How seriously should we take AI extinction risk?
Take the possibility seriously. Be precise about what the evidence can establish.
RESEARCH + EDITORIAL ANALYSIS · Updated 2026-10-07
Start by separating the questions.
When might AI systems gain a capability? How could that capability contribute to harm? What conditions would make that harm widespread or irreversible? These are different questions. A prediction about mathematical performance or research automation cannot answer all three.
Explore the AI Futures timeline to see how widely capability forecasts can extend. Each band has a source and a specific interpretation.
Open AI Timeline in the footer to explore the companion forecasts.
Expert concern is evidence about beliefs.
AI Impacts’ December 2024 survey reports substantial uncertainty and concern among responding researchers. Its catastrophic-outcome questions include extinction or similarly severe, permanent disempowerment. That combined outcome should not be shortened to an extinction-only probability. Responses also depend on wording, assumptions and the researchers who participate.
A survey can tell us how respondents judge a risk. It does not measure the future frequency of an event that has not occurred. Neither dismiss the concern nor turn an aggregate estimate into an established fact.
Examine the pathway to harm.
Google DeepMind’s technical safety discussion distinguishes misuse, misalignment, accidents and structural risks. These categories invite different questions: who can use a system, what it can do, whether its behavior can be supervised, and how deployment changes institutions or incentives.
For any alarming claim, ask which pathway it describes, which parts have been observed, and which remain assumptions. A laboratory demonstration, an extrapolation, and a real-world incident warrant different interpretations.
Source: Google DeepMind, Taking a responsible path to AGI (2 April 2025).
https://deepmind.google/blog/taking-a-responsible-path-to-agi/
Uncertainty calls for proportionate preparation.
Our editorial judgment is that uncertainty is a reason to investigate and prepare, not a reason to announce inevitability. Useful questions concern evaluation, independent scrutiny, human authority to intervene, security, incident reporting, and responsibility when safeguards fail.
These are ways to examine and reduce possible harm. They are not proof that a catastrophic outcome is either inevitable or impossible.
Keep the metaphor separate from the measurement.
The Bulletin describes the Doomsday Clock as a metaphor. It combines judgments about several threats. Its setting does not supply an AI-specific probability or a date when catastrophe will happen.
Source: Bulletin of the Atomic Scientists, Doomsday Clock FAQ.
https://thebulletin.org/doomsday-clock/faq/