The federal government recently ordered Anthropic to cut off access to Fable 5 — the most capable model it had ever released to the public — days after launch. The stated reason was national security. Anthropic disputes that the flagged technique is dangerous, noting rival models can do the same and that the government has shown almost none of its evidence. Set aside who is right. The world’s leading AI developer and the U.S. government are in a public standoff over whether a single product is too dangerous to sell — and we have no neutral way to settle it.
That is the uncomfortable truth beneath the AI debate. However fast or slow you think we should move, we can’t actually measure the speed at which we’re moving. We rely almost entirely on what developers — racing to justify billions in investment — choose to tell us about what they’re building. A few independent labs are chipping away, but we’re nowhere near the science serious oversight needs. We are flying blind.
Recommended Stories
This is why the Great American AI Act, the bipartisan bill from Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA), matters. It builds on a model for regulating complex, fast-moving technologies that my colleagues at Fathom and I developed a decade ago: the Independent Verification Organization, or IVO. The idea is already advancing in the states — through bills, study measures, and pilots from California to Connecticut — and now enters the national conversation.
Transparency laws in California, New York, and Illinois make developers publish a safety framework and disclose how they handle catastrophic risks. That’s progress, but it leaves them grading their own homework. Independent verification is the next step. It is not an audit of whether a lab lived up to its word. It is an independent crash test of the products themselves. Crucially, what counts as “acceptable” risk is set not by the company but by the CAISI, working with the White House science office, national security agencies, and outside experts.
A voluntary version is already forming, with states building markets companies can opt into for everyday but serious risks — unreliable enterprise agents, or chatbots that might steer a teenager toward self-harm. But catastrophic risks are too consequential to leave optional. The Obernolte-Trahan bill makes verification mandatory for the most powerful models and puts the government in charge of the standard. And it has teeth: IVOs monitor risk continuously and report to CAISI at least twice a year, and if a verifier finds a product too dangerous to release, a state or federal attorney general can ask a court to block deployment.
RURAL AMERICA WAS TOLD ‘LEARN TO CODE’ — NOW IT MUST POWER AI
Why license private verifiers — why not just fund CAISI to run the tests? Because the tests don’t exist yet. Our most powerful models remain mysterious. We can red-team them and learn a great deal, but still can’t say when we’ve tested enough to call a system safe. The science has to be built fast, and rebuilt as capabilities cross new thresholds. The strongest independent work comes from a few small outfits such as METR, Apollo Research, and Transluce that survive on whatever access they can negotiate. The IVO model turns it into a genuine sector, pulling investment and talent toward the measurement science we lack. Tens of billions flow into building artificial intelligence; a rounding error goes into overseeing it. This begins to change the ratio.
The cliche is that we’re building the plane while flying it. It’s true. The Obernolte-Trahan bill gives government a window into frontier AI now — and builds the science we’ll need to use it well. That it asks both parties to set aside their reflexes — the Right’s distrust of government, the Left’s distrust of markets — is what makes it fit for the moment. Washington should take it up.
Gillian Hadfield is the Bloomberg Distinguished Professor of AI Alignment and Governance at Johns Hopkins University.
