Language models put AI extinction risk at ten per cent
A new preprint compares the answers of 241 models, but does not measure the actual chance of extinction.
Listen to this article
There is no audio version yet. Request one and an AI voice will read the article aloud.
Read by an AI voice.
A new preprint asked 241 language models how likely it is that future AI will lead to humanity’s extinction or leave it permanently powerless. The median answer was 10 per cent, but that figure is not a scientific risk assessment.
Researcher Julian Kauk of Friedrich Schiller University Jena put a question to language models about an extreme scenario: how likely is it that future artificial intelligence will wipe out humanity or leave it permanently powerless? According to a report on the study, the median of 241 answers was 10 per cent.
The study has initially been published as a preprint. Publication as a preprint in itself offers no guarantee that the research has been fully assessed by independent peers. The result should therefore primarily be read as an analysis of the answers models give, not as a measurement of the likelihood that the scenario will occur in reality.
A model’s answer to a hypothetical question is not an observation of a future event. Differences between the answers therefore do not in themselves show how likely that scenario is.
The research is nevertheless relevant to the debate about AI safety. Developers, policymakers and researchers use such questions to examine how models deal with uncertainty, catastrophic risks and moral dilemmas. A model that gives a high percentage therefore provides no evidence that the danger is actually that great.
The broader scientific debate is divided. The International AI Safety Report 2026 states that experts disagree about the likelihood and severity of scenarios in which people lose control of AI systems. The report therefore does not describe a unanimous assessment of these risks.
The main limitation of the new study is therefore the distance between answers and reality. Estimating a risk requires data on system capabilities, development paths, security, human behaviour and possible countermeasures. None of these factors can be summarised in the median of answers to a single hypothetical question.
The result therefore does not justify the conclusion that ‘AI has a 10 per cent chance of wiping out humanity’. The research describes language models’ answers to a question about an extreme scenario. That is useful for research into model behaviour, but insufficient as a basis for policy or personal conclusions resting on a single percentage.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The text makes the necessary distinction between a preprint, model output and an empirical risk assessment. Its substantive interpretation is cautious and consistent with the international safety report.
- confirmed 241 language models gave a median of 10 per cent in response to the question posed. — Report on the preprint. source
- confirmed The study is a preprint. — The publication is available as a preprint. source
- confirmed Experts disagree about the likelihood of loss-of-control scenarios. — International AI Safety Report 2026. source
Editor's note
The research is a preprint and measures model answers, not an empirical probability of extinction. Its central finding is therefore informative about model behaviour, but unsuitable as a prediction.Sources
- Asked the AI extinction question, 241 language models gave a median answer of 10% — On The Wire
- International AI Safety Report 2026 — International AI Safety Report
- International AI Safety Report 2026 — arXiv
More on this in Dutch media
- NU.nl — „kunstmatige intelligentie”
- De Telegraaf — „kunstmatige intelligentie”
- AD — „kunstmatige intelligentie”