Nvidia's Jensen Huang Rejects AI Doom Predictions, Urges Full-Speed Building
Nvidia CEO Jensen Huang rejected predictions that artificial intelligence will destroy humanity, telling CBS Sunday Morning there was a “0% chance” of AI bringing about the end of the world.
The remarks came in an interview with CBS News correspondent Jo Ling Kent, published as an extended video on Sept. 20, 2026. The conversation focused directly on fears about AI capabilities and timelines for catastrophic outcomes. CBS News
Huang dismissed those timelines directly. He told CBS News that predictions AI could destroy humanity within a few years are irresponsible and not based on science. In a separate account of the same interview, he was quoted as saying of those sounding the alarm about AI dangers: "Scaring people is unnecessary. It is irresponsible." The Verge
He used the same language for proposals to slow down. Huang said calls by CEOs like Anthropic's Dario Amodei and OpenAI's Sam Altman to slow AI development are "not grounded in science." He rejected those calls and said AI should be developed "as fast as we can," according to CBS News accounts published Sept. 18 and Sept. 20. CBS News
Huang also rejected additional oversight as unnecessary. He argued there was no need for new rules, laws, or guidelines for AI.
On extinction scenarios specifically, Huang described AI extinction warnings as "doomsday narratives" in comments to CBS News. He told the outlet there is a "0% chance" that 2030 will be the end of the world. In the video interview, he said he completely disagrees with the idea that AI will destroy the world by the end of the decade. CBS News
The broader context here starts with the wording itself. Zero does not mean low risk. It means no risk. Much of the safety debate has centered on capability thresholds, meaning preset capability levels that trigger extra review, evals, meaning standardized tests of what a model can do, deployment controls, and post-training safeguards, meaning safety checks added after initial training. Huang is shifting the question from how to manage risk to whether slowdown arguments meet the standard of evidence engineers would normally require before limiting a system.
Looking at what this means in practice for builders and operators, the stakes come down to speed and permission. A slowdown approach points to gated releases, stricter review, and possible outside limits on training and deployment. A full-speed approach keeps responsibility with developers and existing laws, with testing in live use as the main way to find failure modes, much as road testing reveals issues lab tests miss.
In my view, both approaches involve tradeoffs familiar to anyone who has shipped complex systems. Moving fast surfaces problems quickly, but it passes those problems on to users. Moving slowly allows more structured testing, but it concentrates decisions about acceptable risk in fewer hands and slows learning from real use.
Worth flagging is Huang's choice to treat public alarm itself as a problem. He describes fear not as a natural side effect of uncertainty but as irresponsible conduct. For a technical reader, the distinction matters. There is a difference between measuring a hazard and amplifying it. Mixing the two harms calibration, meaning the ability to judge risk accurately, and that leads to weaker design and policy decisions.
In my view, none of this settles the underlying technical questions. Claims about 2030 are not testable in the way claims about loss curves, meaning graphs that track training error, eval scores, or inference reliability, meaning consistent performance in live use, are testable. They are forecasts about emergent behavior, meaning new abilities that can appear as systems scale up and get widely deployed. Readers should treat them as forecasts, however confidently stated, and keep attention on what can be measured: system behavior, misuse vectors, meaning ways the system could be abused, operational controls, and incident data.
In practical terms for teams deciding roadmaps, hiring, and infrastructure commitments, what remains concrete is Huang's stated preference. Develop without artificial delay. Do not add new AI-specific rules. Treat catastrophic forecasts as narratives rather than findings. That preference is worth weighing on its own, apart from whether the zero-percent figure holds up.
