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 combine capability evaluations, scenarios, expert judgments and forecasts—each of which addresses a different part of the question and carries important limitations.
What “AI existential risk” means
The term describes humanity-level outcomes: human extinction, or a permanent and drastic loss of humanity’s ability to shape its future. “Existential” marks the severity and potentially irreversible scope of the outcome; it is not a synonym for every serious problem involving AI.
As an Amazon Associate I earn from qualifying purchases.
Fraud, disinformation, biased decisions, cyber incidents and job disruption can cause substantial harm. They are not automatically existential risks. Nor does a forecast of a “catastrophic” outcome necessarily mean extinction: studies may use that word for outcomes defined by deaths, financial losses or other severe impacts.
The 2026 International AI Safety Report focuses on general-purpose AI and emerging risks associated with frontier capabilities. It synthesizes research on capabilities, risks and risk management, drawing on scenarios and forecasts from organizations including the OECD and Forecasting Research Institute. More than 100 experts authored the report, which was backed by over 30 countries and international organizations.
#1 Best Overall
How researchers assess possible existential risk
No single instrument can read off the probability of an existential catastrophe. Researchers use different approaches to examine capabilities, possible causal pathways and judgments about likelihood and severity.
Capability and risk evaluations
Evaluations test what systems can do and how they behave, including whether they show capabilities or failure modes that could matter in dangerous settings. A test result is evidence about performance under the test conditions—not a guarantee about behavior after deployment. Conclusions depend on how well the tests represent real uses, tools, users and operating environments.
Scenario analysis
Scenarios lay out a proposed route from system capabilities and human incentives to harmful outcomes. For example, one research preprint 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. This is an argument whose premises and links need evaluation, not a demonstrated chain of events or proof that current systems are on that path.
Rank #2
Structured expert elicitation
In a structured elicitation, researchers ask experts to assess defined risks under specified conditions. A Delphi study by MIT FutureTech and the University of Queensland considered 24 risk categories and compared “business as usual” with “pragmatic mitigations.” Its results apply to the study’s categories, thresholds and scenarios—not to an unspecified probability of human extinction.
Probabilistic forecasting
Forecasters estimate the chance of clearly defined events by a stated date, sometimes under explicit scenarios. The Longitudinal Expert AI Panel presents group-median forecasts for global AI-related catastrophe across slow, moderate and rapid AI-progress scenarios. A median summarizes a group’s estimates for that question and scenario; it is not a direct measurement of the underlying risk.
Studying disagreement
Researchers can compare estimates and investigate which assumptions or observations account for differences. The Forecasting Research Institute’s Existential Risk Persuasion Tournament (XPT) involved 169 forecasters in a multi-stage tournament about existential risks over the next century. Its follow-on Roots of Disagreement project examined the beliefs and possible cruxes behind differences in views about AI risk.
What the reported numbers do—and do not—tell you
| Study and result | What was assessed | What the number does not mean |
|---|---|---|
| MIT FutureTech and University of Queensland Delphi study, 2026: 18 of 24 risk categories | Under the study’s business-as-usual scenario, experts rated 18 categories as more than 10% likely to cause a catastrophic outcome. The study defined catastrophic as more than one million deaths, more than $100 billion in financial losses, or comparable harms. | This is not a 10% probability of extinction or existential catastrophe. The threshold includes outcomes that can be catastrophic without being existential. |
| MIT FutureTech and University of Queensland Delphi study, 2026: 272 experts across 37 countries | The press release gives these figures to describe the international expert sample. | They describe who took part, not how likely a risk is. |
| Forecasting Research Institute, Roots of Disagreement, 2024: 0.10% to 0.12% | The median forecast of the 11-person “AI skeptic” group for AI existential catastrophe by 2100, at the start and end of the project, respectively. | This is the estimate of a small, intentionally recruited group—not a representative poll of AI researchers. |
| Forecasting Research Institute, Roots of Disagreement, 2024: 25% to 20% | The median forecast of the 11-person “AI concerned” group for the same event and 2100 horizon, at the start and end of the project, respectively. | This, too, is a small, intentionally recruited group’s forecast, not a consensus probability. The study found that beliefs did not substantially converge. |
The Delphi figures and Roots of Disagreement forecasts are not competing estimates of the same quantity. They use different methods and outcomes: the Delphi study assessed study-defined catastrophic harms across risk categories, while Roots of Disagreement asked small groups to forecast AI existential catastrophe by 2100. Treating them as interchangeable would confuse severe harm with an existential endpoint.
More generally, an estimate is interpretable only alongside its event definition, time horizon, respondent population, elicitation method and assumptions about the scenario and mitigation. “P(doom)” is informal shorthand, not a standard event definition: different speakers may mean different outcomes, horizons or scenarios by it.
Why forecasts can differ—and why they remain uncertain
Forecasts of unprecedented, long-term events are difficult to validate: the events have not occurred often enough to provide a reliable track record. Judgments can also change with assumptions about how quickly capabilities develop, how systems are deployed, what safeguards are used and which causal pathways are plausible.
In Roots of Disagreement, the median forecast for AI existential catastrophe by 2100 shifted from 0.10% to 0.12% among the “AI skeptic” participants and from 25% to 20% among the “AI concerned” participants. The gap remained large. Those figures document disagreement and limited belief revision within that project; they should not be presented as the median view of researchers or the public.
For any forecast, distinguish the group statistic from a consensus, and the event being forecast from broader labels such as “AI risk.” A probability attached to a study-specific definition of catastrophe does not become an extinction probability simply because the possible harms are severe.
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 matchWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallWhat current assessment can establish
The UK Government-hosted international interim report describes assessment of general-purpose AI as an unsettled area of science. It points to limited understanding of model internals, the difficulty of assessing downstream effects across varied uses, and the lack of rigorous, comprehensive assessment methodologies. Current technical methods have limitations and cannot provide strong assurances against most harms.
Best Value
The report states: “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 caution about the limits of assurance, not evidence that a particular catastrophe will occur.
Keep three kinds of claim separate when weighing the issue: observations about what present systems can do; conditional arguments about how future capabilities and incentives might create dangerous pathways; and subjective forecasts about long-term outcomes. Evidence in one category does not, by itself, settle the others. Researchers have not established a reliable single probability for existential risk, and loss of control remains a debated possibility rather than an established outcome.
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.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minute
