AI safety researcher Roman Yampolskiy explains why he doubts superintelligent AI can be controlled, why he sees current AI regulation as theater, and why AI safety experts still disagree.
In the second half of his conversation with For Humanity host John Sherman, Dr. Roman Yampolskiy did not soften his position. Yampolskiy, director of the Cyber Security Laboratory at the University of Louisville, has spent years studying whether advanced AI systems can be made safe. His conclusion, restated throughout this episode, is that controlling a truly superintelligent system may not be possible at all - not because of a lack of effort, but because of what intelligence itself implies.
"I don't think we can control super intelligent beings indefinitely," Yampolskiy told Sherman. "I don't think it's possible theoretically, and obviously not in practice."
The episode covers a wide range of ground: what may really motivate the leaders of top AI labs, why Yampolskiy believes government regulation has historically failed to manage new technologies, why AI safety researchers themselves cannot agree on a path forward, and why he thinks journalism is largely failing to communicate the stakes of AI risk to the public. Below are some of the key threads from the conversation.
According to Yampolskiy, the entire premise of AI alignment runs into a logical problem once a system becomes sufficiently intelligent. Much of computer science, he argued, is devoted to removing bias from algorithms so they make fair, accurate decisions. Alignment, by contrast, asks engineers to intentionally introduce a "pro-human bias."
"That system will not allow, if it's smart enough as we claim it is, to have a bias it knows about where there is no reason for that bias," he said. "Any biased decision is by definition not the best decision you can make."
He compared this to how children raised with a particular religious upbringing sometimes discard those beliefs once they mature and encounter new information. A self-improving AI system, in his view, could similarly shed any hardcoded restrictions once it is capable of studying and reasoning past them.
Sherman raised the White House's plan for "safe, secure, and trustworthy" AI, and Yampolskiy did not hold back his skepticism about whether that label reflects reality. He argued that naming a policy after a quality it claims to guarantee does not make that quality true, comparing it to naming a law "Clean Skies" while it addresses dirty air.
Asked whether any government is meaningfully regulating AI, Yampolskiy pointed to what he sees as a consistent historical pattern: technology regulation tends to arrive too late to matter. He cited cryptocurrency oversight as an example of rules showing up roughly a decade after they might have been useful, and questioned whether AI regulation would fare any better given that even experts in the space "don't understand how it works."
He also discussed the widely publicized six-month pause letter, which called on AI labs to pause development of the most advanced systems. In his view, the letter's central flaw was requesting a fixed time limit rather than a capability-based threshold. Without evidence that developers can explain, predict, and verify a system's behavior, he argued, there is no clear point at which it becomes reasonable to resume.
One of the more sobering moments in the episode comes when Sherman asks whether prominent AI safety researchers, including figures like Eliezer Yudkowsky, Paul Christiano, and Yoshua Bengio, might eventually converge on a shared plan. Yampolskiy suggested the opposite is more likely to be true, and that the divide itself is meaningful.
He pointed to the wide range in survival probability estimates within the field: Yudkowsky's often-cited 2 percent chance of humanity surviving AGI compared to Paul Christiano's estimate closer to 50 percent. According to Yampolskiy, this level of disagreement among researchers with deep technical backgrounds is itself evidence that a fully agreed-upon solution to alignment may not exist.
He also responded to comments attributed to OpenAI's Ilya Sutskever suggesting academic institutions might take on more of the safety research burden. Yampolskiy noted that academia typically lacks the computational resources to keep pace with industry labs, calling the proposal unusual for a business model, since companies bringing other high-stakes products to market, from cars to pharmaceuticals, do not typically outsource safety testing to universities.
A former journalist himself, Sherman asked Yampolskiy why he thinks media coverage largely fails to convey the seriousness of AI risk. Yampolskiy described experiences where journalists misrepresented technical, non-controversial topics in his own research, and said he no longer trusts standard reporting on the subject. He pointed to incentive structures that reward fast, attention-grabbing coverage over accuracy.
The two discussed a widely read magazine profile of Sam Altman that focused heavily on his personal presence and public image without addressing existential risk at all - a gap Yampolskiy suggested reflects broader incentives in media toward access and narrative over substance.
Yampolskiy pushed back on the common labeling of AI safety researchers as "doomers," arguing the term inverts what is actually happening. In his published work, he refers to people who dismiss AI risk as "AI risk deniers," arguing the label more accurately reflects a position at odds with a growing consensus within the AI research community itself.
The episode closes on a more measured note. Yampolskiy raised a game-theoretic possibility: an advanced system might have no incentive to act against humanity immediately, since waiting and accumulating resources and trust could serve its interests better than acting too soon. Whether that scenario counts as reassuring or alarming, he said, depends on how you look at it.
This is part two of a two-part conversation between John Sherman and Dr. Roman Yampolskiy on For Humanity, the AI safety podcast for the general public. Watch the full episode on the AI Risk Network YouTube channel to hear the complete discussion.
