AI existential risk is the possibility that AI could contribute to human extinction or permanently and drastically curtail humanity’s future potential. Researchers do not measure it with one definitive test or agree on a single probability. They assess different parts of the problem through capability evaluations, scenario analysis, expert elicitation and probabilistic forecasting—each with important limits.
What makes an AI risk “existential”?
The term refers to outcomes at the level of humanity’s continued existence or long-term potential. It is not a synonym for every serious harm caused or enabled by AI. Fraud, disinformation, bias, cyber incidents and labor-market disruption can be damaging or systemic without, by themselves, amounting to existential risks.
Researchers may discuss human extinction, or a lasting loss of human agency and ability to shape the future. Those are grave but distinct possible outcomes, and neither should be casually substituted for broader terms such as “catastrophic harm.” A forecast about catastrophe only answers a question about existential risk if the event being forecast is actually defined in existential terms.
How do researchers assess the possibility?
No instrument can directly read off the probability of an existential catastrophe. Assessment instead combines evidence about what systems can do and how harms might arise with judgments about likelihood and severity. Different methods answer different questions:
#1 Best Overall
Capability and risk evaluations
Researchers test AI systems for capabilities, behaviors and failure modes that could matter for safety. Such evaluations provide evidence about tested systems under tested conditions; how representative those tests are of real-world deployment is a separate question. A test result is not, by itself, a long-term probability of catastrophe.
Scenario analysis
Scenarios describe a possible causal route from capabilities and incentives to harm. For example, the preprint Is Power-Seeking AI an Existential Risk? presents a conditional argument in which powerful agentic systems, incentives to deploy them, difficulty building aligned systems and power-seeking could culminate in human disempowerment. The argument’s premises can be examined, but the proposed chain is not a demonstrated sequence of events.
Structured expert elicitation
Studies can ask experts to assess specified risks under specified conditions. A Delphi study is one way of collecting and comparing judgments across rounds. Its results apply to the risks, thresholds and assumptions that the study actually asked about; they do not automatically estimate extinction risk.
Rank #2
Probabilistic forecasting
Forecasting asks respondents to assign probabilities to defined events by specified dates. The Longitudinal Expert AI Panel’s Wave 9 displays group-median forecasts for global AI-related catastrophe under slow, moderate and rapid AI-progress scenarios. Scenario-dependent medians are conditional judgments, not a universal probability for existential risk.
Recommended Free Tools
Studying disagreement
Comparing estimates can reveal where judgments diverge and which assumptions may drive the gap. The Forecasting Research Institute’s Existential Risk Persuasion Tournament (XPT) and its follow-on Roots of Disagreement project examine such differences. Disagreement is relevant evidence about uncertainty; it is not resolved simply by averaging estimates from groups with different assumptions or recruitment methods.
What do prominent estimates actually measure?
The figures below come from studies with different endpoints and methods. They should not be read as competing measurements of one common quantity.
| Study | Population and method | Reported result | What it does—and does not—mean |
|---|---|---|---|
| MIT FutureTech and University of Queensland Delphi study, 2026; reported by MIT Sloan | 272 international experts across 37 countries; expert Delphi elicitation covering 24 risk categories | Under the study’s business-as-usual scenario, experts rated 18 of 24 categories as more than 10% likely to cause a catastrophic outcome. | The study defined catastrophe as more than one million deaths, more than $100 billion in financial losses, or comparable harms. The figure is not an estimate that AI has an 18-in-24 chance of causing catastrophe, nor an extinction probability. MIT Sloan quoted Principal Research Scientist Neil Thompson calling the breadth of the assessments “incredibly worrisome”; that comment concerns the study’s catastrophe thresholds. |
| Roots of Disagreement on AI Risk, Forecasting Research Institute, 2024 | 11 participants recruited as “AI skeptics” and 11 as “AI concerned”; forecasts were recorded at the beginning and end of the project. | For AI existential catastrophe by 2100, the skeptic group’s median moved from 0.10% to 0.12%; the concerned group’s median moved from 25% to 20%. | The project found little convergence between these intentionally opposing, small groups. These are not population-representative estimates or the median view of AI researchers. |
| Existential Risk Persuasion Tournament, Forecasting Research Institute, 2023 | 169 forecasters in a multistage tournament concerning existential risks over the next century | The project documented substantial disagreement and limited convergence. | The participant count describes the tournament, not the size of a consensus or a single probability for AI extinction. |
| Longitudinal Expert AI Panel, Wave 9 | Group-median forecasts shown under slow, moderate and rapid progress scenarios | Forecasts concern global AI-related catastrophe; the panel page presents scenario-dependent medians. | “Global AI-related catastrophe” should not be treated as synonymous with extinction without checking the event definition. |
These comparisons show why a probability needs its full context. Before interpreting an estimate, check the event definition, time horizon, respondent population, sample size, elicitation method and assumed scenario or mitigations. A probability of “catastrophic outcomes” under a study-specific threshold cannot be directly compared with a probability of human extinction.
Why do researchers disagree?
Long-term estimates depend on more than observed system performance. People may differ over how quickly capabilities will advance, whether future systems will act as agents, how deployment incentives will shape use, whether technical safeguards will work, and how much time there will be to respond. Changing the outcome definition or the forecast horizon also changes the question.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11The Roots of Disagreement project illustrates the scale of divergence: its two recruited groups ended with sharply different median forecasts for the same stated outcome and date. Those results reveal disagreement within that project, not a representative distribution of all experts’ beliefs. The informal phrase “p(doom)” has the same interpretive problem: speakers may mean different outcomes, probabilities and time horizons by it.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What are the limits of current assessment?
The International scientific report on the safety of advanced AI: interim report, hosted by the UK Government, describes assessment of general-purpose AI as an unsettled area of science. It notes limited understanding of model internals, difficulty assessing downstream impacts across varied uses, and a lack of rigorous, comprehensive assessment methodologies. Technical methods have limitations and cannot provide strong assurances against most harms.
“At present, computer scientists are unable to give guarantees of the form ‘System X will not do Y’ about general-purpose AI (artificial intelligence) systems.”
This is a statement in the interim report’s discussion of assessment and assurance, not a finding that a particular catastrophic outcome will occur. Tests can inform judgments, but they cannot establish a guarantee that covers every system, context or future use.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Best Value
How should the international AI safety assessment be read?
The International AI Safety Report 2026 synthesizes research on general-purpose AI capabilities, risks and risk management, with a focus on emerging risks associated with frontier capabilities. The report was authored by more than 100 experts and backed by over 30 countries and international organizations. It draws on more specific scenarios and forecasts, including work from the OECD and Forecasting Research Institute.
Its treatment of loss of control is a debated possibility, not an established outcome. Keep three kinds of claims distinct when reading the report or any risk assessment:
- Evidence about current systems: what evaluations have observed under their test conditions.
- Arguments about possible future mechanisms: conditional accounts of how capabilities, incentives and safeguards might interact.
- Forecasts: subjective probabilities assigned to defined events, over stated horizons and under assumptions.
Those forms of evidence can inform one another, but one does not prove the others. In particular, a present-day capability result is not proof of a future loss-of-control pathway, and a forecast about catastrophic harm is not automatically a forecast about extinction.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
