Trump’s AI Safety ‘Accord’ Is a Fancy Pinky-Swear

Picture a Tuesday afternoon in Washington, D.C., where the leaders of the world’s most powerful artificial intelligence companies are gathered around a White House luncheon table with President Donald Trump. On the surface, it looks like a moment of unity: Google, Anthropic, Meta, OpenAI, xAI, and Nvidia all signed a document called “The White House Accord on Super Intelligence.” The occasion, presented with all the ceremony of a major policy breakthrough, was meant to show that the people building our most advanced technologies are willing to police themselves. President Trump praised the effort as an act of “tremendous self-regulation,” and the room likely felt charged with optimism. But if you read the fine print, this accord is not the sweeping, enforceable rulebook that many safety advocates, and even some AI company insiders, have been asking for. It is a collection of promises, not laws. It cannot force anyone to do anything, and it does not create a new government agency with power to inspect models before they reach the public. Instead, it asks companies to voluntarily commit to a set of good practices, in the hope that they will take their responsibilities seriously. The scene is a snapshot of where we are with AI policy right now: big names, big handshakes, and a lot of hope that private corporate commitments will be enough to keep us safe.

So what exactly did the companies agree to? The accord’s language is gentle and advisory, full of phrases like “should implement” rather than “must do.” For starters, it says companies should establish “robust internal controls” to monitor their models, which means they need to understand not just what their AI can do, but also what it might do when pushed in unexpected ways. In plain language, they should have systems in place to notice if a model is becoming too good at hacking, manipulating, or causing harm. Second, the accord tells companies to “empower” an internal team to do this monitoring work and to fix problems when they arise. That sounds straightforward, but it is actually a big ask: it means giving real authority and resources to people whose job is to say no, slow down, or delay a product launch if something looks dangerous. Third, the companies are encouraged to bring in an external monitor, someone independent who can assess whether all of this internal work is actually happening. Finally, the accord asks each company’s board of directors to form a committee that receives reports about safety efforts. In theory, this creates a chain of accountability from the engineers working on the models all the way up to the people who govern the company. But because none of this is enforceable, the entire system depends on trust: trust that companies will hire the right people, trust that those people will have genuine authority, and trust that boards will actually pay attention.

The White House Accord is a mix of progress and disappointment. On one hand, it shows that the frontier AI labs are finally trying to figure out how to make products that are safe for the general public. It is meaningful that executives met at the highest levels of government to talk about AI danger, rather than pretending the issue doesn’t exist. On the other hand, the accord falls far short of what many people inside these companies have been asking for. For years, researchers and safety engineers have called for mandatory regulations: pre-market testing, licensing requirements, legal liability for harm, and independent audits with teeth. The accord offers none of that. It is also worth noting that this is a voluntary agreement, not a contract. The main legal risk for companies is not that they will be fined for violating the accord, but that they could be accused of deceptive marketing if they publicly promise safety practices they don’t actually follow. Under consumer protection law, making a bold claim about your product and then failing to deliver can count as lying to customers. The Federal Trade Commission, which protects consumers from unfair or deceptive business practices, reportedly intends to investigate several AI companies over potential consumer protection issues. That investigation could lead to fines or forced changes, but it is not the same as a comprehensive regulatory framework designed to prevent the next catastrophic AI failure. So while the accord is a nice symbol, it is not a safety net.

To understand why this moment matters, you have to look at what has been happening behind the scenes. For weeks before the accord was announced, there was intense debate about how to regulate AI. Some of the most powerful AI companies were reportedly seeking an antitrust exemption, which would allow them to coordinate on safety research without being accused of colluding to restrict competition. On its face, that request seems reasonable: if companies can share information about dangerous model behavior, they can all build safer systems. But antitrust experts and practitioners argue that the exemption is actually unnecessary, and perhaps even misleading. Their reasoning is both clever and darkly funny: if AI ends humanity, as some researchers have warned could happen if we don’t act, then that would be the ultimate anticompetitive outcome. Why? Because humans are an essential part of every marketplace. If there are no humans, there is no market, no customers, no workers, and no businesses. So, in a strange way, preventing AI catastrophe is already consistent with preserving competition. The fact that companies are asking for special legal protection to coordinate on safety suggests that they do not trust the current rules to allow them to be cautious together. But it also highlights a deeper tension: private companies are being asked to protect the public from risks that they themselves are creating, and they want to do so on their own terms, without the inconvenience of tight government oversight.

This is not the first time that these companies have put their names to a voluntary safety agreement. In early 2025, the United Kingdom and the Republic of Korea announced a set of “Frontier AI Safety Commitments,” and the same group of companies, or at least most of them, signed on. Those commitments included promises to conduct internal and external red-teaming of AI models, which means having teams of people try to break the AI or discover severe and novel threats. They also promised to work toward sharing information on AI safety issues with each other and with governments. So the White House Accord is part of a growing pattern: international gatherings, signing ceremonies, and polished press releases, all aimed at reassuring the public that AI safety is taken seriously. But the pattern also raises an uncomfortable question. If voluntary commitments are so effective, why do we need a new one every few months? Why are the same companies signing the same kinds of promises again and again, in different cities and under different political banners? Each accord is a step, but repetition without enforcement can start to feel like a ritual. It is a way to show progress without accepting regulation. None of the companies involved, Google, Anthropic, Meta, OpenAI, xAI, or Nvidia, responded to requests for comment about the White House Accord, and the White House also did not immediately respond. The silence is telling. It suggests that even the people responsible for these commitments are not eager to explain exactly how they will be carried out, monitored, or held accountable.

At the end of the day, the White House Accord on Super Intelligence is a human document, with all the strengths and weaknesses that implies. It is a recognition that the people building AI are not monsters hiding in dark server rooms; they are executives and engineers who understand, at least on some level, that their creations carry enormous risk. It is also a reminder that our political system is struggling to keep up with technology that moves faster than legislation. We are trying to solve a problem that is both futuristic and ancient: how do you ask powerful people to protect others from dangers that are hard to see and hard to measure? The accord is a start, but it is not an ending. Real safety will require more than voluntary promises. It will require independent oversight, legally binding rules, and a willingness to slow down when something feels risky, even if that means losing money or losing a competitive advantage. It will require the public to stay engaged and to ask hard questions. And it will require governments to act not just as cheerleaders for innovation, but as guardians of the shared world in which that innovation happens. Until then, we have a piece of paper with six signatures and a lot of hope. That is not nothing. But for a technology that could reshape everything about human life, it is not nearly enough.