The AI regulation draft is markedly different than what Anthropic CEO Dario Amodei advocated for this summer

Anthropic's leaked draft frontier AI legislation is markedly different than the Federal Aviation Administration model that Anthropic CEO Dario Amodei proposed this summer.
Under the Anthropic framework, the labs get to write the standards that they must then comply with. Powerful groups in the AI safety camp are suggesting the government should do so instead.
Anthropic may be trying to curry favor and appear constructive on AI policy with the Trump administration -- after being on the outs for so long.

This is Benjamin Guggenheim. Welcome to WP Intelligence's AI & Tech Brief, where we examine the transformative technology of artificial intelligence at the intersection of innovation, policy and power. Get in touch at [email protected] or on Signal at ben_guggenheim.88.

The Lead Brief

Anthropic's leaked draft legislation deviates in surprising ways from what the company has said in the past regarding frontier regulation.

It's also surprising that the AI company is putting its name to and circulating its own legislation in the first place -- the contents of which have opened the floodgates to criticism from both the AI safety camp and the anti-regulation camp close to the Trump administration.

"I've never seen someone say 'Hey, this is the Google-branded bill or the Meta-branded bill.' They're subtle," said one person who works in AI policy in Washington, who spoke on the condition of anonymity to speak candidly. "They're hopelessly naive on this. You're [putting out] an invitation to just get roasted from the left and the right."

What's in the draft legislation: The bill, reported by Politico on Friday, calls on large AI developers to write and comply with their own safety framework describing how the respective companies deal with risks and respond to critical safety incidents.

Here are four key takeaways from the text, which differs in some notable ways from the most prominent proposals on frontier AI put forward so far by Reps Lori Trahan (D-Massachusetts) and Jay Obernolte (R-California), as well as by the top Democrat on the Senate Commerce Committee, Maria Cantwell (D-Washington):

The legislation is no "Federal Aviation Administration" for AI.

Anthropic CEO Dario Amodei wrote in a June essay that model AI regulation should be similar to how the FAA sets out minimum standards for plane designs.

An FAA-type model reflects a very hands-on approach by the government: The agency not only certifies aircraft model designs but also the factories that build the planes -- and it has significant authorities to revoke certification that planes are airworthy.

But in Anthropic's leaked proposal, the frontier developers would write their own safety plan to determine whether models pose unacceptable risks.

The plan is legally binding (the AI company is supposed to submit to the Commerce Department an annual statement on how it is complying with its own safety framework), but the framework nonetheless gives significant leeway for the labs themselves to self-regulate.

Contrast this with Cantwell's plan, which would have the National Institute of Standards and Technology set standards against which the labs are evaluated. Under Obernolte and Trahan's Frontier Act, labs will write their own plans in accordance with minimum standards outlined by the Commerce Department.

"It feels incongruous with Anthropic's otherwise-strong public statements on the need for robust regulation to promote a 'race to the top,'" said Samuel Hammond, the director of AI policy at the Foundation for American Innovation.

"I have no inside knowledge, but it feels like an exercise of having an attorney write up an older framework when the conversation -- and Overton window -- has already moved on," Hammond said.

The auditing regime: Unlike the Frontier Act -- which would have independent verification organizations (IVOs) for very large AI developers licensed under the Commerce Department -- the Anthropic bill would have auditors registered and rated by Commerce (and AI companies would have to have their guardrails checked by these independent auditors).

In other words, under Anthropic's proposed system, the government would have less direct oversight and power over the evaluators, which could arguably be a good or bad thing.

If the auditors are too dependent on AI companies for their compensation, for example, critics have warned that companies might want to "audit-shop" and pick auditors that will give them more favorable assessments.

The Anthropic proposal potentially sweeps in a larger selection of AI companies.

Anthropic's draft bill uses a lower threshold to define a frontier model that would be subjected to these regulations.

The draft text says that models that were trained using more than 10 septillion calculations (10 to the power of 25) qualify; the Frontier Act and the California AI transparency and safety AI law, by contrast, use a threshold of 100 septillion (10 to the power of 26).

