Jensen Huang fires at "AI will end humanity" talk: fabrication and cheap hype with no scientific basis for the "10% odds" — calling for responsible optimism

"When you hear that AI will end human civilization, do not believe it." Nvidia CEO Jensen Huang trained his fire on doomsday AI rhetoric at a company meeting in Berlin on October 22, calling such claims 'fabrications' and 'cheap hype' chasing headlines: "every one of their predictions has been wrong, and the so-called 10% probability has neither scientific basis nor research behind it... these predictions only hurt people." He urged a stance of "responsible optimism" and declared now the best time to "build AI"^[1].

Huang's syllogism: predictions, motives and a replacement narrative

Unpacked, his remarks form a layered argument. The factual claim: predictions of AI ending humanity within a decade have all failed, and figures like "a 10% probability" lack scientific grounding or research support. The motive analysis: such rhetoric is fundamentally about "making headlines" — attributing doomsday narratives to the economics of traffic rather than rigorous risk assessment. The replacement narrative: swap doom for "responsible optimism" and recast building AI itself as a positive investment in humanity's future — the way to care about that future is to build, not to watch in fear^[1]. The script isn't new, but the venue matters: in Berlin, on Nvidia's global stage, just ahead of GTC Berlin — Huang is clearing the public-opinion path for European AI-compute expansion; his last European swing made "Europe's AI vision runs on plug-in datacenters" the theme.

A fair scale: doomsday and total denial both miss the evidence

Soberly: Huang's position is equally interested — as the world's largest supplier of AI compute, a cooling of AI-risk narratives directly affects both regulation and demand expectations. A fair assessment needs evidence anchors between the extremes. On one side, "AI exterminates humanity in 10 years" indeed lacks operable scientific support — such predictions are unfalsifiable and their probabilities uncalibratable. On the other, the real progress of alignment research — just last week OpenAI announced offline training for frontier models and Anthropic lobbied for a "50 lives / $1 billion" red-line bill — shows the top labs themselves are responding with engineering to the reality of unpredictable model behavior; if the risk were pure fabrication, none of that spend would be necessary. The honest middle ground: near-term risks are real (rogue agents, misuse, labor shocks), while civilization-level risk is uncertain — not impossible.

This week's footnotes: safety spending and hype coexist

The industry's timeline this week supplied ammunition to both sides. For Huang's 'hype' thesis: doomsday forecasts have indeed missed repeatedly while harvesting traffic. For the cautious camp: Moonshot AI proactively limited Kimi K3.1 Pro's cyber-attack capabilities, setting safety boundaries for open weights — a leading Chinese lab turning safety from slogan into product constraint; and the UK pledged over £600 million for AI, with $134 million earmarked for an AI Biosecurity Lab — government-side risk hedging landing in real money^[1]. In other words, the industry mainstream has moved past 'whether to manage' to 'how to manage' — the 'cheap hype' Huang attacks is real, but the prudent voices he would sweep away are precisely the ones labs and governments backed with action this week.

What to watch

First, the GTC Berlin keynote — whether Huang's 'responsible optimism' converts into concrete European AI-infrastructure commitments. Second, the academic response to the '10% odds' fight — will AI scholars publish methodological critiques of doomsday probability estimates? Third, the actual operations of hedging projects like the UK's AI Biosecurity Lab. Fourth, the signatory list of the next frontier-model safety pledge — the labs that vote with their feet say more about the industry's true risk judgment than any speech^[1].

The fight over the '10% probability'

The '10% probability' Huang singled out did not come from nowhere — it belongs to a tradition of probability estimates in AI-risk research about 'frontier AI leading to extinction-scale outcomes,' numbers that have appeared repeatedly in scholarly letters and interviews. The debate over them is at bottom methodological: critics say such figures cannot be calibrated and lack falsifiable models — essentially 'the intuitions of respected people'; defenders counter that for low-frequency, high-impact events, expert probability aggregation is the least-bad tool available — early climate science walked the same path. Huang's blanket verdict of 'no scientific basis' is rhetorically effective but epistemically shaky: for events that have not occurred and cannot be experimented on, the standard of 'scientific basis' itself needs defining. The genuinely useful debate is not whether '10% is right or wrong' but 'what evidence structure should move that number' — a question neither side has yet seriously answered.