Hassabis proposes US AI regulator modeled after FINRA

Key Takeaways

• Hassabis proposes a new US AI standards body modeled after financial regulator FINRA, starting voluntary and becoming mandatory
• The framework exempts non-frontier models from startups and academia to avoid regulatory capture accusations
• DeepMind's co-founder Shane Legg considers a minimal AGI possible as early as 2028

Demis Hassabis, CEO of Google DeepMind, has published a detailed AI governance framework proposing a new US standards body modeled after the financial regulator FINRA. The agency would develop evaluation protocols for frontier AI models, fund itself through industry contributions, and eventually make compliance mandatory. Hassabis argues that AGI is likely just a few years away, with impacts ten times greater than the Industrial Revolution arriving ten times faster.

The proposal comes weeks after Hassabis declared humanity "in the foothills of the singularity," a statement that drew both agreement and sharp criticism from peers. His framework attempts to thread the needle between enabling continued AI development and preventing catastrophic outcomes nobody can reliably predict.

What does the proposed AI standards body look like?

The FINRA model is deliberate. The Financial Industry Regulatory Authority operates as a self-regulatory organization, funded by the securities industry it oversees but with real enforcement teeth. Hassabis envisions something similar for AI: an agency that develops regularly updated benchmarks for frontier models, starts with voluntary participation from major labs, and transitions to mandatory compliance as the field matures.

Critically, the framework exempts non-frontier models. Academic research projects, startup experiments, and open-source efforts below a certain capability threshold would not face regulatory burden. This design choice directly addresses the "regulatory capture" accusation that has dogged previous governance proposals. Critics have long worried that Google, OpenAI, and Anthropic might support regulation precisely because it would crush smaller competitors under compliance costs.

Hassabis also suggests the body could coordinate a development slowdown if necessary, pointing to Anthropic's recent public consideration of similar measures. International coordination would follow, with the US agency setting standards that other nations could adopt or adapt.

Why now? The timing reveals the stakes

Days before Hassabis published his framework, a group of prominent AI researchers and economists released a letter warning of sweeping consequences from AI-driven job losses. Hassabis did not sign that letter, though his own rhetoric about the scale of disruption sounds nearly identical. The difference: his proposal focuses on countermeasures rather than alarm bells.

“Nobody in the world knows for sure what is going to happen from here, and even the experts disagree. When there is a large degree of uncertainty and the stakes are this high, proceeding with cautious optimism is the sensible and correct strategy.”

— Demis Hassabis, CEO of Google DeepMind

That expert disagreement is not hypothetical. Last December, Yann LeCun publicly called the concept of general intelligence based on language models "complete BS" and "completely delusional." Hassabis fired back, saying LeCun was "just plain incorrect." The spat highlighted how even the field's most credentialed figures cannot agree on basic questions about what current AI can and cannot become.

Where do other AI leaders stand?

Gemini co-lead Oriol Vinyals offers a middle position: today's models are strong in narrow domains, but genuine innovation capacity remains absent. Deep learning pioneer Richard Sutton holds a similar view and recently launched Oak Labs to tackle that specific gap. DeepMind co-founder Shane Legg goes further, placing a "minimal AGI" as possible by 2028.

The 2028 timeline matters. If Legg is right, the window for building governance infrastructure is extremely short. Traditional regulatory processes take years. A FINRA-style body could potentially move faster because it operates outside the congressional calendar, but even voluntary industry coordination requires consensus that does not yet exist.

What are the real objections to this proposal?

Three critiques will likely emerge. First, who defines "frontier"? The exemption for smaller models depends on drawing a capability line that the industry has not figured out how to measure reliably. Current benchmarks are gamed almost as fast as they are published.

Second, FINRA's track record is mixed. The organization has faced criticism for being too cozy with the industry it regulates. An AI version might inherit that dynamic, especially if the same companies funding the agency are also its primary subjects.

Third, international coordination is hand-waved. Hassabis says other nations "would need to follow suit and find consensus," but offers no mechanism. China, the EU, and the US have fundamentally different approaches to AI governance. Consensus seems unlikely without a forcing function.

What happens if the framework is adopted?

Assuming voluntary adoption by Google, OpenAI, Anthropic, and Meta, the first phase would establish baseline evaluation protocols. Labs would submit frontier models for benchmarking before major releases. The agency would publish results, creating a transparency layer that currently does not exist.

The transition to mandatory compliance would require either congressional action or a creative interpretation of existing regulatory authority. Neither is fast. The more likely near-term outcome is a patchwork: voluntary US participation, EU AI Act compliance for European markets, and bilateral agreements with select allies.

Whether that patchwork is enough depends on what the next generation of models can actually do. And as Hassabis himself admits, nobody knows.

Frequently Asked Questions

What is Hassabis's proposed AI governance model?

Hassabis proposes a US standards body modeled after FINRA, the financial self-regulatory organization. It would develop evaluation protocols for frontier AI models, start with voluntary participation, and eventually become mandatory.

Would all AI companies face regulation under this framework?

No. The proposal specifically exempts non-frontier models from startups and academic research, aiming to avoid regulatory capture that would disadvantage smaller players.

When does DeepMind expect AGI to arrive?

DeepMind co-founder Shane Legg considers a minimal AGI possible by 2028. Hassabis describes AGI as likely just a few years away.

How does this proposal differ from EU AI regulations?

The EU AI Act focuses on risk categories and mandatory compliance from day one. Hassabis's proposal emphasizes industry self-regulation with a phased transition to mandatory benchmarks.

What is the main criticism of a FINRA-style AI regulator?

Critics argue that self-regulatory bodies can become too close to the industries they oversee, potentially favoring incumbents over genuine safety enforcement.

Source: The Decoder / Matthias Bastian