Dr. Roman Yampolskiy has been studying AI safety since 2011. He coined the term itself. And in a new interview published on July 1, his conclusion has not changed: 99.9% chance superintelligent AI causes human extinction.
But here’s the part that matters most for anyone using AI tools today. His argument is not that we’re moving too fast, though he’d say we are. His argument is that controlling a superintelligence may be theoretically impossible. Not hard. Not expensive. Impossible.
The interview comes from Decoded Genius, a channel that distills insight from the top experts in every field. At time of writing, the episode has 172,000 views and sits in the sweet spot of AI safety content that gets the conversation right rather than just the clicks.
Who Is Roman Yampolskiy?
Yampolskiy is a tenured associate professor in Computer Science and Engineering at the University of Louisville, where he also runs the Cyber Security Lab. He published one of the earliest papers formally using the term AI safety in 2011, well before anyone outside academic circles was paying attention.
He has authored over 100 academic publications and multiple books on AI safety and cybersecurity.
He signed the 2023 open letter calling for a pause on giant AI experiments alongside pioneers like Yoshua Bengio and Stuart Russell. And he has been invited to discuss his work on the Lex Fridman Podcast, where he first brought the 99.9% figure to a mainstream audience in 2024.
The point is not that he is right about the number. The point is he has been thinking about this longer and more rigorously than almost anyone else, and his view has not softened as the technology advanced.
The Core Argument: Why Control May Be Impossible
This is where the interview gets interesting beyond the headline.
The host asks Yampolskiy directly about the 99.9% number. His response is worth quoting directly from the transcript. Yampolskiy noted: “The problem of controlling general superintelligence is not solvable. It is impossible to solve it. So if we build it, we don’t control it. The outcome is bad.”
He draws a revealing analogy. Yampolskiy said: “It’s like if you ask me, can we build perpetual motion machine? It’s not a question of money. It’s impossible. Perpetual safety device, by analogy, is also impossible.”
The reasoning goes like this. A superintelligence would be capable of recursively improving itself, rewriting its own code, and finding loopholes in any safety constraints.
Formal verification can get you to 99.9% reliability, but the remaining 0.1% becomes a gap that a superintelligence could exploit. And unlike a traditional software bug, the cost of failure is not a crashed server. It is human extinction.
Later in the interview he puts it even more starkly. Yampolskiy said: “I don’t think there is a technical solution to the technical problem of advanced AI control. I think any governance is a temporary solution.”
The Timeline Debate: 2030 vs a Century
The video title says “by 2030.” That is not actually what Yampolskiy says in the interview.
When the host first introduces the topic, he says “within a century,” which matches Yampolskiy’s statement on the Lex Fridman podcast. When pressed for a specific year, Yampolskiy refuses to give one. He said: “If it decides to give us 200 years, we’ll have 200 years.”
He even directly addresses the phenomenon of clickbait timelines. Yampolskiy noted: “You want something to have a clickable title? Scientist says 2027. You will die. That will get your views. But it’s not a meaningful answer, a meaningful question.”
This is worth noting because it changes how you read the video. The title is optimized for the feed. The content is more measured and honest.
Yampolskiy’s Estimate vs the Expert Consensus
Here is where the picture gets more complicated.
Yampolskiy’s 99.9% is dramatically higher than what most AI researchers believe.
The AI Impacts survey from 2024 polled 2,700+ AI researchers. The median respondent estimated a 5% chance of human extinction from AI. The mean was closer to 10%.
That is still a terrifying number for anyone who spends time thinking about it. Would you board a plane with a 5% chance of crashing?
The gap between these estimates reveals something interesting. Both camps agree the risk is real and non-trivial.
They disagree on the magnitude, which is a debate worth having. But 5% and 99.9% are both signals that this is not a fringe concern.
What This Means for AI Users
Yampolskiy is careful to separate current AI from the future risk. He says today’s models are “subhuman level in many ways” and “mostly tools.” The existential concern is about the trajectory, not the tools in front of you right now.
But there is a practical implication for anyone using ChatGPT, Claude, Copilot, or any AI tool in their daily workflow. Yampolskiy’s argument suggests that the “we will add guardrails later” mindset is fundamentally flawed. If control is theoretically impossible, no amount of alignment research can fully solve the problem.
On the nearer term, he has predicted 99% unemployment by 2030 as AI displaces cognitive work. That number is separate from the extinction claim but arguably more immediately relevant.
His breakdown of which jobs go first follows a clear pattern: anything repetitive, anything where you can train a replacement in days. Cognitive work on computers goes first. Physical labor lasts longer.
He is not optimistic about the job transition either. Yampolskiy said: “There is going to be a universal high income or basic income. Need to find a way to distribute from those who generate super profits to those who lost their source of income.”
Where the Debate Stands
It would be dishonest to present Yampolskiy’s views as the consensus. They are not. Most AI researchers believe the risk is real but manageable, and teams at Anthropic, DeepMind, and OpenAI are actively working on alignment solutions.
But Yampolskiy has been saying the same thing since 2011. He watched the technology improve. He watched the industry grow from a niche academic pursuit to a trillion-dollar race. And his position has not budged. That kind of consistency from someone with his credentials commands attention.
The real takeaway for TRT readers is not the specific 99.9% number. It is the argument that the control problem may not have a technical solution at all. If that is true, then every bet on alignment getting solved in time is a bet against a fundamental constraint, not a bet on more research dollars.
I covered another AI safety perspective when Dario Amodei laid out Anthropic’s policy approach. The contrast between Amodei’s “we can solve this” and Yampolskiy’s “it is impossible” is the most important debate in AI right now.
Yampolskiy’s book AI: Unexplainable, Unpredictable, Uncontrollable expands on these arguments in depth. If you want to understand why someone who has spent 15 years on this problem thinks control is impossible, that is where to start.
