Trump and Tech CEOs Sign a "Morally Binding" AI Code of Conduct

President Trump and a group of the most powerful figures in artificial intelligence signed an AI code of conduct at the White House that the agreement itself describes as only "morally binding." The signatories include Meta CEO Mark Zuckerberg, OpenAI President Greg Brockman, Nvidia CEO Jensen Huang, and Elon Musk. No enforcement mechanism accompanies the pledge.

The language matters. "Morally binding" is a legal term of art, not a slogan. It signals that the signatories accept an ethical obligation to behave in certain ways, but that the obligation carries no contractual force, no regulatory penalty, and no court-enforceable duty. In practice, the code of conduct functions as a public promise. Companies can be shamed for breaking it. They cannot easily be sanctioned.

What the agreement actually contains

According to the report, the code of conduct rests on two structural elements.

First, outside auditors are supposed to check whether AI models work as intended. This is the only verification mechanism described in the agreement. External auditing would, in principle, give the public and the government some independent signal about whether the systems these companies build behave the way their makers claim. The report does not specify who those auditors are, how they would be selected, what access they would receive, or what standards they would apply.

Second, and critically, it is unclear what would happen if a company broke the rules. That gap defines the entire arrangement. A code of conduct with no defined consequences for violation is a norm-setting document, not a regulatory one. It establishes an expectation of behavior and depends entirely on reputational pressure and voluntary compliance to hold.

The contrast between the two elements is stark. Auditing implies scrutiny. Undefined penalties imply that scrutiny leads nowhere in particular. An auditor could find that a model does not work as intended, and the agreement as described provides no answer to the obvious follow-up question: then what?

A renaming buried in an executive order

Trump also signed an executive order renaming artificial intelligence "Super Intelligence." The renaming is symbolic, but symbolism at the level of an executive order is not nothing. It changes how the federal government's own documents may refer to the technology, and it signals how the administration wants the public to think about it.

The choice of the term is revealing in two directions at once. "Super Intelligence" borrows the vocabulary of AI safety research, where it traditionally refers to a hypothetical system that vastly exceeds human cognitive capabilities across all domains. Applying the label to today's AI systems — capable and commercially significant, but not superhuman in any general sense — inflates the technology's image. It also raises the stakes rhetorically: if the government calls the technology "Super Intelligence," the argument for binding oversight becomes harder to dismiss, even as the actual conduct code signed the same day declines to impose any.

The signatories and what their presence means

The list of executives at the signing reads like a map of the industry's power centers. Mark Zuckerberg leads Meta, which builds and open-weights its Llama family of models. Greg Brockman co-founded and serves as president of OpenAI, the company behind the GPT model line and the ChatGPT product. Jensen Huang runs Nvidia, the chipmaker whose hardware underpins the training of nearly every frontier model. Elon Musk owns and runs xAI, and simultaneously serves as a senior adviser to the Trump administration — a dual role that has drawn scrutiny since he entered government.

Their voluntary presence at the signing tells the reader something about the market calculus. Companies building frontier AI systems face intense public pressure to demonstrate responsibility. Signing a conduct code at the White House, alongside the president, produces exactly that demonstration at minimal cost, because the code's provisions bind only morally. For the administration, the signing delivers a visible governance achievement without requiring Congress, without a rulemaking process, and without a fight over enforcement powers.

That symmetry explains why voluntary frameworks of this kind recur in the history of technology governance. Industry gets goodwill. Government gets a headline. The public gets a promise whose violation has no defined price.

Why the stakes are real anyway

The context makes the weakness of the mechanism consequential. The companies represented at the signing are the same ones racing to build the most capable AI systems in existence. Their models are already deployed to hundreds of millions of users and embedded in business software, search products, and consumer applications. If those systems fail — if they work in ways their makers did not intend or did not disclose — the affected parties include essentially everyone who touches the internet.

Outside auditing, even imperfect auditing, addresses a genuine problem: today, the public has almost no independent verification of claims made about frontier models. Safety evaluations are largely run by the companies themselves, and results are released at their discretion. A regime in which external parties check whether models work as intended would, if implemented seriously, be a structural change in how the industry is held accountable.

The failure mode is equally clear. An audit regime with no consequence attached to a failed audit produces information without power. Companies would know what auditors find. The public might learn some of it. Nothing in the agreement, as described, would compel anyone to act on the findings.

The open questions

Several questions now determine whether this agreement matters beyond the signing ceremony.

Who audits the auditors? The report describes outside auditors checking models but leaves their identity, independence, and mandate unspecified. An auditor chosen and paid by the companies it audits inherits an obvious conflict of interest.

What counts as working "as intended"? Intent is declared by the model's maker. If a company's stated intent is narrow, the audit's scope is narrow with it.

And what does "morally binding" survive first contact with a competitive market? If one signatory concludes that a rival is cutting corners and the code imposes no penalty for corner-cutting, the rational pressure runs toward quiet defection, with the code serving as cover rather than constraint.

The renaming order raises its own question: whether a government that calls the technology "Super Intelligence" can long tolerate governing it with a handshake. A voluntary code, with audits but no teeth, is a first draft of AI oversight rather than the final one. Whether a binding version follows — through legislation, regulation, or a future executive order with actual enforcement power — will decide whether this signing was the beginning of accountability or a substitute for it.