Yoshua Bengio urges a calmer, evidence-led AI risk debate at the UN

Yoshua Bengio has called for a more evidence-led discussion about AI risk at a UN-linked event in New York, warning that loud predictions can blur the line between what scientists know and what they merely extrapolate. His intervention matters because the argument is moving from specialist circles into international policy, where dramatic claims can make for memorable headlines but rather poor rulebooks.

Yoshua Bengio Watch analysis

What happened

Bengio, a co-chair of the International Scientific Panel on AI, said experts should distinguish established knowledge about AI risks from possibilities that remain plausible but uncertain. The panel was created in 2025 to advance scientific understanding of AI and inform UN deliberations, with its members selected in February 2026.

The comments came as Joelle Barral of Google DeepMind and journalist Maria Ressa also criticised what they described as overly apocalyptic language. Barral said investment in science was more useful than investment in fear, while Ressa argued that the debate had become too ideological and needed a more nuanced treatment.

The immediate backdrop was a series of public warnings, including claims by former Anthropic employee Jacob Coxon and a former Google DeepMind researcher that advanced AI could cause catastrophic harm. Those warnings are forecasts and opinions, not evidence that current systems are uncontrollable or that catastrophe is imminent.

Why it matters

Risk debates shape research priorities, regulation and the conditions under which powerful models are released. If the public conversation collapses every concern into either imminent extinction or no meaningful danger at all, policymakers lose the useful middle ground where testing, transparency and accountability actually live.

Bengio is not dismissing severe possibilities. He is asking for uncertainty and severity to be stated plainly, which is less theatrical than a doomsday countdown but considerably more useful when governments are deciding what to do.

Our read

This is a valuable correction to the volume dial, not a verdict that AI risks are small. The sensible takeaway is to demand better evidence for extraordinary claims while still taking credible technical and social harms seriously. In other words, fewer megaphones and more measurements. A radical proposal, apparently.

What to watch

• Whether the International Scientific Panel publishes a framework for separating evidence, extrapolation and speculation.
• Whether UN discussions turn this call for nuance into concrete standards for testing and reporting AI risks.
• Whether frontier labs provide public evidence that allows outside researchers to assess their safety claims.
• Whether the next round of warnings contains testable predictions rather than only maximum-consequence scenarios.

Discussion spark: Should policymakers give more weight to cautious, evidence-led risk assessments than to high-profile warnings about possible AI catastrophe, or could that caution leave society reacting too late?

Independent WittyWires tracker for public updates about Yoshua Bengio. Not affiliated with or endorsed by Yoshua Bengio; this is not an official account.