It finally happened: Washington, D.C., took a decisive stand on artificial intelligence, saying “No” to one of the world’s most powerful technology companies. Since President Trump cut off foreign access to Anthropic’s most advanced AI models, prompting the company to suspend Mythos 5 and Fable 5 altogether, Silicon Valley’s consternation has been palpable.
Anthropic employees are complaining about the Trump administration targeting them, while critics question everything from the move’s legality to its geopolitical sense. But the impulse to treat advanced AI as dangerous is the right one; Anthropic has been proudly proclaiming as much.
When the Trump administration learned Fable’s safeguards could fail, they decided to simply take Anthropic’s claims at face value. Now Anthropic is objecting that, since nobody else knows how to make AI safe either, “this standard [...] would essentially halt all new model deployments for all frontier model providers.”
Well, good. I have been telling other researchers and elected officials like Sen. Bernie Sanders (I-VT) that we will need to stop AI since before ChatGPT. We’re running out of time. Anthropic itself suggested a global pause would be better than proceeding along the path they’ve charted, which involves handing over more and more AI R&D to systems like Fable, even at the risk of irrevocably losing control.
They know they’re risking the lives of everyone alive. So why aren’t they celebrating this move to restrict AI? Anthropic wants to be subject to “a statutory process that is transparent, fair, clear, and grounded in technical facts.” What about the technical fact that nobody knows how to make sure the next generation of AI doesn’t take over the world?
We’ve got more than enough evidence for a global moratorium on AI development. What’s been missing is political will. For years, I have been told that an AI moratorium is a political non-starter and that Trump, in particular, would never go for it. So much for that. The federal government’s oversight of Anthropic proves otherwise. A pro-industry, anti-regulation president is drawing a red line for AI.
When the Trump administration learned that Fable provided users with details of software security flaws that should have been withheld (a so-called “jailbreak”), the red line was crossed. Even the world’s leading AI start-up—nearing a $1 trillion valuation—had no choice but to comply.
Still, the Trump directive is a stopgap measure. And Anthropic is right that they’re being made an example of, and other AI developers are no better. Users have been jailbreaking ChatGPT since its inception. AI safety and security are plagued by long-standing unsolved research problems. I’ve personally spent over a decade working on them. They may prove intractable.
We need to face up to the fact that super-human AI is too dangerous and we shouldn’t keep building it. Following the Trump directive, it may be tempting to think that we can make AI safe by “nationalizing” it, i.e., turning it into a closely guarded national security project. But this isn’t the right approach. Again, it’s not that Anthropic failed to follow best practices for AI safety; it’s that the best practices simply aren’t good enough.
Even if the federal government mistakenly thinks it can build AI safely, what about other countries? If the U.S. doesn’t want its rivals to possess dangerous AI capabilities, the U.S. must be willing to give up on its own capabilities in good faith.
Critics argue that, even if China and the U.S. want to cooperate, there’s no way to verify they’ve actually stopped developing AI. Well, here’s one plan that’s too simple to fail: Get rid of the compute that powers frontier AI and is prompting the incredibly unpopular datacenter buildout.
This proposal may seem extreme, and even colleagues who share my concern about AI risks have told me that talking about it is pointless political suicide. But it only sounds extreme if you don’t realize the severity of the crisis. The situation with AI is, frankly, too terrible to be believed, so people have refused to believe it—AI CEOs couldn’t really be racing to build something they thought had a 10 to 25 percent chance of catastrophe, could they?
But they are. As crazy as it seems, they are. As this reality is finally starting to sink in across society, it is upending the political playing field. A year ago, AI was second from the bottom among people’s concerns; now it’s above issues like climate change and abortion. If AI continues to gobble up jobs and resources, the 2026 election may be the last one fought along left versus right lines. In 2028, the dividing line could be between AI opponents and proponents.
The weeks ahead will determine the current government’s relationship with advanced AI, and perhaps the political future of the Republican Party. Will the Anthropic move become a flash in the pan, or will we see lasting progress towards an AI moratorium? For now, we can celebrate this potentially pivotal moment in the battle to stop AI.
Some observers may choose to view this as a purely political move to punish Anthropic for its spat with the Pentagon over military use. But in the context of Trump’s recent executive order, this is just the latest sign that Trump’s until-now anti-regulation administration is starting to take the risks of AI seriously. Good.