A Framework for Frontier AI and the Dawning of a New Age

In an X article, Google DeepMind CEO Demis Hassabis proposes a U.S. standards body for frontier AI. He discusses AGI's potential, catastrophic-risk testing, pre-release review, independent audits, international coordination, and how AI's benefits should be distributed.

Demis Hassabis defines AGI as a system with all the cognitive capabilities of the human brain. He expects it within a few years and compares its potential impact with electricity and fire. His estimate of ten times the Industrial Revolution at ten times the speed is a forecast, not a measurement.

He expects AGI to accelerate drug discovery, clean energy and materials research. AlphaFold 3 already predicts structures involving proteins, nucleic acids and small molecules. GNoME generated candidate stable crystal structures. These systems show narrow scientific gains. They do not establish AGI or post-scarcity. A survey of 2,778 AI researchers also found much later median estimates for human-level performance.

Hassabis says capabilities are advancing faster than researchers understand them. He names cyber, biological and nuclear threats, then the longer-term risk of losing control over more agentic or self-improving systems. Commercial and geopolitical competition gives labs less time to study those risks.

He calls for cautious optimism: keep innovation moving while rewarding security, responsibility and international cooperation. The 2026 International AI Safety Report found AI assistance in cyber operations, but no fully autonomous, end-to-end attack against a real target. It reports better biological knowledge performance but uncertain practical uplift. Current systems show some capabilities relevant to loss of control, not the full set needed to cause it.

His main proposal is a US Standards Body. It could be a federally supervised public-private partnership or a self-regulatory organisation modelled on FINRA. Independent technical experts and open-source representatives would sit on its board. Industry would fund much of its staff and compute.

The body would set changing thresholds for “Frontier-class” models. Labs that cross them would publish model cards, strengthen cybersecurity, vet key staff and fund safety research. They would also work with federal agencies and US National Laboratories on national-security tests.

NIST’s Center for AI Standards and Innovation already develops voluntary standards and evaluates national-security capabilities. NIST also connects government, laboratories and industry. A FINRA-style regulator would need statutory authority, federal oversight, stable funding and enforcement powers that these programmes do not have.

Demis proposes giving the Standards Body access to frontier models up to 30 days before release. Participation would begin voluntarily. Passing the review could later become a condition for US deployment. Labs would help fix critical flaws found after release.

Tests would cover cyber, biological, agentic and deceptive capabilities. He also mentions image watermarking and human-readable reasoning as possible standards.

NIST signed testing agreements with Anthropic and OpenAI in 2024. US and UK institutes later ran joint pre-deployment tests. Those voluntary agreements are narrower than a universal 30-day approval rule.

NIST describes watermarking as an imperfect provenance tool, not a test for catastrophic capability. Research from OpenAI and Anthropic also finds that written reasoning can aid monitoring without faithfully showing how a model reached its answer.

Static tests expire. Hassabis wants regular updates, perhaps quarterly at first. Saturated benchmarks would be retired. The Standards Body would build private tests and support independent auditors instead of relying on labs indefinitely.

NIST’s AI Risk Management Framework also treats measurement as continuous. Private tests can still leak, reward narrow optimisation or miss deployment failures. Auditors can depend on labs for access and compute. Independence therefore needs reproducible methods, conflict rules, controlled access and repeated testing before and after release.

Frontier Lab status would carry prestige. Any organisation could qualify by crossing the capability thresholds. The same process would cover foreign, domestic, open and closed models. Startups and academic teams below the frontier would be exempt.

Rules could tighten as risks rise. Hassabis includes a coordinated slowdown among Frontier Labs as an emergency option. Under the voluntary Seoul Frontier AI Safety Commitments, developers agreed to stop development or deployment when severe risks could not be controlled. The commitments did not create an authority that could order or verify a common slowdown.

The EU’s general-purpose AI guidance combines compute thresholds with capability and impact reviews. It also changes some duties for open models. A US system would need similar review procedures and different access rules for open and closed models.

He sees the US body as a first step. Frontier systems will cross borders, so severe-risk standards and access to benefits must also become international.

The International Network of AI Safety Institutes links national technical bodies. NIST has a global standards plan. The OECD runs voluntary reporting. The UN Global Digital Compact addresses governance and capacity. The Council of Europe opened a binding AI treaty on rights, democracy and law.

None forms a global frontier-testing regime. Agreement on tests, terms and reporting may come before agreement on thresholds or enforcement. Other countries and affected communities would need power over the rules. Access also requires compute, infrastructure and trained institutions.

Hassabis ends with the choices that remain after technical safety. Who receives the gains? Which economic systems fit a less resource-constrained world? Which values guide society? What gives people purpose?

Current evidence is mixed. A study of 5,179 customer-support agents found higher average productivity with an AI assistant, especially among less experienced workers. Another macroeconomic estimate projected smaller gains from current systems than transformational forecasts imply. The IMF estimates that AI exposes almost 40% of global employment and may deepen inequality.

He says technologists cannot make these decisions alone. UNESCO’s AI ethics recommendation puts dignity, rights and broad participation at the centre of governance. The UN has created an Independent International Scientific Panel and Global Dialogue on AI Governance. Neither settles distribution or enforcement. They give more people a formal role in the decisions.

Subject: Demis Hassabis’s plan for governing frontier AI

Preview text: A US standards body, pre-release tests and international rules.

Demis Hassabis expects AGI within a few years. He thinks it could accelerate science, medicine, energy and materials research. Current systems show narrower gains. They do not prove AGI or post-scarcity.

He also sees cyber, biological, nuclear and future control risks. Commercial and geopolitical competition gives labs less time to understand them.

His answer is a US Frontier AI Standards Body. Government would supervise it. Industry would fund much of it. Independent technical and open-source representatives would help govern it. The body would set thresholds for frontier models and work with federal agencies and National Laboratories.

Qualifying labs would strengthen cybersecurity, vet key staff and fund safety work. They would initially provide models for review up to 30 days before release. Passing could later become a condition for US deployment.

Tests would cover cyber, biological, agentic and deceptive capabilities. The body would replace saturated benchmarks, build private tests and support outside auditors. Existing government evaluations cover parts of this work but remain voluntary and limited.

The same frontier rules would cover foreign, domestic, open and closed models. Lower-capability startups and academic teams would be exempt. Rules could tighten as risks rise, including a coordinated slowdown. That would require clear authority and a way to stop labs from defecting.

Hassabis wants the US system to seed international standards. Existing institutes, treaties and UN programmes provide pieces, not a global testing regime.

Technical safety would still leave decisions about distribution, economics, values and human purpose. Demis says society must make them together before the technology sets the terms.

Read the original article