The median AI researcher in a major 2024 survey assigned a 10% probability to AI causing human extinction or permanent disempowerment, double the median in earlier rounds. The average estimate was about 18%, showing that recent public warnings from laboratory staff are not isolated views even though researchers disagree widely.
The survey, published by AI Impacts, covers more than 1,500 researchers and is the latest round in a series running since 2016. Respondents have also repeatedly moved their forecasts for human-level performance closer. Fifty-seven percent considered it unlikely that users would understand the real reasons behind AI decisions by 2029. Several current and former researchers have separately argued that increasingly capable systems can develop unintended goals, recognize aspects of their operating situation and become difficult to evaluate. Those are risk arguments, not evidence that an extinction event is imminent.
For the next 30 years, respondents ranked misinformation and manipulation of public opinion above misaligned-system scenarios, with access to powerful tools by dangerous groups also prominent. They overwhelmingly supported more safety research. Survey probabilities capture expert beliefs under deep uncertainty; they are not actuarial forecasts and can be shaped by question wording and the selected population. Their policy significance is that a substantial part of the field sees both ordinary misuse and low-probability catastrophic outcomes as worthy of preparation.