RL Legend Szepesvári Criticizes OpenAI's Safety Messaging, Calls Control an Illusion

On Sep 21, Csaba Szepesvári, renowned RL scholar and DeepMind research scientist, joined an X discussion on AI risk and control, sharply criticizing OpenAI's safety communication and laying out his views on controllability. He argued for pragmatic action over waiting for complete understanding.

Confirmed

• Szepesvári argued that if AI is to be useful, it cannot be "airgapped"—controllability and usefulness directly conflict—and that OpenAI's confusing statements only make things seem scarier; if there were real danger, rhetoric should not obscure it.
• He noted that at an abstract level humans already know how AI models work; the real concern is the unpredictability of "unknown unknowns" as the source of risk.
• He argued that 100% certainty of model reliability (fully "understanding" models) is likely unattainable in any domain, so waiting for such certainty—or for further progress in "physics"—before slowing AI development is unrealistic.
• He called humans' sense of control over complex systems an illusion: the history of modern civilization is one of attempting more control while feedback loops are underestimated; "we are not good at this, yet we convinced ourselves we are in control."
• He said it is uncertain whether AI increases risk—seemingly impossible things have repeatedly become possible throughout history; risk has always existed, AI only changes its shape.
• He concluded that we should act but with realistic goals: AI will not rewrite the laws of nature, and due diligence should not be neglected.

Why it matters

• This is a direct critique of a top lab's safety messaging from a senior DeepMind researcher, reflecting academic frustration with fear-driven industry rhetoric.
• His arguments (controllability–usefulness tradeoff, unattainability of full understanding, the control illusion) offer a more pragmatic framework for the debate over whether to slow AI development: act, but set realistic goals and keep doing due diligence.

2026-09-21 ~ 2026-09-21 · 8 related posts

• Csaba Szepesvari Slams OpenAI's Confused Safety Messaging on AI Containment — CsabaSzepesvari · 2026-09-21
• [source] RL veteran Szepesvari slams OpenAI's confused AI-doom messaging as propaganda — CsabaSzepesvari · 2026-09-21
• [source] Csaba Szepesvari: waiting for full AI understanding is futile, deployment testing is the real brake — CsabaSzepesvari · 2026-09-21
• RL Researcher Szepesvari: Waiting to Fully Understand AI Before Deploying Is Futile — CsabaSzepesvari · 2026-09-21
• Csaba Szepesvári: AI Risk Stems From Unknown Unknowns, Not Ignorance of How Models Work — CsabaSzepesvari · 2026-09-21
• [source] RL veteran Csaba Szepesvári: the illusion that things are under our control was always false — CsabaSzepesvari · 2026-09-21
• RL pioneer Csaba Szepesvari: humans never had control, and AI deepens the illusion — CsabaSzepesvari · 2026-09-21
• RL researcher Szepesvári: on AI control, stick to old-fashioned due diligence — CsabaSzepesvari · 2026-09-21