AI safety researcher Geoffrey Irving puts extinction-from-superintelligence odds at ~50%
AI safety researcher Geoffrey Irving publicly stated he believes there is roughly a 50% probability that superintelligence kills all humans, with the risk hinging mainly on actions over the next few years to a decade. He does not expect his estimate to converge below 10% or above 90% before the outcome arrives — in effect, humanity must act under uncertainty.
Confirmed
- Irving has worked at Google DeepMind, OpenAI and Google Brain, and served as chief scientist at the former UK AI Security Institute; some coverage also refers to him as an Anthropic researcher. He posted the estimate himself on X.
- His rationale for 50%: even after years of thought, model uncertainty remains large; years ago he gave a 10–90% range in which he did not consider himself calibrated, so he judged taking the mean the most honest option.
- He places the key risk window within the next few years to a decade.
Why it matters
- This is an explicit, high point estimate of extinction-level risk from a prominent frontline AI safety researcher, not a vague expression of concern.
- The post sparked controversy, with some users accusing the estimate of being fraudulent, illustrating how contentious high p(doom) statements remain in the community.
- By insisting the estimate won't converge before the outcome, Irving is effectively arguing that uncertainty itself should be an input into current AI safety decisions, with direct implications for policy and research priorities.
2026-09-10 ~ 2026-09-11 · 6 related posts
- Episode 1: Anthropic Researcher's High-Profile Exit Over AI Doom Sparks Conspiracy Claims and Safety Debate(2026-09-09, 238 posts)
- Episode 2: Ex-DeepMind researcher says colleagues genuinely believed AI could kill everyone(2026-09-09, 2 posts)
- Episode 3: Anthropic's Pre-IPO Doom Talk Draws Wave of Mockery Over Conflicting Incentives(2026-09-09, 11 posts)
- Episode 4: AI Insiders Fear Loss of Control as Anthropic Races Toward IPO(2026-09-09, 2 posts)
- Episode 5: AI Community Debates P(doom): Show Your Reasoning(2026-09-09, 3 posts)
- Episode 6: AI Extinction Warnings Spark UK Parliament Debate(2026-09-09, 2 posts)
- Episode 7: Ex-Anthropic researcher's resignation warning sparks cross-media and political push for AI regulation(2026-09-10, 32 posts)
- Episode 8: Debate Rages Over Whether AI Doomers Truly Believe Their Warnings(2026-09-10, 4 posts)
- Episode 9: Cambridge Professor Puts AI Doom Odds Above 50%(2026-09-10, 2 posts)
- Episode 10: Anthropic Researcher Says AI Developers Believe Extinction Risk Is Real(2026-09-10, 2 posts)
- Episode 11: Anthropic's Risk Statement Slammed as Corporate Speak, Safety Credentials Questioned(2026-09-10, 5 posts)
- Episode 12: Wildeford Urges Anthropic to Replace PR Team with Researchers(2026-09-10, 3 posts)
- Episode 13: Anthropic Researcher's Viral Resignation Post Sparks Coordinated-PR Allegations(2026-09-10, 16 posts)
- Episode 14: AI safety researcher Geoffrey Irving puts extinction-from-superintelligence odds at ~50%(2026-09-10, 6 posts)
- Episode 15: Top Lab Insiders Warn of AI Extinction as Coxon's Resignation Post Tops 160M Views(2026-09-10, 23 posts)
- Episode 16: Ex-Anthropic Researcher Warns NBC of Recursive Self-Improvement AI Risk(2026-09-11, 3 posts)
- Episode 17: Ex-DeepMind researcher reveals $500K+ salary, quantifying the cost of whistleblowing(2026-09-12, 5 posts)
Primary sources
- [source] AI safety researcher Geoffrey Irving puts ~50% odds on humanity dying from superintelligence — geoffreyirving · 2026-09-10
- [source] Geoffrey Irving puts 50% odds on superintelligence killing everyone, citing deep model uncertainty — geoffreyirving · 2026-09-10
- Anthropic researcher defends 50% doom estimate as critics call it fraud — TinfoilTricorn · 2026-09-11
- Researcher estimates ~50% chance superintelligence kills everyone, sparks pushback — TinfoilTricorn · 2026-09-11
- Safety researcher puts ~50% odds on superintelligence wiping out humanity within a decade — inductionheads · 2026-09-11
1 near-duplicate retellings: zetalyrae