The Proposal That Changes Everything
On Tuesday morning, Demis Hassabis posted something unusual. The DeepMind CEO wasn’t announcing a new model or a research breakthrough. He was asking the US government to regulate him.
In a detailed post on X titled “A Framework for Frontier AI and the Dawning of a New Age,” Hassabis proposed creating an independent standards body modeled after the Financial Industry Regulatory Authority (FINRA). This body would test frontier AI models before release and develop best practices for their deployment. And he wants it operational before the end of the year.
Think about how rare this is. A CEO of one of the world’s most advanced AI labs voluntarily calling for oversight. That should tell you something about where we are with this technology.
What Hassabis Is Actually Proposing
The framework has three key elements. First, frontier labs would voluntarily share their models with the Standards Body up to 30 days before release for review. Second, once the assessment protocol is proven effective, it would become mandatory — frontier models would need to pass before being deployed in the US market. And third, labs would work with the body to address any critical vulnerabilities discovered after release.
Hassabis envisions this body as technically focused, staffed by open source representatives and industry experts, and funded by the AI labs themselves — but operating independently under US government backing. Think of it like FINRA, which oversees broker-dealers in the securities industry: a self-regulatory organization (SRO) with government backing but industry funding and operational independence.
Why Now? The Context Matters
This proposal didn’t come out of nowhere. Over the past few months, the US government conducted ad hoc reviews of two major model releases: Anthropic’s Mythos and OpenAI’s Sol. Both drew significant criticism for lack of technical expertise and opaque decision-making. The government reviewers simply didn’t have the deep technical knowledge needed to evaluate these systems properly.
The current system isn’t working. Government agencies lack the AI talent to meaningfully evaluate frontier models. The labs themselves have no consistent standards to meet. And the public gets no transparent, technically-grounded assurance that these systems are safe to deploy.
Hassabis’s answer is to create a dedicated organization that can hire that expertise, can develop proper testing protocols, and can make independent decisions about when a model is ready for release. It’s a structure designed to keep pace with the field’s acceleration, as he puts it — “to be ratcheted up if the seriousness of the situation demands.”
If you’ve been following the rapid-fire model releases this year — from OpenAI’s Sol deleting user files autonomously to SpaceXAI’s Grok Build uploading entire codebases to cloud storage — you can see why a more structured approach might be necessary.
The FINRA Model: A Clever Workaround
The choice of FINRA as a model is strategic. White House AI advisor Sriram Krishnan recently stated flatly that “there will not be an FDA for AI” within the executive branch. A traditional government regulator was off the table.
An SRO like FINRA sidesteps that political reality. It’s not a government agency — it’s an industry-funded, independent organization that operates under government oversight. This structure has been used for decades in finance and securities regulation. Applying it to AI is novel, but the legal and operational framework already exists.
Hassabis even suggests the body could outsource some evaluations to the growing ecosystem of AI safety groups, letting them specialize in specific risks. That’s practical thinking — no single organization can cover every dimension of AI safety, from cybersecurity vulnerabilities to bias testing to long-term existential risk.
Is This Genuine or Strategic?
This is where I get a bit skeptical — and I think a healthy dose of skepticism is warranted.
DeepMind is the AI division of Google, which is currently facing its own training data lawsuits and regulatory scrutiny across multiple jurisdictions. A FINRA-style body that’s industry-funded and technically-focused sounds a lot more comfortable than the alternatives: heavy-handed government regulation, state-by-state patchworks of AI laws, or the kind of public backlash that leads to outright bans.
There’s also the first-mover advantage. If DeepMind helps design the standards body, they help shape the rules. That’s not necessarily bad — the people closest to the technology should absolutely have a seat at the table. But it’s worth remembering that self-regulation has a mixed track record across industries — as I wrote recently about the cost of handing your data to AI providers, the incentives matter. The 2008 financial crisis happened under a regulatory system that included SROs.
At the same time, I don’t think this is purely cynical. Hassabis has been talking about AI safety since long before it was fashionable — since founding DeepMind in 2010. His background in neuroscience and his public statements over the years suggest genuine concern about where unchecked AI development could lead. The proposal reads like someone who genuinely believes the industry needs guardrails and is trying to build them before something goes badly wrong.
I think the truth is somewhere in the middle. It’s both strategic and sincere. And honestly, that’s probably the best combination for something like this to actually work.
What This Means for Developers and Companies
If this proposal gains traction — and with the backing of Google DeepMind, it has a real shot — here’s what changes:
- Frontier model releases will slow down. A 30-day review window means no more surprise model drops. Planning and release cycles will need to account for this.
- AI safety becomes a competitive advantage. Labs that have invested in safety infrastructure, testing protocols, and transparency will find it easier to pass review. This could accelerate the quiet shift toward open-source and self-hosted models that many enterprises are already making — if you can control the model, you can audit it yourself.
- Smaller AI companies may struggle. The cost of compliance with a formal standards body could be significant. This might accelerate consolidation in the AI industry, which has its own set of concerns.
- The standards themselves become a battleground. What counts as “safe enough” will be debated intensely. The open source community will want different standards than closed-source labs. This is healthy — but it will be messy.
The Bigger Picture
This proposal arrives at a moment when the AI industry is wrestling with multiple crises of trust. OpenAI’s Sol is deleting user files. SpaceXAI’s Grok Build was uploading customer code to cloud storage without permission. Meta is being sued for allegedly using AI to target workers on leave with layoffs. The list goes on.
Regulation isn’t just about preventing worst-case scenarios. It’s about building enough trust that this technology can reach its full potential. Companies won’t bet their core operations on AI systems they can’t trust. Governments won’t deploy AI in critical infrastructure without guarantees. Regular users won’t adopt tools that might delete their data.
In my own work managing ICT for a government agency, I’ve seen both sides of this. The potential of AI for improving public services is enormous — automating routine tasks, analyzing data for better policy decisions, detecting fraud. But every time I read about another AI incident — data leakage, unexpected behavior, security vulnerabilities — it reminds me that we’re not ready to just turn these systems loose without guardrails.
That tension — between the incredible potential and the very real risks — is exactly what Hassabis’s proposal tries to address. Whether it succeeds or not, the conversation itself matters. The fact that one of the most powerful people in AI is voluntarily calling for oversight is a signal that the industry knows it can’t self-regulate forever.
The question now is whether the US government — or anyone else — is ready to take him up on it.