The difference may play into the arguments of the anti-regulation crowd, such as former White House AI czar and co-chair of the President's Council of Advisors on Science and Technology David Sacks, who has argued that Anthropic wants to create a "regulatory moat" around its business that sinks smaller competitors with compliance costs.

"As written, however, it will eventually sweep in most of Silicon Valley, damaging the start-up ecosystem," said Sacks in a social media post responding to the framework, and noting that the definition includes calculations from fine-tuning and modifying models.

"In practice, that captures almost all meaningful post-training -- work that is becoming commonplace in the enterprise."

It's important to note that several of the draft bill's provisions, such as those pertaining to the safety framework, only apply to large labs with revenue greater than $500 million or expenditures in excess of $1 billion (though reporting requirements also apply to smaller frontier developers).

According to AI research firm Epoch AI, several models released in 2024, including those in Google Gemini and Meta Llama families, were trained with between 10 and 50 septillion operations, which falls below California's threshold but above Anthropic's threshold. (The Washington Post has a content partnership with OpenAI.)

By contract, Epoch estimates that OpenAI's recent model, GPT-6 Astra, was trained using 940 septillion operations.

Under the new bill proposed by Anthropic, the Commerce Department needs to clear a high bar in order to force a pause on frontier AI development.

Under the Anthropic bill draft, the Commerce Department must first go to a court and establish by "preponderance of evidence" that the model poses an imminent risk of death or serious injury to more than 50 people or more than $1 billion in property damage.

"[The U.S. government] should retain more emergency optionality without having to make an unreasonably detailed affirmative case that predicts whether 50 people will die or 49," said Sneha Revanur, founder and president of AI safety advocacy organization Encode Action, in a social media post.

"It's nakedly self-interested for Anthropic to shop around provisions that serve to protect its own wiggle room."

On the other hand, it's understandable that Anthropic might want protections from a capricious shutdown of its model after the Trump administration used a legally dubious export-control directive to shut down Anthropic's Fable model earlier this year.

Emil Michael, a senior Department of Defense official and a vice chair of Trump's Super Intelligence task force, told the Information Monday that the administration reserves export-control authority to get AI labs to retract their models.

Food for thought: The proposal is similar to legislation that was under negotiation between Sen. Amy Klobuchar (D-Minnesota) and Senate Majority Leader John Thune (R-South Dakota) under the auspices of Senate Commerce Chair Ted Cruz (R-Texas).

Cruz is a China hawk who, along with Sacks, has previously pushed a preemption on state AI laws and favored an extremely light-touch approach to federal AI regulation.

"They're so desperate to be seen as constructive," the person who works on AI policy said of Anthropic's federal affairs team.

The preemption of state laws is significant enough to draw the ire of AI safety advocates.

Currently, Anthropic's draft text would preempt laws governing public disclosure, incident reporting and third-party evaluation. It would not touch other laws governing children's safety and civil rights.

As written, however, it's unclear whether the bill would also block states' abilities to shut down models with emergency authorities.

"Some of the preemption provisions are remarkably far-reaching and consequential," said Nathan Calvin, general counsel at Encode, in a social media post. "Most outrageously, it has a provision that looks like it intended to stop states from enforcing laws or judicial decrees that 'suspend the development, deployment or internal use of an AI model.'"

As I've noted previously, the trade-off in any upcoming AI legislation will be a national standard for some extent of preemption requested by the "accelerationist" camp.

I've heard from several Capitol Hill staffers, policy advocates and lobbyists to expect another stab at preemption from that camp in the lame-duck session.

"If preemption is included in federal legislation, the trade should be to achieve a nationwide policy that's generally stronger and broader than what can be implemented through individual state legislative provisions," said Anthony Cimino, the head of federal affairs at Anthropic, in a social media post responding to the criticisms.

"We've emphasized that any preemption in the AI safety context should be narrow, even scalpel-like."

This newsletter is published by WP Intelligence, The Washington Post's subscription service for professionals that provides business, policy and thought leaders with actionable insights. WP Intelligence operates independently from The Washington Post newsroom. Learn more about WP Intelligence.

AI & Tech Brief: Anthropic's leaked bill