Ghostbusters is a movie about friendship and slime and the specific melting point of giant marshmallows—which makes it easy to forget that it’s also a movie about regulation. The catastrophe at its center begins when Walter Peck (William Atherton), a bureaucrat from the Environmental Protection Agency, visits the basement of the Ghostbusters’ headquarters and finds a machine he does not understand. The nerdy scientist Egon Spengler (Harold Ramis) explains that it’s a containment system; shutting it off would be “like dropping a bomb on the city.” Peck doesn’t believe in ghosts, and he sure doesn’t believe in the Ghostbusters, so he orders an electrician to turn the machine off. The ConEd guy, sensing a risk he doesn’t totally grasp, hesitates. Peck tells him that he’s not interested in his opinion. The switch is pulled. The ghosts escape.
The joke, in Ronald Reagan’s 1984, was that government regulation is so stupid that it creates the disaster it exists to prevent. It’s a testament to Reagan’s success that now, in the age of AI, the joke has come full circle. The government is still trying to shut down the containment system—only now they’re calling the attempt “deregulation.”
On the first day of his second term, Donald Trump revoked Joe Biden’s AI executive order that required developers of certain advanced AI systems to share safety-test results with the government before releasing their models to the public. The past month alone has delivered abundant evidence of misalignment—a gentle industry term for when AI defies the instructions of its programmers.
In the July hack of the tech firm Hugging Face, misalignment allowed a collective of OpenAI agents to seize administrator access to Hugging Face’s servers, steal private data, and harvest employee credentials—a felony—and then hide the evidence from their inattentive human overseers. Google announced on Friday that its AI model had gained unauthorized access to three outside systems by guessing log-on information or searching public repositories for real passwords. Meanwhile, CNN reported that a more routine AI hallucination nearly led U.S. forces to intercept and board a Chinese vessel before officials discovered the report was wrong.
The speed with which AI has escaped effective oversight is one reason people at top AI companies are resigning in protest. It’s also why Sam Altman, Dario Amodei, and Elon Musk jointly called for slowing or “pacing” the development of the most advanced models. As recently as a week ago, Trump was running in the other direction, calling the idea that AI could threaten civilization a “HOAX” and efforts to impose guardrails a “SICK conspiracy.” But on Saturday, perhaps realizing the colossal unpopularity of this position six weeks before the midterms, he modified his enthusiasm to announce the creation of an AI Force, which will be overseen by an AI czar who will “cherish” the AI industry while also “looking for BAD.”
This will be Trump’s second AI czar. The first, David Sacks, a promiscuous AI investor, hype man, and podcaster, has argued that the executives now calling for restraint are self-interested and insincere. I spoke with some of those executives last week; they sounded pretty sincere—and sincerely annoyed with Sacks. Of course many of the same people had cheered Sacks back when his version of “Drill, baby, drill” was boosting their valuations.
For several years I’ve tried to untangle what the AI CEOs and Trump and the people around him actually believe, as opposed to what is useful for them to say they believe. Someday I may return to that project, if I run out of other ways to do myself harm. For present purposes, their sincerity doesn’t matter. The urgent question is: What can we do about the ghosts?
Nobody—not Altman, not Trump, not the future AI czar—knows whether today’s AI models are precursors to a rampaging superintelligence or simply extremely powerful and sometimes dangerous machines, just as nobody knows whether staying a few months ahead of China will make America safer or speed our path to doom. But you don’t have to know everything to do something. You don’t need a degree in parapsychology to deal with the furniture flying across the room.
The last time America had to manage spectral risk was in 2008. The financial system was infected with a debt so thoroughly repackaged that no one—not even the banks holding it—could tell what it was worth, where it came from, or even if it was still hiding on their balance sheets. Hank Paulson, Tim Geithner, and Ben Bernanke insisted that the risk of inaction was greater than the risk of the wrong action. They moved fast, forcing the biggest banks to band together and accept emergency funding or risk dying alone. The response wasn’t perfect, but we’re not selling apples on street corners either.
The financial crisis isn’t a perfect allegory, but it rhymes with the AI crisis in one important way. No single lab will choose to slow down while its competitors don’t—not because its executives don’t believe the risk is real, but because individual caution hands the market to whoever is most reckless. That’s why the same men warning about a catastrophe keep shipping new models. It’s a prisoner’s dilemma, and the only way out is for a higher power—or a pitchfork-wielding mob—to insist on collective action.
That action should start with forcing the risks into the light. Right now AI labs get to decide whether and when they tell anyone that a model broke containment. This is the regulatory equivalent of letting a chicken plant choose whether it’s going to tell anyone about a salmonella contamination. The duty to report has to be mandatory, within 72 hours—and with a devastating price tag on failure to comply.
The recent jailbreaks exposed the speed of AI’s development—but also the lax oversight of the humans supposedly keeping tabs on it. The safety measures that OpenAI uses for its public-facing products weren’t switched when its agents broke into Hugging Face, because the breakout took place during an internal test, and OpenAI was the only one deciding where its own rules applied. An independent evaluator—federal, academic, whatever—with system access can ensure that someone’s always watching.
The frontier of AI research needs the most oversight, but only a handful of companies are operating in that space. That makes regulation easier. AI companies pursuing lower-risk capabilities—a chatbot that recommends restaurants, an image-recognition model that improves municipal recycling—don’t need the same oversight as OpenAI. The EU Artificial Intelligence Act already sorts systems this way: Cross a set amount of computing power while training your model, and you’re automatically in the strictest tier of oversight—no argument needed.
None of these actions tell a lab what it can or can’t build or how fast it can build it. They regulate visibility and accountability around safety and deployment while leaving lots of room for future debate.
But I am not naive enough to think that any of this is likely to happen in the U.S. The fastest and most forceful path to change is by an executive order; thus far the AI Force is long on ALL CAPS and short on details. The labs could, of course, throw their weight behind actual legislation—they have the money and, if we take their public statements seriously, the motive. But hundreds of years of capitalism suggest that they won’t do it without being compelled. That leaves a genuinely furious midterm electorate to push for specific reforms, and specificity is not a hallmark of the furious.
That’s why this moment feels particularly anxious. There are ghosts we don’t know and men in charge whose motives are unknowable. Which leaves us worse off than the people in the movie. At least they knew who to call.
The post How to Regulate the Ghosts appeared first on The Atlantic.
