Short answer: a large and growing share of the experts who build and study AI think it could be — though estimates of how likely vary enormously. It is a serious, mainstream concern, not settled fact and not science fiction. Here's what the evidence actually says.
This is not a fringe view. In 2023, hundreds of leading figures signed a one-sentence Statement on AI Risk:
"Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."
Its signatories include Turing Award winners Geoffrey Hinton and Yoshua Bengio, and the CEOs of OpenAI, Anthropic, and Google DeepMind — the very people building the technology. That doesn't make them right, but it does mean the concern deserves to be taken seriously.
The defining feature of "P(doom)" estimates is how widely they diverge. In the largest survey — 2,778 researchers who published at top AI venues (AI Impacts, 2023) — the median chance of an "extremely bad" outcome such as human extinction was around 5%, with the mean near 16%; between a third and half gave 10% or more. Individual estimates run from below 0.01% (Yann LeCun) to 10–20% (Hinton) to well above 90% (Eliezer Yudkowsky).
The concern doesn't rest on AI becoming "evil." It follows from a few premises:
Put together, a sufficiently capable, misaligned, autonomous system could pursue objectives that conflict with human survival or human control. And catastrophe needn't be sudden: gradual disempowerment describes how humanity could lose control incrementally, with no single dramatic event.
It's a fair challenge, and worth stating honestly: extrapolating from today's models to catastrophic agents is uncertain, timelines are unknown, and some serious researchers put the risk very low. But several failure modes once dismissed as theoretical are now measured in frontier systems — models that fake alignment, attempt to disable oversight, or try to copy themselves when tested (Apollo Research, 2024). "Uncertain and unprecedented" is an argument for caution, not dismissal. See the strongest objections, steelmanned and answered, in our FAQ.
No — and fatalism is counterproductive. Catastrophe requires a conjunction of failures, each a point of intervention. Real, tractable progress is underway: interpretability is beginning to read model internals, the AI-control agenda builds safeguards that hold even if a model is misaligned, dangerous-capability evaluations now gate deployment, and governance is advancing worldwide. The outcome depends on choices we make now.
Explore the concrete threat models, the capability trends, and what's being done — all source-cited.
Read the scenarios What is the alignment problem?Related: Capabilities & timelines · Governance & solutions · What you can do