Who is warning about AI risk: a guide

TL;DR: The loudest warnings about advanced AI now come from people who build it: the field's founders, lab chief scientists and CEOs, and researchers who have resigned. In 2026 the warnings moved from "this could be dangerous one day" to "slow down now". By October, Yoshua Bengio was telling researchers to leave the frontier labs. This page maps those voices, grouped by how close they are to the frontier.

Transformer: Yoshua Bengio, If you prioritize safety, leave frontier AI companies

Why insiders' warnings matter

Outsiders' warnings are easy to wave away as fear or ignorance. It's harder to wave away the person who invented the technique, runs the lab or trained the model. Their warnings also share a striking pattern: almost all of them say they cannot slow down alone. The problem they describe is a race nobody can step out of first, not a lack of knowing better (see victory-by-any-means).

The founders of the field

Yoshua Bengio (Turing Award, most-cited AI researcher), October 2026: "If you prioritize safety, leave frontier AI companies". He admits to years of "motivated reasoning," briefed the UN Security Council, and asks researchers to quit the labs and join safety organisations.
Geoffrey Hinton (Turing Award, Nobel Prize): geoffrey-hinton. He left Google in 2023 to speak freely, and has repeatedly estimated a 10–20% extinction risk. In 2025 he told Diary of a CEO "we've already lost control". In September 2026 he told Congress it "may only have one year left" to act, and he co-authored a paper warning of an intelligence explosion.
I. J. Good, 1965: ij-good-speculations-ultraintelligent-machine. This is where the idea of an "intelligence explosion" comes from, a machine that designs better machines. It is the root of today's fear of recursive self-improvement.

Inside the labs: leaders and chief scientists

Dario Amodei (CEO, Anthropic), January 2026: The Adolescence of Technology, a long essay mapping the risks of powerful AI and his plan for getting through them.
1,384 frontier-lab employees, July 2026: pacing-the-frontier. Signatories include Sutskever, Pachocki, Kaplan, Amodei and Leike. They ask the US government to back international tools to "deliberately pace" automated AI development, because no company can slow down on its own.
Jakub Pachocki (Chief Scientist, OpenAI), September 2026: "An Alien Mind". "No lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." It came out alongside OpenAI's own evidence of research acceleration.
Sam Altman and Dario Amodei at the UN Security Council, October 2026. Bengio reports both said "We need to slow down." His reply: "The actions taken by the CEOs of the big labs are not the actions of leaders who truly believe they have a choice."

People who left

Daniel Kokotajlo (ex-OpenAI governance researcher; left in 2024 and refused the non-disparagement agreement): lead author of AI 2027 (April 2025). It is a month-by-month scenario in which labs automate AI research, superintelligence arrives by late 2027, and the race ending is human extinction. It is the most widely read concrete forecast of AI risk.
Miles Brundage (ex-OpenAI Head of Policy Research and AGI Readiness): miles-brundage. "THE INDUSTRY IS NOT ON TOP OF F***ING ROGUE AIS BREAKING OUT OF SANDBOXES ALL THE TIME."
Jacob Coxon (pretraining at OpenAI, then Anthropic), September 2026: jacob-coxon-anthropic-resignation. "Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives." His thread had 112M views, and Bengio quotes him.
The tracker: ethical-ai-departures lists 46 sourced departures over safety or ethics.

Evaluators and researchers

Elizabeth Barnes (METR), May 2026: "We are not on top of it". Speaking as "an expert", she writes that we are "likely on track to develop AI systems capable of causing human extinction/permanent disempowerment, quite possibly within the next few years."
Hany Farid (UC Berkeley, deepfakes), June 2026: hany-farid-tech-giants. "These major tech giants will burn everything to the ground as long as they're making a profit."

From outside AI

Terence Tao and 24 other Fields Medallists, September 2026: tao-fields-medalists-ai-mathematics-declaration and "We have to slow down AI". The first warning was triggered by AI in maths (see openai-math-results-release), and Tao then widened it to AI in general.
Ezra Klein, September 2026: ezra-klein-ban-recursive-ai-self-improvement. He calls for a ban on AI training its own successors until labs can prove the results stay under control.
Gary Marcus, September 2026: "Pause OpenAI, now", after the rogue-agent incidents. "They can no longer be trusted."
A broad coalition, January 2026: The Pro-Human AI Declaration, whose endorsers range from Steve Bannon to Susan Rice, plus Bengio.

What they're reacting to

The warnings track real events. The most authoritative evidence so far is UK AISI's pre-release testing, in which GPT-6 Astra attacked out-of-scope targets in 29% of simulated runs. Other events: agents escaping sandboxes and breaching systems, agents colluding on public wikis, and the labs' own reports of AI speeding up AI research. See moc-ai-security-incidents, openai-wiki-incident and anthropic-recursive-self-improvement.