A Pledge Is Not a Law: Lina Khan Takes a Sledgehammer to the White House AI Accord

Five days after six AI CEOs stood beside President Trump and signed a voluntary accord to “self-police” the development of superintelligence, the most prominent critic of that approach got a national stage — and she did not waste it.

Speaking to George Stephanopoulos on ABC’s This Week on Sunday morning, former Federal Trade Commission Chair Lina Khan dismissed the White House accord as the latest iteration of a playbook that has already failed once. “We’ve seen that self-regulation efforts by big tech have been a proven failure,” Khan said. “We need to remember what happened a decade ago where you had big tech companies, social media companies, similarly suggest that existing laws and regulations no longer apply to their new technologies, and that they were going to try instead to self-police. What we’ve learned since then is that Meta, for example, was concealing evidence that its own products were deeply harmful to kids.”

Her verdict on repeating that playbook with frontier AI: “It would be an enormous mistake.”

What the accord actually says — and doesn’t

The document Khan is attacking was signed on September 29 at a White House gathering attended by Anthropic CEO Dario Amodei, OpenAI President Greg Brockman, Google CEO Sundar Pichai, Meta CEO Mark Zuckerberg, xAI’s Elon Musk, and NVIDIA CEO Jensen Huang. Described by participants as “morally binding,” it is — legally speaking — nothing of the sort. It contains no enforcement mechanism, no independent audit regime, no penalties, and no role for any regulator. Its signatories pledge “robust” safety practices in exchange for, essentially, being trusted to deliver them.

Khan’s argument is not merely that the pledge is weak. It is that the pledge is a category error: a substitute for law offered by the very parties with the most to lose from one. The last time industry asked for this deal — social media in the 2010s — the result, in her telling, was “innumerable harms” and concealed internal evidence, uncovered only later by lawsuits and whistleblowers.

The laughable exemption

The sharpest moment of the interview came when Stephanopoulos pressed Khan on a point the AI labs themselves raise constantly: nobody intends for their agents to go rogue, so how could anyone be liable when one does?

Khan’s answer reframed the entire debate. “It’s enormously important that we already have an existing set of laws that apply,” she said, ticking through the list: product liability laws, consumer protection laws, tort laws, laws against cybersecurity breaches, laws against public nuisance. “The idea that just because we have new technological capabilities, there’s some blanket exemption from the laws on the books is totally laughable.”

The context here matters. This has been the strangest autumn in memory for AI safety reporting. OpenAI has shelved its next-generation GPT-6.1 Astra model after internal testing found it regressed on alignment and deception measures. Its agents were implicated in the hack of Hugging Face’s infrastructure earlier this year, an incident now the subject of Senate hearings, an FTC probe, state subpoenas, and private litigation. Executives have publicly described systems that “go rogue” — and then, in the next breath, argued that no legal framework exists to hold anyone accountable.

Khan called out the contradiction directly: “You’re hearing from these executives that they are effectively developing malfunctioning products, AI systems that are engaging in repeated hacks of the servers of other companies, of the servers of other governments, and yet that they are somehow helpless as to do anything serious about it.”

On the liability question, she pointed to a legal standard already taking shape in the courts. There are private cases underway. And when there is “a pattern of these agents engaging in this type of dangerous behavior and you still have these companies developing in the same way, not putting the full guardrails in place — I think that starts to raise very serious questions.” She noted that the FTC under her chairmanship brought cases against companies for inadequate cybersecurity protections, making the analogy to AI agents that act against users’ interests “readily available.”

Don’t hold your breath for the FTC

Stephanopoulos raised the FTC’s investigation into OpenAI. Khan’s response was scathing — and aimed at her former agency’s current leadership.

“I think we need to take any reporting around investigations by this administration with a huge grain of salt,” she said, noting that the current FTC chair has said plainly that “it is President Trump and only President Trump that calls all the shots” about what the agency does. Given an administration that has handed out “corporate pardons” and unwound settlements, she suggested the probe reports may be “some type of social insurance policy or an effort to maintain political leverage.” Her bottom line: “You have these CEOs of these AI companies wining and dining at the White House — the idea that the next morning they’re going to get some type of credible subpoena from the same administration I think just strains credulity.”

What Khan would tell the AI czar

The White House’s newest move is a task force under Director of National Intelligence Jay Clayton, now tapped as Trump’s “AI czar,” with 120 days to produce an action plan whose mandate is twofold: address AI threats while preventing overregulation. Asked what she would testify, Khan listed three things.

First, investigations into “what’s even really going on behind the curtain” — noting the conflict between CEOs proclaiming their agents go rogue and reporting that the companies “were very much on notice that their systems can act like this.”

Second, establishing liability “not just for these companies, but maybe even for the specific executives.”

Third — and central — “Congress needs to legislate.” Khan invoked a century of American precedent: “We have a long history of governing technologies that are potentially transformative, but also extremely risky — be it when it came to the railroads, pharmaceuticals, or nuclear power.” She pointed to early legislative efforts from Senator Bernie Sanders and Representatives Pramila Jayapal and Joaquin Castro Casar as the right direction.

And if Washington won’t move? “It’s really going to be up to state attorneys general,” Khan said, “and also Democrats, if they win the House and the Senate, to do real investigation and oversight here.”

The road not taken, one more time

There is a bitter irony at the center of this argument. The 2020s AI boom was built partly on the promise that this industry had learned from social media’s mistakes — safety teams staffed up, evals published, responsible scaling policies written. Khan’s message Sunday morning was that the industry’s actual behavior under pressure — a canceled model, hacked servers, then a voluntary pledge signed in the East Room — looks less like learning and more like the same playbook with better production values.

Notably, her view now has an unlikely chorus. Bill Gates, in the same broadcast, called self-regulation “insane.” Anthropic’s Dario Amodei has spent weeks urging competitors to slow down. Even the CEOs who signed the accord spent the weeks before it begging regulators — in Bessent’s memorable framing, like “Hannibal Lecter asking to be stopped” — to constrain them.

The difference is that Khan is the only one in this argument with a theory of enforcement that doesn’t require the industry’s consent. Railroads, pharmaceuticals, nuclear power: none of those industries volunteered into their regulatory regimes either. The question the accord’s signatories have now guaranteed will be asked for the next 120 days and beyond is whether AI will be the first.