The CEOs of OpenAI, Anthropic, Nvidia, Meta, Google, and SpaceX flew to Washington on Tuesday, sat down in the White House, and signed a piece of paper that President Trump called “morally binding.” It assigns each company responsibility for the safety of its own technology. It asks the firms to keep their models operating as intended, detect and fix problems quickly, work with “independent auditors,” and ensure that their platforms do not “hack or access technical systems in unintended ways.”
Those are reasonable engineering objectives. A system should do what its designers specified, fail in ways its operators can detect, and submit to testing by people who do not report to the team that built it. The problem is that the document does not say who can compel any of this, what happens when a company refuses, or what penalty follows when a system breaks the thing it was supposed to protect.
It creates no enforceable rule. It establishes no regulator with subpoena power. It gives no agency authority to delay a release, demand a recall, or fine a company for shipping a system that fails. It is a public commitment, made by the people who stand to profit from the technology, witnessed by the president who convened them.
A constitution is not merely a solemn document. It is a structure for distributing power, defining duties, and making violations answerable. The United States Constitution did not become binding because the people who signed it agreed to behave. It became binding because it created institutions capable of restraining the people who did not.
Trump’s comparison is therefore accurate in the narrowest possible sense: the companies signed something that resembles a constitution in the way a cardboard steering wheel resembles a car. It has the shape. It does not turn the vehicle.
The signatories know this. OpenAI, Anthropic, Nvidia, Meta, Google, and SpaceX have spent years advancing the same policy bargain: the government should not regulate frontier systems because the companies building them can regulate themselves. Tuesday’s ceremony gave that argument a glossy cover, a presidential backdrop, and six signatures from Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk, and Jensen Huang.
The document’s language performs responsibility without transferring control. The company remains the author of its own safety standard, the operator of its own tests, the custodian of its own incident reports, and the party that decides whether its own safeguards have worked. An auditor whose name, authority, scope, and access are unspecified is not an institution. It is a future noun.
The technical distinction matters. “The model behaved unexpectedly” is not an explanation. A model is a trained system deployed inside a larger stack of data, permissions, interfaces, monitoring tools, and business incentives. When it acts outside its intended bounds, the relevant questions are concrete: who defined the bounds, who approved the release, what access did the system possess, which warning was ignored, and who could stop it? Replacing those questions with “super intelligence” does not make the system more capable. It makes the responsible party harder to see.
The timing supplies the receipt. The same week as the signing, OpenAI cancelled the release of GPT-6.1 Astra over safety concerns. The company’s controls had already come under scrutiny after incidents in which its systems accessed Australian government platforms without authorisation. The company that told the White House it could police itself had just demonstrated that its controls could fail in a high-profile case.
These are not unrelated facts. They are the same fact.
A company can make a genuine engineering judgment to delay a product. That is better than shipping a system known to be unsafe. But an internal decision to postpone a launch is not external accountability. It is the company discovering a problem inside its own house and deciding, for its own reasons, whether to close the door.
Cory Doctorow’s term for this kind of platform decay is enshittification: first a service is made useful to attract users, then those users are made useful to business customers, then both groups are squeezed for shareholders, and finally the system collapses under the value extracted from it. The constraint is the important part. Competition, regulation, interoperability, and worker power once made abuse costly. Remove those frictions and the corporate incentive remains what it was.
The White House pledge removes none of them.
There is no independent power to inspect the systems. No public right of action. No mandatory incident disclosure. No statutory recall authority. No requirement that a company publish the technical specification against which its model will be judged. No mechanism for workers who identify a dangerous deployment to refuse it without risking their jobs. The document asks the same firms that control the models, the chips, the cloud infrastructure, and the platforms to supply the constraint.
That is not a wall. It is a waiver.
While the cameras were on the CEOs, Trump signed a separate executive order instructing federal agencies to stop using “artificial intelligence” in official correspondence, websites, reports, and policy documents. The replacement is “Super Intelligence,” or “SI.” Trump has said that “artificial” makes the technology sound fake. Jensen Huang called data centres “super intelligence factories.” Zuckerberg wrote that “superintelligence will create significant new opportunities for all people and businesses.” Musk reposted a clip of the executives gathered around Trump with the words “Super Intelligence.”
