AI Deal — Anthropic's CEO unloads on OpenAI in a leaked memo, accusing the rival lab of misrepresenting its Pentagon AI deal.
- 'Right-to-Compute' Laws May Be Coming to Your State This Year — Montana passed the first "right-to-compute" law shielding AI from regulation. Now similar bills are spreading.