What is P(doom)? — AI Glossary

What it is: P(doom) is shorthand for “probability of doom” — specifically, the probability that advanced AI causes human extinction or civilizational collapse. AI researchers and commentators publicly share their P(doom) estimates, which range from near-zero to greater than 50%.
Who it is for: Anyone following AI safety discussions, podcasts, or social-media debates about AI risk.
Best if: You’ve seen the term in a tweet, podcast, or article and want to understand what it actually means.
Skip if: You’re uninterested in long-term AI risk — P(doom) is unrelated to using AI tools day-to-day. Want one practical AI workflow every morning? Subscribe to our free daily newsletter.

What is P(doom)?

P(doom) — pronounced “p of doom” — is shorthand for “probability of doom,” the estimated likelihood that advanced AI causes human extinction or civilizational catastrophe. The term emerged from AI safety communities and entered broader discourse around 2023, especially after the public “AI doom” debates triggered by ChatGPT-4 and the warnings from prominent AI researchers like Geoffrey Hinton.

People in AI publicly share their P(doom) estimates as conversation starters. Numbers range wildly: some researchers say <1%, others 10-25%, and some at the high end estimate 50%+. The variance reflects genuine disagreement about both how powerful AI will become and how hard alignment will be.

Why does P(doom) matter as a concept?

P(doom) is the most compressed way to communicate someone’s overall view on AI existential risk. Saying “my P(doom) is 10%” instantly conveys a position — concerned enough to take it seriously, not so concerned to advocate stopping AI development. It’s a useful shorthand for fast disagreement.

The numbers also drive policy and investment positions. Someone with P(doom) of 1% supports modest AI safety investment. Someone with P(doom) of 25% argues for major safety-first regulation and possibly pauses on frontier development. People at the high end (P(doom) above 50%) sometimes advocate stopping certain AI research altogether.

What are some publicly-stated P(doom) estimates?

Some notable public statements (numbers shift over time; these are representative):

• Yann LeCun (Meta Chief AI Scientist) — close to 0%; argues current AI architectures don’t lead to existential risk
• Geoffrey Hinton — somewhere around 10% (and notably revised upward after leaving Google)
• Dario Amodei (Anthropic CEO) — has cited 10-25%
• Eliezer Yudkowsky (Machine Intelligence Research Institute) — very high, often above 90%; widely considered an outlier
• Many median AI researchers — surveyed estimates often cluster around 5-15%

Critics push back that P(doom) as a single number is unfalsifiable, overweights one specific failure mode (existential risk) versus other AI harms (bias, misinformation, labor disruption), and rewards confident pronouncements over careful uncertainty. Defenders argue it’s a useful shorthand for an unavoidable conversation. Reasonable people disagree on both the methodology and the number.

Related terms

• AGI
• Superintelligence
• AI Safety
• AI Alignment

Learn more on Beginners in AI

• AI Glossary

Sources and further reading

• AI Impacts — research on AI timelines and risk
• Anthropic — Core Views on AI Safety

Last reviewed: May 2026. AI terminology evolves quickly — verify specifics on the official source pages above.

You may also like

• AI Glossary
• AGI
• Superintelligence
• AI Safety
• AI Alignment