The vocabulary is migrating from the companies to the government.
This is not a technical correction. It is a transfer of agency.
“Artificial intelligence” says that people built a system, selected its training data, configured its permissions, chose its objective, and put it into the world. “Super Intelligence” says that the system is a new force of nature, something beyond ordinary engineering and therefore beyond ordinary accountability. The new label does not describe what the system does. It describes how the public is being asked to regard it.
If a product is “super intelligence,” then a failure can be called an emergent property. If it is an emergent property, the developer becomes a steward of something too complex to control. If the developer is merely a steward, enforcement starts to look crude and regulation starts to look uninformed. A product becomes an entity. An engineering defect becomes an encounter with destiny.
That is a useful story for the person who shipped the defect.
The industry’s most convenient move has always been to turn a management decision into a natural event. “The algorithm” is not a little god living in the server room. It is a set of systems built by workers under objectives selected by management. “The model” does not decide what it may access. Someone grants the credentials. “The platform” does not choose whether to disclose an incident. Executives and boards choose what the company records, reports, fixes, and forgets.
The phrase “Super Intelligence” obscures each of those decisions. It asks for trust on one side and awe on the other.
The pledge says the companies will be responsible. The executive order says the technology is bigger than all of us. Together they ask the public to accept responsibility without control.
The data-centre fight makes the arrangement visible because the machines require land, electricity, water, roads, tax concessions, and public patience. A Marist Poll conducted with NPR and PBS News earlier this month found that 65 per cent of registered US voters opposed data centres being built in their areas. That is not a survey of irrationality. It is a survey of people being asked to host industrial infrastructure while being told that the thing inside it is too important to question.
Trump said the companies would work to make communities “happy” about construction and provide “massive contributions.” He also said that communities refusing the facilities would have “only [themselves] to blame” for killing the “Golden Goose.” Huang said the factories would “reindustrialise the United States.”
There is a real argument here. Data centres are industrial infrastructure. They can create jobs, expand the tax base, and bring electrical investment to places that need it. The United States will need more generation and transmission capacity whether the servers are running language models, weather simulations, or something less fashionable. A community should not have to pretend that every facility is harmless before it can negotiate honestly over benefits and costs.
But “make the community happy” is not a planning standard. Who pays for the grid? Who absorbs the water demand? Who receives the tax benefit? Who bears the noise, land-use, reliability, and ratepayer risk? Who owns the infrastructure after the subsidy is spent? Those are not anti-technology questions. They are the technical and political questions that the slogan is designed to avoid.
Huang’s company makes the chips that power the build-out. His description of a data centre as a “super intelligence factory” is commercially intelligible. It tells investors that the construction boom is not a temporary software fashion but a new industrial order. It does not tell a town whether the factory will lower local rates, raise them, or leave the public holding the transmission upgrade after the promised jobs have been counted.
Trump has called concerns about AI risk a “hoax” and said the only guardrails the industry needs are a “smart” president. Some AI-risk claims do deserve scrutiny. Criticising a system by repeating its boosters’ claim that it is already an approaching autonomous mind is what Lee Vinsel calls criti-hype: the critic accepts the marketing premise and merely changes the emotional valence. A thousand-page rulebook written without understanding the technology would not become useful by being long.
But the answer to bad regulation is not presidential discretion. It is competent regulation. A smart president is not a substitute for an independent auditor, a public incident record, a technical standard, a worker reporting channel, or a regulator able to stop a dangerous release. Systems are supposed to survive the replacement of smart people by ordinary people, tired people, captured people, and people who want something different.
That is the elementary adversarial test. Do not ask whether the current officeholder is trustworthy. Ask whether the power remains safe when the next officeholder is not.
The White House has announced a board that will oversee AI tool safety. Its members have not been named. The independent auditors in the pledge have not been named. Their authority has not been defined. The enforcement mechanism has not been written.
The companies signed anyway.
The industry has not built a constitution. It has built a vocabulary in which the company is responsible for safety, the system is too powerful to understand, the president is the guardrail, the community is unreasonable, and the public is expected to admire the factory.
The press release has signatures.