AI safety researcher Dr. Roman Yampolsky explains why the public underestimates AI extinction risk, and why researchers keep building it anyway.
Walk down any city street and you will not see people rushing to deal with artificial intelligence risk. According to Dr. Roman Yampolsky, director of the Cyber Security Laboratory at the University of Louisville, that is not because the risk is small. It is because the human mind is built to look away from things it cannot fix.
In part one of a two-part conversation on For Humanity, host John Sherman asks Yampolsky the question at the center of the AI safety movement: if the danger is real, why does almost nobody act like it? The answer, according to Yampolsky, has less to do with AI and more to do with human psychology, incentives, and the limits of imagination.
Yampolsky compares public indifference to AI risk with how people treat mortality itself. "We know that everyone's going to die and we do absolutely nothing about that," he says. Even at conferences where he presents on existential risk, he says attendees ask about their jobs or their art, then go home and change nothing. He describes this as a kind of protective bias, a built-in mechanism that keeps people functional instead of paralyzed. Without it, he argues, "everyone would just be depressed and maybe suicidal."
He is careful to note his own uncertainty. Yampolsky recalls being convinced that earlier AI milestones, like Deep Blue beating a chess champion or IBM Watson winning Jeopardy, would change everything, and they did not. That history makes him cautious about overconfidence in either direction. But he says the trend across GPT-2, GPT-3, and GPT-4 is impossible to ignore, even if any single model turns out to be a "nothing sandwich."
Sherman raises a question he says he hears constantly: why can't people imagine AI actually killing them? Yampolsky's explanation is that most people project the risk onto software they already understand and control. "How would Microsoft Word kill you? I can just turn off the computer," he says, describing the mental shortcut people reach for. He compares the situation to an ant trying to predict the many ways a human could end its life. The ant has no comprehension of the possibilities. Yampolsky argues humans are in a similar position relative to a system smarter than themselves, unable to predict the method because they cannot model the mind doing the predicting.
Sherman cites recent polling showing roughly seventy percent of people do not support building superintelligence. Yampolsky is skeptical that this number means much to AI companies, for two reasons. First, he says polling is easy to manipulate depending on how a question is framed. Second, he argues the public does not have the technical grounding to give an informed answer in the first place. "It's like asking consent from a five-year-old," he says. "It just doesn't mean anything." Even the experts building these systems, he adds, do not fully understand how they work or what they are capable of, which raises a deeper problem: if nobody fully understands the technology, nobody can meaningfully consent to it being deployed.
One of the more striking parts of the conversation covers Yampolsky's research into the people building advanced AI systems. Sherman asks how anyone can go to work each day knowing the stakes described in their own research. Yampolsky says the answer is unremarkable. "Very normal average human beings with exactly the same values," he says. "They want to have a good job, they want to be respected, promoted." He describes an incentive structure where an individual researcher's decision to stop feels irrelevant, since another engineer will simply take the role. Yampolsky pushes back on the idea, raised by Sherman, that childhood science fiction fantasies explain why people in tech keep building toward these outcomes, pointing out that most science fiction actually depicts AI going badly.
Yampolsky is blunt about the current state of AI alignment research. He argues the field has not even settled on a definition of what alignment means, since it is unclear which agents are being aligned to, whose values are being used, and whether those values are fixed or constantly shifting. He challenges the idea that "preserve human life" could serve as a simple universal value, noting that humans themselves disagree on what that phrase covers, from questions about consciousness to disability to the earliest stages of life. In his view, much of what is currently called alignment research amounts to filtering a base model so it avoids saying anything politically incorrect or embarrassing, which he describes as "putting lipstick on a pig" rather than solving the underlying problem.
Sherman closes part one on a personal note, asking Yampolsky how he discusses these risks with his own children. Yampolsky says his kids grew up alongside his research and have their own reactions to it, including genuine confusion about why anyone healthy and successful would want to risk destroying their own existence. Sherman shares his own struggle with the topic, drawing a comparison to growing up in a household where his father worked on nuclear arms control. He frames both conversations the same way: difficult, but worth having openly rather than avoiding.
Yampolsky does not offer easy reassurance, but he also does not claim certainty. He is candid that he has been wrong about AI trajectories before and could be wrong again. What comes through across the conversation is less a prediction than a description of why the public conversation lags so far behind the technical one. As Yampolsky puts it, there is no simple set of instructions for what an individual should do differently after hearing this. What he asks for instead is honesty about the scale of the problem.
Part two of this conversation continues next week, covering what Yampolsky sees at the core of AI safety skepticism and a step he believes the federal government could take to make a real difference.
Watch the full episode on the AI Risk Network YouTube channel.
