There is something quietly comforting about sorting existential dread into categories. When the conversation turns to artificial intelligence and the many ways it could go wrong, the possibilities can feel overwhelming—bioweapons, cyberattacks, nuclear launches, autonomous robots, all of it swirling into one undifferentiated blur of doom. That is why the simple act of making a list, of drawing lines between one nightmare and another, can feel almost therapeutic. In a recent conversation, the three of them—Leah Feiger, Zoë Schiffer, and Brian Barrett—found themselves doing exactly that. Zoë had arrived with a framework, a way of breaking down AI risk into three distinct buckets, and both Leah and Brian latched onto it immediately. “I love a bucket,” Leah said, and Zoë agreed. There was something almost joyful in the admission, as if they had discovered a small, sturdy tool for making sense of a very large and frightening problem. The buckets gave them a way to talk about the end of the world without being paralyzed by it. They could hold each scenario up to the light, examine it, argue about it, and decide how much worry it actually deserved. It was a way of saying: we can handle this, one piece at a time.
The three buckets themselves are the kind of thing that AI researchers will talk about if you push them hard enough. Zoë laid them out plainly: AI creates a bioweapon, AI hacks something critical, and AI does something else catastrophic that we haven’t fully imagined yet. These are the scenarios that keep experts up at night, the ones that appear in white papers and congressional hearings and anxious op-eds. But Zoë was careful to note that these are not the only possibilities. If you keep talking to researchers, if you really press them, they will eventually admit that there are other scenarios, ones that feel even more like science fiction. Imagine AI not just sitting in a server room, generating text or code, but embodied in a robot, walking around, interacting with the physical world in ways that are deeply, viscerally scary. That version of the future is further away, and the researchers themselves will concede that it is more farfetched. But it is still out there, lurking at the edges of the conversation, a fourth bucket that nobody has quite named yet. For now, the three buckets are enough. They give structure to the anxiety, a way to separate the plausible from the merely possible, and a starting point for asking the question that matters most: what do we actually need to worry about?
That question leads directly to the nuclear codes. Leah asked the obvious, slightly absurd question: where does the nuclear button fit into all of this? Brian answered with a reassuring confidence. “That’s the one that we actually don’t have to worry about,” he said, and Zoë agreed. Leah, understandably, wanted more. The exchange that followed was a perfect illustration of how to think clearly about AI risk. Brian reached back to the 1983 film War Games, a movie that has become shorthand for the fear that machines might accidentally start a nuclear war. In the film, a supercomputer nearly triggers Armageddon because it convinces the United States military that it is under attack by Soviet missiles. The computer never actually pushes the button itself, but it creates the conditions under which humans might. Brian’s point was that the real risk is not a rogue AI deciding to launch nukes on its own. It is the possibility that AI could generate a false alert, a piece of misinformation, a signal that looks like an attack, and set off a chain of events that spirals out of control before anyone can stop it. He hoped, he said, that there are enough safeguards in place, enough satellite technology and human verification, to prevent that from happening. But the vector of concern is real. It is not the machine making the final decision; it is the machine creating chaos that humans then have to interpret under unimaginable pressure.
This is where Brian introduced a distinction that felt like a sub-bucket, or, as he put it, a little pail within a bucket. The difference is between human-guided AI danger and fully autonomous AI danger. When people hear that AI could create a bioweapon, they often react with disbelief. How could a machine, even a very smart one, manufacture a deadly pathogen and release it into the world? That does sound crazy. But there is another version of that scenario, one that is much more grounded in reality. There are bad people in the world, people who are willing to do horrible, dangerous, malicious things. Those people could use AI as a tool, as a way to access information and instructions that might otherwise be difficult to obtain. They could use it to design a bioweapon and then release it themselves. In that scenario, the AI is not acting on its own; it is an enabler, a weapon in human hands. That is not farfetched at all. It is, unfortunately, entirely plausible. As Zoë put it, if you give people access to information about how to create a bioweapon, someone is going to be dumb enough and reckless enough to actually do it. People are already doing all sorts of horrible things, often with inspiration from the internet. Why would AI be any different?
Zoë’s response to Brian’s distinction was immediate and emphatic. She bought it completely. The human-guided scenario, the one where a bad actor uses AI to amplify their own destructive capabilities, is totally valid. It is real, it is present, and it is dangerous. But the fully autonomous version—AI deciding on its own to create a bioweapon and release it, without any human prompting—is a different story. That seems more farfetched, at least for now. It is not impossible, and it is worth thinking about, but it is not the thing that keeps her up at night. The same logic applies to the other buckets. AI hacking something critical, for example, could happen in two ways. A human hacker could use AI to find vulnerabilities and break into systems, which is a clear and present danger. Or AI could decide on its own to hack into infrastructure, motivated by some goal or directive that we do not fully understand, which is a much more speculative scenario. The distinction matters because it changes how we think about solutions. If the danger is human-guided, then the answer lies in controlling access, regulating powerful tools, and holding people accountable. If the danger is autonomous, then the answer lies in alignment, in teaching AI to want what we want, in building safeguards into the technology itself. Both are important, but they are not the same problem.
In the end, the conversation circled back to the idea that buckets, even imperfect ones, are useful. They force us to be specific, to separate the realistic from the speculative, to avoid the trap of treating every AI risk as equally likely. The nuclear codes scenario, for all its cinematic drama, is probably not the thing we need to worry about most. The human-guided bioweapon scenario, on the other hand, is uncomfortably plausible. And the autonomous robot scenario, the one where AI is walking around and interacting with the physical world, is further away but still worth keeping in mind. The real lesson is that we need to be honest about where the danger actually lies. It is easy to fixate on the most dramatic possibilities, the ones that make for good movies and good headlines. But the more mundane dangers, the ones that involve human malice combined with machine intelligence, are often the ones that deserve our attention. That is not to say we should ignore the bigger questions. We should not. But we should approach them with clarity, with nuance, and with a willingness to admit that some things are more farfetched than others. As the conversation showed, even a simple framework can help. A bucket is just a container, but it can also be a way of keeping our fears from spilling over into panic. And in a world where AI is changing faster than we can fully understand, a little structure can go a long way.
