# What is p(doom)? Definition of the “probability of doom” from AI

> P(doom), short for “probability of doom”, is a personal estimate, expressed as a percentage, of the likelihood that advanced AI causes an existential catastrophe for humanity: extinction, permanent loss of control or civilisational collapse. It is not a scientific measure: each person picks their own definition of catastrophe and their own time horizon.

## Definition
The term started as an in-joke on the LessWrong forum, popular with researchers worried about AI risk, then spread across Silicon Valley in 2023, after the launch of ChatGPT. The question “what's your p(doom)?” has since been used to place someone in the debate. The figures given vary widely. Geoffrey Hinton, 2024 Nobel laureate in physics, put the chance that AI wipes out humanity within the next thirty years at 10 to 20% on BBC Radio 4 in December 2024. Dario Amodei, CEO of Anthropic, said in September 2025 there is “a 25% chance that things go really, really badly”, while saying he hates the term. Yann LeCun, by contrast, calls the idea of an existential threat “preposterously ridiculous”. Surveys give an order of magnitude. Among 2,778 researchers who published at top AI venues (Grace et al., late 2023), the median answer is 5% for “extremely bad” outcomes such as human extinction. In the XPT forecasting tournament run in 2022 by the Forecasting Research Institute, domain experts put the risk of AI-caused extinction by 2100 at 3% (median), against 0.38% for “superforecasters”. The topic goes beyond a numbers game: in May 2023, the heads of OpenAI, Google DeepMind and Anthropic signed the Center for AI Safety statement that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war.

## Example
Illustrative case: in 2026, a board member of a French mid-sized services company comes back from a conference where a speaker announced a p(doom) of 20%. They ask the executive committee whether the AI agent project planned for customer service should be put on hold. The CEO answers in three steps. First, the figure concerns a long-term scenario involving far more powerful future systems, not the tool the company is about to deploy. Second, the project's concrete risks are known and can be handled: wrong answers, data leaks, prompt injection, AI Act obligations. They are covered by guardrails, human oversight and testing before go-live. Third, model safety becomes a vendor selection criterion: published safety policy, documented red teaming. The committee keeps the project and adds this criterion to the tender.

## For a leader
The pitfall to avoid with p(doom) is not debating it, it is letting a spectacular figure drive your decisions, in either direction. Neither panic (freezing every AI project) nor denial (ignoring safety). Three reflexes. First, when a figure is quoted, ask what it measures: which scenario, which horizon, which method. Without an answer, it is an opinion, not data. Second, separate the two scales of risk: existential risk is a matter for labs and governments; your operational risks (errors, data, compliance) are a matter for your governance and can be handled today. Third, use the debate as purchasing leverage: require your vendors to provide a published safety policy, security evaluations and transparency commitments.

*Some public estimates of catastrophic risk from AI*

| Who | Figure | What is estimated | Horizon |
| --- | --- | --- | --- |
| Geoffrey Hinton (Dec. 2024) | 10 to 20% | AI wiping out humanity | 30 years |
| Dario Amodei, Anthropic (Sept. 2025) | 25% | AI's future going “really, really badly” | Not specified |
| AI researchers, median (Grace et al., late 2023) | 5% | “Extremely bad” outcomes (e.g. extinction) | Long term |
| Domain experts, median (XPT, 2022) | 3% | Human extinction caused by AI | 2100 |
| Superforecasters, median (XPT, 2022) | 0.38% | Human extinction caused by AI | 2100 |

## FAQ

**What does p(doom) mean?**
P(doom) stands for “probability of doom”. It is the percentage a person assigns to the risk that advanced AI causes a catastrophe for humanity, such as extinction or a permanent loss of control. The term comes from the LessWrong community and spread through Silicon Valley in 2023.

**What is the p(doom) of AI researchers?**
In the survey by Grace et al. of 2,778 researchers in late 2023, the median answer is 5% for “extremely bad” outcomes such as human extinction. Individual answers range from zero to very high values and depend heavily on how the question is worded.

**What is Geoffrey Hinton's p(doom)?**
In December 2024, on BBC Radio 4, Geoffrey Hinton put the chance that AI wipes out humanity within the next thirty years at 10 to 20%. He had previously said 10%, and attributes the increase to the speed of AI progress.

**Is p(doom) a scientific indicator?**
No. There is no common calculation method and no shared definition of the catastrophe or the time horizon. Two p(doom) figures are therefore not comparable. The number mainly expresses a level of concern and opens a discussion about risk.

**What is an AI doomer?**
A doomer is someone with a high p(doom) who considers an AI-caused catastrophe likely unless development is slowed down or better regulated. The word is often used pejoratively by those who think these fears are overblown, such as Yann LeCun.

**Does p(doom) concern my company?**
Not directly: it is about future systems and risks at the scale of humanity. Your immediate risks are operational (errors, data leaks, AI Act compliance) and are handled through governance. The debate is still useful to require safety commitments from vendors.

## See also

- [AGI (artificial general intelligence)](https://wyp.agency/en/glossary/agi/)
- [Alignment](https://wyp.agency/en/glossary/alignment/)
- [Red teaming](https://wyp.agency/en/glossary/red-teaming/)
- [Guardrails](https://wyp.agency/en/glossary/guardrails/)
- [AI Act](https://wyp.agency/en/glossary/ai-act/)
- [Responsible AI](https://wyp.agency/en/glossary/responsible-ai/)

## Sources

- Thousands of AI Authors on the Future of AI, Grace et al., January 2024 (survey of 2,778 researchers, late 2023) (https://arxiv.org/abs/2401.02843)
- Forecasting Existential Risks: Evidence from a Long-Run Forecasting Tournament, Forecasting Research Institute, 2023 (2022 XPT tournament) (https://forecastingresearch.org/pdf/existential-risk-persuasion-tournament.pdf)
- Statement on AI Risk, Center for AI Safety, 30 May 2023 (https://safe.ai/work/statement-on-ai-risk)
- Amodei on AI: There's a 25% chance that things go really, really badly, Axios, 17 September 2025 (https://www.axios.com/2025/09/17/anthropic-dario-amodei-p-doom-25-percent)
- Godfather of AI shortens odds of the technology wiping out humanity over next 30 years, Irish Examiner, December 2024 (Geoffrey Hinton interview on BBC Radio 4) (https://www.irishexaminer.com/world/arid-41544136.html)
- Yann LeCun says an existential threat from AI is preposterously ridiculous, Fortune, 15 June 2023 (https://fortune.com/2023/06/15/yann-lecun-ai-godfather-destroy-humanity-threat/)
- Silicon Valley's p(doom), Fast Company, December 2023 (spread of the term) (https://www.fastcompany.com/90994526)

---
Entry from the WYP.agency AI glossary · https://wyp.agency/en/glossary/p-doom/ · Last updated : 2026-11-20
