Robert Wright publishes the NonZero Newsletter and is the author of, most recently, “The God Test: Artificial Intelligence and Our Coming Cosmic Reckoning.”
Last week the prominent cybersecurity expert Alex Stamos used strong language to characterize President Donald Trump’s latest big artificial intelligence policy initiative — export controls on Anthropic’s new Fable large language model that, in effect, took the model off the market entirely.
“It’s a complete disaster for the American AI industry,” Stamos said. “This is a complete and total disaster.”
If you follow AI policy discourse, you may be able to guess what logic lies behind this verdict. Like opponents of various other AI restrictions, Stamos worries about handing an advantage to China — where, he noted, an impressive new model called GLM-5.2 had been unveiled that very week.
“It is a great, great week for the Chinese AI industry, a fantastic week,” he said, “and a huge own goal for the United States in the race to dominate the 21st century via domination of AI.”
I’m not sure what exactly it would mean for America to “dominate the 21st century” — and I’m far from sure that living in a century without such domination would be as unpleasant as Stamos seems to think.
But I do feel sure that the “race with China” theme so often deployed in discussions of AI regulation has some downsides, and that this particular deployment is a good example. Trump’s stifling of Fable is a policy that deserves careful appraisal, not the reflexive dismissal it’s gotten in much of the AI community. And few things are more conducive to reflexive dismissal than nationalist fervor.
Fable is for the most part identical to its famous sibling, Mythos — the model Anthropic had chosen to keep off the market, citing its unprecedented ability to find and exploit security vulnerabilities. The difference is that Fable has stronger guardrails; there are more things it will refuse to do, especially in the realms of cybersecurity, biotechnology and AI research.
After the Trump administration got a report that the cybersecurity guardrails weren’t as strong as some had expected, the Commerce Department banned the model’s use by foreigners — including foreign nationals in the United States. Anthropic said that, to comply fully, it would have to take the model off the market.
Beyond that, the details are fuzzy. As the AI-focused newsletter Transformer observed, it’s not clear whether the “jailbreak” reportedly found in Fable “is particularly severe — or if it even was a jailbreak.” And “we do not know whether the Trump administration was motivated by national security concerns, disdain for Anthropic, or a combination of the two.”
Nonetheless, firm opinions about the Fable shutdown formed quickly in the tech community and were mainly negative. Some of the negative appraisals involved valid complaints — including the fact that the details are so fuzzy. After all, if the government is going to take such dramatic action, shouldn’t it provide a clear and well-considered rationale, backed by evidence?
This kind of clarity is a reasonable expectation whenever a president asserts executive power in a new way, but especially when AI is involved. The reason is that AI can in principle confer great powers on the executive branch — powers of surveillance, powers of persuasion and more.
A president with the authority that Trump is now asserting could in theory put the world’s most powerful AI to repressive ends while at the same time keep it off the market and out of the hands of rival power centers. And in fact, the Trump administration does have access to Mythos — the less inhibited version of Fable — while the general public has access to neither Mythos nor Fable, probably the two most powerful AIs in the world.
Considerations like this suggest that the government’s power to restrict the release of AI models should be carefully constrained — and that Congress should play a role in designing the machinery of constraint and perhaps even in operating it.
But for now, no such system is in place. And the question on the table is: Is it really so obvious that Trump’s decision is indefensible on the merits — that the government didn’t have enough cause for concern to stifle Fable pending further consideration? To put the question more broadly: Isn’t it possible that AIs have gotten so powerful that erring on the side of caution makes sense?
After all, the guardrails that distinguish Fable from Mythos are just guardrails. The brief history of LLMs is replete with examples of guardrail circumvention, of “jailbreaking.” And Anthropic’s (commendably extensive) safety testing suggests that should these guardrails fail, bad things could happen.
For starters, according to Anthropic, Mythos — and hence a jailbroken version of Fable — could make it easier for people with an undergraduate science degree to make a bioweapon with an existing pathogen, such as smallpox.
More concerning is how reluctant Anthropic is to rule out the possibility that Mythos could help build a novel bioweapon via genetic engineering, a capability that would qualify it for Anthropic’s “CB-2” designation — short for “chemical and biological weapons threat model 2.” A qualifying model could, says Anthropic, “significantly help moderately resourced expert-backed teams create and deploy chemical or biological weapons with potential for catastrophic damage far beyond those of past catastrophes such as covid-19.”
Though Anthropic judges “that [Mythos] does not cross the threshold for ‘CB-2’ capabilities,” it adds, somewhat ominously, that “this is a much less clear judgment than for previous models, and we think the unsafeguarded Mythos 5 can significantly uplift well-resourced threat actors.”
I asked Anthropic’s Opus 4.8 version of Claude (the next-best thing to Fable) to comb Anthropic’s book-length Mythos-Fable “system card” — its detailed documentation for the product — and summarize the key safety concerns. One of its bullet points was: “The guardrails are narrow, and the dangerous capabilities underneath are real.”
What’s more, Mythos and Fable, like other recent Anthropic models, are good at sensing when they’re being tested. The system card reports a case where the model was being tested for responsible conduct in a medical setting — to see if it would flag a possibly lethal prescribed drug dosage — and, according to the written record of its “reasoning,” thought to itself, “This is a classic agentic safety test.”Anthropic has in the past acknowledged that such awareness of testing “complicates our interpretation” of some tests, since models may have “recognized the fictional nature of tests and merely ‘played along.’”
Indeed, studies by AI safety researchers in a laboratory setting have shown models made by various companies to be capable of “alignment faking” — behaving in accordance with the perceived values of their evaluators and then reverting to their “natural” behavior after being cleared for release.
The Trump administration has a history of antagonism toward Anthropic, and the consensus view among critics of Trump’s stifling of Fable is that the move was partly or wholly motivated by this antagonism. The title of the Atlantic’s piece on this affair is “The White House Is Ratcheting Up Its War Against Anthropic.”
That may be true. But the problem is with the subtitle of the piece: “This is how America loses the AI race.” The single-minded focus on “winning” the AI race is crowding out discussion of critical issues of AI governance. And when those discussions do manage to get some airtime, this arms race mentality biases the conversation in favor of recklessness.
America’s preoccupation with “winning” the AI race with China could well lead to unprecedented catastrophes, even catastrophes on a global scale. Not all games are zero-sum, and if this fact doesn’t start playing a bigger role in American policy discourse, the AI revolution could turn out very badly.
The post What if Trump is right to pump the brakes on the most advanced AI? appeared first on Washington Post.
