White House AI 'Superintelligence' Accord Criticized as Toothless Self-Regulation

Full analysis

Gary Marcus took a swing at a White House "superintelligence" accord that Trump praised as a "tremendous" exercise in "self-policing." Marcus reads it as three lines: we agree not to be regulated, we agree not to give the public a voice, trust us. He names Dario Amodei, Sam Altman, and Elon Musk as the ones who "chickened out" of backing anything with teeth. For anyone buying or building on frontier models, the takeaway is plain: binding US federal AI rules are not coming in the next year, and the labs like it that way.

This is easy to undo in theory and hard to undo in practice. An accord is not a law. A future administration or Congress could replace it overnight. But the thing being decided is not really "should there be an accord." It is who audits frontier labs, and the accord answers that with subcontractors the labs pick themselves. Once that becomes the norm, it sets the reference point for every procurement conversation for years. No deadline forces anyone's hand here, which is exactly the point: the absence of a decision gate is the decision.

The Skeptic

Marcus is right and also late to his own party. Labs resisting binding rules is not news, it is the default whenever a young industry grabs its regulatory moment before regulators can spell "eval." The accord buys the labs 18 to 24 months of clear road while the EU AI Act generates the first real enforcement data. Fine. But here is the harder question Marcus skips: what would binding oversight even measure? Nobody has a credible test for "superintelligence risk at scale." You cannot audit a threshold nobody can define. Regulation without a measurement is theater too, just with a government badge on it. The accord is toothless. So is the alternative on offer.

The Safety Lens

Buyers should worry about the structural failure, not the politics. In aviation, pharma, and nuclear, the one institutional move that made safety real was separating the thing being tested from the people doing the testing. This accord reverses that. The "independent" reviewers are subcontractors chosen by and paid by the labs being reviewed. That is self-certification with a nicer letterhead. The quieter loss is the "Pacing" framework, discussed weeks earlier and now gone. Pacing was the only proposed mechanism that could have created a checkpoint before a model ships, rather than an inquest after something breaks. Kill the checkpoint and you are left trusting that internal safety teams will overrule their own launch calendars. They rarely do.

The Enterprise Buyer

Here is what a chief AI officer actually loses. A "trust us" accord gives procurement nothing to put in front of a legal and risk committee. When you evaluate a vendor's model, you want an audit trail from someone who does not report to the vendor. This accord produces the opposite: a legitimacy shield the labs can wave around while handing you no independent artifact. So the practical move does not change. Anchor your compliance program to the EU AI Act, which is real and dated, and to your own contract terms: indemnification, data handling, audit rights you negotiate directly. Do not treat federal silence as cover. State attorneys general, FTC Section 5 actions, and ordinary tort liability all remain live, and none of them care that Washington threw a party.

The Compute Pragmatist

No binding rules means no mandatory reporting on training runs, no compute threshold that triggers a review, no external checkpoint on scaling timelines. Regulatory friction was the one plausible outside brake on how fast the next generation of models gets trained. It is gone. Anthropic, OpenAI, and xAI all have massive cluster builds committed or underway, and the accord tells them to floor it. For buyers that means the capability curve keeps steepening and the GPU market stays tight, which keeps NVIDIA's pricing power intact through 2026. If you were hoping a compute cap might slow the arms race and cool chip prices, that hope just left the building.

The Researcher

The "independent review" language is the giveaway. Independence means the reviewer does not answer to the reviewed. Subcontractors picked by the labs fail that test on day one. The academic and independent research community, the people who actually publish adversarial red-team work, get no formal role. Marcus is right that Pacing's disappearance is the bigger story than the accord's weak wording, because Pacing was a mechanism and the accord is a mood. What this produces is a legitimacy shield, not a safety architecture. When self-assessment replaces outside review, you lose the one thing that makes any technical claim trustworthy: someone with no stake in the answer checking the work.

Where they disagree

The real split is between the Safety Lens and the Skeptic, and it is not about whether the accord is weak. Everyone agrees it is. It is about whether stronger rules were even possible. The Safety Lens says the mechanism existed, Pacing was right there, and the labs killed it. The Skeptic says Pacing was a gate with no gauge, because nobody can measure the risk it was supposed to hold back. If the Skeptic is right, the outrage is misplaced and the EU experiment will teach us more than any US mandate could. If the Safety Lens is right, a real brake was within reach and got quietly removed.

The second tension is the Enterprise Buyer against the Builder's status-quo comfort. Federal inaction reads as permission to defer governance spend nobody wanted to make anyway. The Buyer's answer is that liability did not go anywhere. The absence of a federal rule is not the absence of exposure.

What it hinges on

One belief: will US federal AI policy stay permissive through the next year? Every persona leans yes. The accord, the vanished Pacing framework, and three named CEOs declining to back accountability all point the same direction. The labs won this round, and nothing on the calendar reverses it soon. The move for buyers is to build governance off the EU AI Act's real deadlines and off contract terms you negotiate yourself, because that is the only oversight anyone can currently show a risk committee.

Prediction: No US federal binding rule imposing independent third-party safety audits or a mandatory pre-deployment review on frontier AI labs (OpenAI, Anthropic, xAI, Google DeepMind) will be enacted by 2027-04-04.

Confidence: High. The accord, the killed Pacing framework, and three CEOs declining teeth all point one way.

Why: The White House accord replaces oversight with labs picking their own reviewers, and the one mechanism that could have created a real pre-deployment checkpoint, the Pacing framework, quietly vanished from the conversation weeks after it was floated. Dario Amodei, Sam Altman, and Elon Musk all declined to back stronger accountability, which removes the only inside pressure that could have pushed a binding rule. Federal legislation needs either sustained industry support or a triggering crisis, and this story shows the industry aligned against it with the administration cheering the self-policing framing. The opposite outcome, a binding federal audit rule in six months, would require Congress to move fast on a technical mandate nobody can yet define a test for, which is the least likely path given every signal here.

Revisit by 2027-04-04: We're right if no US federal statute or binding executive rule mandating independent (non-lab-selected) third-party safety audits or pre-deployment review of frontier models is in force. We're wrong if any such federal rule is enacted and binding by that date.

The Skeptic's point cuts in the labs' favor here. You cannot pass a binding audit rule when nobody, including the critics, can name what the audit measures. That gap is the labs' best protection, and they know it.