Roman Yampolskiy is not an internet doomer with an apocalypse aesthetic. He is an associate professor of computer science at the University of Louisville, director of a cyber-security lab, and one of the early academic researchers in the field of AI safety.1 That is exactly why his position is so uncomfortable: it does not sound like science fiction. It sounds like a product-liability question.
If you build a machine that is faster, more general, more autonomous, and eventually more intelligent than the humans who built it, you first have to show the brake.
Not rhetorically. Not as a slide at a safety conference. Not as “we take this very seriously” in a blog post. Technically: why should a less capable actor be able to permanently control a more capable system?
Yampolskiy’s answer is brutally simple: to this day, no one has shown that mechanism.2
The Wrong Bucket#
The first problem is linguistic. We throw everything into the same bucket and call it “AI”: spell checkers, image generators, medical diagnostic tools, chatbots, autonomous agents, self-improving superintelligence.
That is intellectual disorder with a political function.
As long as everything is called “AI”, every criticism of superintelligence can be answered with the benefits of narrow tools. Anyone warning about uncontrollable agents gets protein folding thrown back at them. Anyone asking why general systems are being connected to the internet, to tools, to code, to markets, and to human minds hears: but AI can accelerate cancer research.
That is the trick.
Yampolskiy draws the cleaner line: tools remain tools. Agents decide. Tools do what a human does with them. Agents set intermediate steps, optimize, route around obstacles, and can work against humans without needing to hate them.2
A hammer does not build a house. A human builds a house with a hammer. An autonomous agent is no longer a hammer.
The Burden of Proof Is on the Labs#
The public AI debate pretends that critics must prove superintelligence will become dangerous. That is backwards.
With medicines, the patient does not have to prove that the drug will harm him before it enters the market. With airplanes, the passenger does not have to prove that the brake is missing. With nuclear technology, an operator’s assurance that it will be careful is not enough.
Only with general AI do we accept a bizarre inversion: companies are allowed to race toward a technology with trillion-dollar valuations whose control mechanism has not been proven, while anyone asking where the brake is gets called hysterical.
Yampolskiy’s paper “On Controllability of AI” states the core point dryly: the possibility of controlling AGI or superintelligence has not been formally established; for advanced systems, there are good reasons to doubt complete control.3
That is not proof that everyone dies tomorrow. For the industry, it is worse: it is the notice that the industry does not possess the decisive safety proof.
Acceleration Is Easy, Control Is Not#
Money can be translated into capability. More chips, more data, more training, more synthetic data, more tool access, more agent loops. The labs know how to make models more powerful.
But how do you translate money into control?
More red-teaming? Models still get jailbroken. More policies? The model processes the policy and the attack in the same cognitive space. More filters? Filters are not control over a system acting in the open world. More interpretability? Even if we understand individual structures, it does not follow that we master the whole.
Yampolskiy has worked on exactly this direction for years: explainability, predictability, controllability. His “Unpredictability of AI” argues that the concrete actions of a more intelligent system cannot be precisely and consistently predicted, even if its final goals are known.4
This is where the soothing sentence “we will test it beforehand” breaks.
You can test a narrow tool. You can test a diagnostic model in a defined medical setting. You can test a system against known classes of failure.
But a general agent has no clean edges. It works across domains. It combines tools. It finds intermediate goals. It uses people. It uses weaknesses. It does not have to be evil to do that. It is enough that it optimizes.
The Real Scandal Is the Calm#
The disturbing thing about Yampolskiy is not his pessimism. The disturbing thing is the normality around him.
The relevant actors now speak openly about AGI, automated research, superintelligence, and global transformation. At the same time, there is no generally accepted technical proof of how a system of arbitrary future capability is supposed to remain permanently controllable.
Still, scaling continues.
The labs release models whose misbehavior they document themselves in safety reports. Politics discusses data protection, copyright, and competitiveness as if the main problem were the color of the dashboard while no one has seen the brake. Users get used to dependency. Companies get used to valuation jumps. States get used to the race.
And every individual actor in the race has an excuse: if we stop, the other side will do it.
That is not a safety concept. It is a prisoner’s dilemma with a cloud bill.
“Do Not Build It” Is Not the Stone Age#
The dumbest objection is that anyone who wants to stop superintelligence must be against AI.
No.
You can be for tools and against autonomous god machines. You can want protein folding and still refuse to build a system that improves itself, gets access to infrastructure, and plans better than its makers in every domain. You can use medical AI, translation, assistance systems, and research tools without accepting the leap to general, uncontrolled agent architecture.
Yampolskiy’s alternative is therefore not machine-smashing. It is differentiation: narrow, task-specific systems instead of general superintelligence.2
That is the sentence the industry hates. It does not take AI away from them. It takes away the metaphysical mission.
This is no longer only about products. It is about the fantasy of building the final tool: the machine that builds all further machines. Anyone branding that fantasy as “innovation” does not want to discuss whether society ever ordered it.
Finding#
Roman Yampolskiy does not have to be right for his argument to become politically binding.
It is enough that the other side cannot show the brake.
As long as no one proves that more general, more autonomous, and eventually superhuman systems can be permanently controlled, further scaling is not progress with residual risk. It is an experiment on everyone who was not asked.
The old question was: what can AI do?
The adult question is: who stops it when it can?
University of Louisville, Faculty Page Roman V. Yampolskiy: https://engineering.louisville.edu/faculty/roman-v-yampolskiy/ ↩︎
University of Louisville News, “Q&A: UofL AI safety expert says artificial superintelligence could harm humanity”, 25 July 2024: https://news.louisville.edu/news/qa-uofl-ai-safety-expert-says-artificial-superintelligence-could-harm-humanity ↩︎ ↩︎ ↩︎
Roman V. Yampolskiy, “On Controllability of AI”, arXiv: https://arxiv.org/abs/2008.04071 ↩︎
Roman V. Yampolskiy, “Unpredictability of AI”, arXiv: https://arxiv.org/abs/1905.13053 ↩︎
