Trump's 'morally binding' AI pact lets the labs choose their own auditors

The bill and the accord put obligations in different places. Warner's bill sets federal deadlines on the developer: a Commerce Department safety board gets access to new models at least 45 days before release, and it can enforce technical standards and security controls ^[3]. Incidents must be reported within 30 days, or within 72 hours when an incident poses an imminent threat to national security ^[4]. The accord's commitments sit inside each company: controls that monitor capabilities and alignment, an internal team to confirm the controls work, an independent external auditor or evaluator, and an independent board committee to oversee reports and remediation ^[10]. As implicator.ai reported them, none of the four gives a government body access to a model before release or sets a deadline for reporting an incident to one ^[1].

The board-deck version is that Anthropic, Google, Meta, Nvidia, OpenAI and SpaceXAI have accepted outside evaluation and board-level oversight under a document the President called "morally binding" ^[9]. That version leaves out enforcement. "There's a belief that there should be tremendous self-regulation, and we automatically have regulation with the Department of Justice, the FBI, all of that," Trump said on Tuesday ^[12]. "But the self-regulation is very important," he said ^[21]. Asked why the executives should be trusted to regulate themselves, he replied: "Because they're outstanding people" ^[13].

Cruz made the skeptic's case against the bill. He chairs the Senate Commerce Committee and said the proposal would give the safety board and the Commerce Department too much authority. He said he preferred a role for "independent judges" ^[6]. He said "federal standards on advanced AI must also address the burgeoning patchwork of state laws" ^[7]. "Congress must not legislate on the issue of artificial intelligence hastily or in a closed manner," he said ^[20]. The trade-off is speed against design. Warner, joined by Brian Schatz and Andy Kim, sought passage by unanimous consent, a procedure that one senator's objection stops ^[19][1]. The catastrophic-risk bill Cruz is writing with Majority Leader John Thune and Amy Klobuchar is still a work in progress, and its terms are not known ^[8][15].

Mandatory testing before release has Republican support too. Sen. Josh Hawley wrote on Tuesday that models should undergo mandatory testing before deployment, and said he plans legislation making AI firms liable for reckless design ^[14]. "I've not talked to a single member on either side of the aisle that doesn't think we need to move," Warner said ^[5]. I think that narrows the fight after the election to two questions Cruz raised himself: whether an agency or the courts enforce the rules, and whether federal rules override state laws ^[6][7].

The calendar decides this quarter. House members are not expected back until after the Nov. 3 midterms, and senators could leave as soon as this week ^[11]. Speaker Mike Johnson did not commit to legislation ^[16]. Until Congress acts, review of capability reports at the six signatories sits with the board committees and outside evaluators the accord asks them to appoint ^[10].

Next quarter depends on whether anything goes wrong in the meantime. "If Congress goes out, God forbid we have an incident, AI is already suspect by most Americans, we could see the stop of innovation if an AI tool took over a hospital or water system," Warner said ^[18]. In Verasight's September national survey, 63% of respondents said AI should slow down and 5% said it should accelerate, a ratio of more than 12 to 1 ^[17][2].