A new survey of 59 leading AI safety researchers reveals a sobering consensus: there is a median 25% probability of human extinction or permanent disempowerment before 2100. Yet the findings also expose a surprising fault line in how the field prioritizes its resources, with experts increasingly concerned about AI-enabled authoritarian control rather than unaligned artificial intelligence running amok .

The survey, conducted in February 2026 among attendees of the Summit on Existential Security, asked leading thinkers in the existential risk and AI safety communities to estimate extinction probabilities, artificial general intelligence (AGI) timelines, and resource allocation priorities. The results paint a picture of a field grappling with competing visions of how advanced AI systems might threaten humanity .

What Timeline Do Experts Expect for Advanced AI?

When asked when there would be a 50% probability of developing AGI, defined as an AI system capable of automating more than 90% of jobs in the 2025 economy, respondents clustered their estimates around the early 2030s. The median estimate was 2033, with a mean of 2034. More strikingly, 73% of the 59 respondents placed their 50% AGI estimate before the end of 2035 .

For a 25% probability of AGI, the timeline compressed even further. The median estimate was 2030, with 22% of respondents assigning at least a 50% chance of AGI by 2030 itself. These timelines suggest that many leading safety researchers believe transformative AI capabilities could emerge within the next five to nine years .

Several respondents noted that the survey's definition of AGI, which included physical labor roles like construction and restaurant work, pushed their estimates later than they would otherwise be. One expert observed that if the question focused only on knowledge work, timelines would shift earlier by one to three years .

How Are Experts Reassessing AI Risk Categories?

Perhaps the most striking finding concerns where the field should direct its resources. When asked whether the AI safety community should allocate more or fewer resources to different risk categories, respondents revealed a significant shift in priorities. The strongest consensus emerged around AI-enabled human takeover scenarios, such as permanent authoritarian states enabled by AI surveillance and control. A substantial 73% of respondents said the community should direct "more" or "much more" resources toward this category, with a mean score of +0.78 on a scale from -2 (much less) to +2 (much more) .

Better futures work also received net-positive support, with a mean score of +0.51. However, the most surprising result was a slight lean against allocating more resources to misaligned AI takeover scenarios, with a mean score of -0.14. This finding masks a critical divide: experts with shorter AGI timelines (2032 or earlier) favored more resources on misaligned AI, while those with longer timelines (after 2032) leaned toward fewer resources .

• AI-Enabled Human Takeover: 73% of respondents said the field should direct more resources toward scenarios where AI enables authoritarian control, such as permanent surveillance states or totalitarian regimes
• Better Futures Work: Respondents showed net-positive support for research on how to ensure advanced AI systems preserve human values and enable flourishing futures
• Misaligned AI Takeover: Contrary to expectations, respondents slightly favored fewer resources on uncontrolled AI systems, though this view was driven primarily by experts expecting longer timelines to AGI
• Advocacy and Policy: The sub-fields respondents identified as deserving significant additional investment include advocacy, policy and governance, and corporate advocacy efforts

What Do Experts Say About Extinction Risk Estimates?

The distribution of extinction risk estimates reveals both consensus and deep uncertainty. While the median probability of human extinction or permanent disempowerment before 2100 was 25%, the mean was higher at 34%, indicating a right-skewed distribution with some experts assigning much higher probabilities. The modal range was 20-29%, but a secondary cluster of 9 respondents (15% of the group) placed their estimate at 70% or above .

Respondents offered important caveats about what "permanent disempowerment" means. Some distinguished between extinction and disempowerment, noting that permanent disempowerment might not constitute an existential catastrophe if human values are preserved in the future. One respondent explained the distinction this way: "I'm not counting it to be human disempowerment if reflected human preferences maintain a major influence on the future, even if humans as they exist today aren't around." Another noted that they placed the chance of human extinction itself below 10%, but saw significant risk of permanent disempowerment where humanity loses control and few of our values are preserved .

Respondents

Steps to Strengthen AI Safety Research and Governance

Based on the survey findings, the AI safety community has identified several priority areas for action and resource allocation:

• Talent Recruitment and Retention: Summit attendees broadly agreed that talent is the binding constraint on AI safety progress, meaning the field's growth is limited primarily by the availability of skilled researchers rather than funding or other resources
• Focus on Aligned AI Risks: Experts emphasized that risks from aligned AI systems, such as authoritarian lock-in where powerful AI systems preserve human control but enforce oppressive values, deserve far more attention than they currently receive
• Expand Policy and Advocacy Capacity: The survey identified advocacy, policy and governance, and corporate advocacy as sub-fields that warrant significant additional investment to shape how AI systems are developed and deployed
• Investigate Automated Alignment Research: Key debates remain over how well alignment is actually progressing and whether automated alignment research, where AI systems help design safer AI systems, represents a genuine strategy or merely a hope

The survey also revealed a weak negative correlation between AGI timeline estimates and existential risk estimates: experts who expect AGI sooner tend to assign higher extinction probabilities, though the relationship is noisy and far from deterministic .

One respondent captured the tension underlying these debates: "Competition between companies and countries means we build superintelligence. And I just don't see how humans retain control of something way more capable and intelligent than us." Another countered with a more optimistic view: "I think we can avoid disempowerment," suggesting that with proper alignment methods and governance, humanity might navigate the transition to advanced AI systems successfully .

One

The survey results underscore a field in flux. While AI safety researchers broadly agree that existential risks from advanced AI warrant serious attention, they diverge sharply on which risks matter most, how much time remains before transformative AI arrives, and whether current approaches to alignment are adequate. As AI capabilities advance rapidly, these disagreements will likely shape which research directions receive funding and attention in the coming years.