AI Hype Goes Nucular
The Viral ‘Study’ That Flattened Everything That Actually Matters
I. A Science-Adjacent Distraction from Real Issues
So I’m at a restaurant with friends, waiting for my fajitas to come out, and the table next to ours already has fajitas, and they smell fantastic and I cannot wait for my food, but then my friend tells me that he read 95% of AI models deploy nuclear weapons in war games. And suddenly, all I can smell is bullshit.
Is this meaningful news, or just an empty scare narrative? Or—and this is a genuinely distinct third possibility—is it a meaningful scare narrative?
As the story’s gotten picked up by more and more outlets, I’ve come to realize that it’s a meaningful scare narrative: It flattens the rational details of maybe a dozen stories, wraps them in their shared emotional content, and draws attention away from all of them. Oh, and when I say the story’s picking up steam, I really mean it.
In the ten days since New Scientist wrote about it, the story has radiated across the media landscape to Axios, Vox, and even (sadly) NPR’s On the Media. The cleanest headline, though, was in Newsweek:
AI Chooses Nuclear Option in 95% of War Simulations
Now, it turns out the operative word in this headline is “Option,” but that’s a scientific point, not a geopolitical one, and I won’t be blowing up the pseudoscientific study behind the headline until a later post.
In any case, the truth or falsehood of the headline is of very little interest to me. It turns out to be false, or possibly just bullshit (if we really want to stress-test that load-bearing use of the word “Option,”) but either way it’s meaningless. It’s meaningless in two ways, which is why I’ll cover it in two separate posts.
The headline is scientifically meaningless, since it’s shorthand for something like:
Uninterpretable, High-Dimensional, 20-Dollar-A-Month Language Model Programmed By Researcher Who Admits He Doesn’t Know How It Works and Given Prompts to Choose Nuclear Option Is Willing to Choose Nuclear Option, Depending on What the Meaning of ‘Is’ Is.
But again: That’s a separate post. For now, I want to explain the sociopolitical meaninglessness of the headline—and it’s worth noting that these stories contain no information beyond their headlines.
Their headlines are profitable distractions from a number of real news stories that just can’t bait as many clicks. Critically, that’s symptomatic of a problem that makes it hard for people who actually want real news to find real news. The AI-nuke story is little more than a shibboleth, a shareable headline that signals membership in part of society that stands against (nuclear) war and against chatbots and especially against warchatbots.
II. Real Issues
There’s a lot going on right now that can’t be summed up in the moments between tortilla-chip pick-up, guac dip, and gobble:
The U.S. and Iran and Israel are bombing like a half-dozen countries right now, which most of us can’t even name
The U.S. has stopped acting like it totally wasn’t us who kidnapped whatever South American leader we just kidnapped
Anthropic informed the Department of War that its models, like all large language models, aren’t yet reliable enough for autonomous, non-human-supervised strikes
The federal government canceled Anthropic’s $200M contract with the military, and also threatened to cancel all its federal contracts, and the federal contracts of any company that even thinks about doing business with Anthropic
OpenAI signed a deal with the federal government that same day, claiming it had secured stronger guardrails than the ones that cost Anthropic its contract—which, once details of that contract were made public, turned out very much not to be true
Shortly after telling Anthropic to take a hike, the U.S. government struck Iran using the very Anthropic software it had just publicly declared it didn’t need or want.
Anthropic is being lauded as ‘the good AI company’ even though it’s expected (in April 2026) to agree to a $1.5 billion settlement for training its models on thousands of books it pirated from a website that has since been seized by the FBI and DOJ.
Etc.
But the AI-nuke story doesn’t require an understanding of Anthropic’s federal contracts, its CEO statements, or Hegseth’s and Trump’s statements—or the ‘subjective interpretation’ thereof. It doesn’t require a consideration of corporate fungibility, or ongoing wars, or the actual Middle East, or anything else ‘political’ or ‘partisan’ or ‘polarizing.’
Not only is the 95% AI-nuke statistic (seemingly) easier to understand than these other stories: It has a number in it produced by a researcher at a university. Talk about true and important, amiright?
This study carries a distinctive air of objectivity, but it is definitely not objective. I mean, if all ‘objective’ means is that something is not open to interpretation or debate, then sure, this research is objective. But that’s not what objective means.
This story is, in a sense, not open to interpretation or debate. But that’s only because it’s about an abstruse, verbose unpublished article written in dense academese—not because it’s objective.
III. The Social and Civic Harms of Science-Adjacent Distractions
These articles about how bad AI is ultimately serve AI hype. I’ve set aside my discussion of the bullshit science here because my friend’s response when I led with that is probably a common response. His response was that the real takeaway, independent of the science, is that chatbots lack the human emotion necessary to avoid nuclear war. Even the academic article on which all these news articles was based makes that claim. And that is the most cynical, subversive kind of hype.
When we pinpoint a problem to be that ‘computers don’t understand emotion,’ we’re giving computers waaaaay too much credit. Computers don’t understand anything. They aren’t capable of cognition, let alone “metacognition,” as these articles claim them to be. Their words are not capable of expressing “reasoning,” as these articles say: they take inputs, they execute operations according to a certain syntax, and 100% of their words are outputs of those operations—not reasons for them.
Articles like these, which say that chatbots are morally bad, reinforce the hype that they are intellectually good—or, at the very least, intellectually real. What great marketing! Ed Zitron made a similar point on NPR’s On the Media about the self-serving way OpenAI CEO Sam Altman frames AI safety as an issue of the models being too good at what they do. Zitron says:
You want to know what the actual AI safety story is? Boiling lakes, stealing from millions of people, the massive energy requirements that are damaging our power grid. That is a safety issue. The safety issue Sam Altman’s talking about is what if ChatGPT wakes up? It’s marketing. Cynical, despicable marketing from a carnival barker.
And don’t even get me started about the Overton window. A yearslong headstart on autonomous nuclear chatbots is a completely artificial—but effective—way to push for chatbot use in more and more contexts. “Hey, compared to giving robots total control over a nuclear arsenal, giving them absolute power over food desert supply chains isn’t that bad, right?”
IV. Root Causes
There’s far more value in meeting with university PR staff to package and ship a press release to the media than there is to, say, making a research article strong enough to be published in an actual journal. And that is exactly the kind of structural problem with Big-S Science that makes science-adjacent garbage go viral while more important, well-crafted research flies under the radar.
I think this is all I deserve to say about science, for now, since I’m setting aside my critique of the paper and its methods until a later post. I’ve read it, it’s bad, there are specifics, we’ll get there.
But what about the news outlets? Beyond the simplicity of this story as compared to actual news, what motivates outlets to publish these kinds of articles as ‘science’ rather than, you know, publishing real science as such? Why, for instance, did New Scientist decide to run this silly story?
I couldn’t tell you, because I couldn’t read the article—it was hidden behind a paywall.