There is no settled, empirically calibrated consensus probability that AI will cause human extinction. A number called “P(doom)” can still help people express concern, but it compresses a chain of uncertain assumptions into a single figure: what counts as catastrophe, when it might happen, how it could happen, and whether safeguards work. To understand AI risk—and decide what to do—readers need those assumptions, the evidence for present harms, and the policy choices alongside any probability.
What does “P(doom)” actually claim?
“P(doom)” is shorthand for someone’s subjective estimate of the probability of an AI-caused existential catastrophe. It is not a measured frequency, and the phrase does not identify one universally agreed outcome. One person may mean human extinction; another may include permanent loss of human control or a societal collapse; someone else may be thinking of a severe but recoverable catastrophe.
The time horizon matters just as much. Asking about the next five years is not the same forecast as asking about the next several decades. Nor are estimates directly comparable if one assumes that safety measures and regulation succeed while another assumes they fail. The Center for Security and Emerging Technology (CSET) notes that even “existential” and “catastrophic” can mean different levels of harm. Its analysis argues that probability alone is a poor fit for questions where we lack knowledge about the possible outcomes and the causal pathways to them (CSET, “Beyond P(doom) for AI Risk”).
That does not make a probability estimate useless. It makes its framing essential. Before comparing two numbers, ask whether they share the same outcome definition, horizon, development assumptions, and expectations about mitigation. If those differ, the apparent disagreement may partly be about what question each person answered.
#1 Best Overall
Separate harms already observed from future catastrophe scenarios
AI risk is not one category. The international expert group convened for the International Scientific Report on the Safety of Advanced AI distinguishes harms that are already occurring from prospective risks whose likelihood and pathways remain debated. It says current general-purpose AI systems do not exhibit the hypothetical catastrophic loss-of-control scenarios discussed in the report. That is an important limit on claims about “runaway” AI: the extreme scenarios are not a description of demonstrated present behavior.
| Risk category | Examples identified in the international report | What the evidence establishes |
|---|---|---|
| Observed or current harms | Biased decisions in high-stakes settings, scams, fake media, and privacy violations | These are present-day concerns, distinct from hypothetical loss of control. |
| Prospective misuse or system risks | AI-enabled cyberattacks, including hacking, and biological attacks | These are risks to assess; the report does not present them as inevitable outcomes. |
| Broader social and economic effects | Labour-market impacts | The scale and distribution of future impacts remain uncertain. |
| Extreme future scenarios | Loss of control over advanced AI systems | The report says these scenarios remain hypothetical and are not exhibited by current general-purpose AI systems. |
The report also discusses capabilities that could matter to future loss-of-control concerns, including exploiting software vulnerabilities, persuasion, automating AI research and development, and autonomous replication and adaptation. It characterizes relevant capabilities as currently limited, not as proof that a runaway scenario is underway. The report’s language is explicit: “These scenarios remain hypothetical as they are not exhibited by current general-purpose AI systems.” (International Scientific Report on the Safety of Advanced AI: Interim Report).
Keeping these categories separate avoids two opposite errors: treating every current harm as evidence that extinction is near, or dismissing current harms because the most extreme scenarios are speculative. Different risks call for different evidence, safeguards, and response timelines.
Why a precise-looking number can hide deep uncertainty
There are two distinct kinds of uncertainty in this debate. Aleatoric uncertainty is variability within a system that is understood well enough to describe—for example, uncertainty about which of several known outcomes will occur. Epistemic uncertainty is missing knowledge about the system itself: what outcomes are possible, which causal pathways matter, or how capabilities and controls will develop.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
A probability can be informative when the outcome and model are sufficiently defined. But if researchers do not know which pathways to include, a figure such as 10% can look more exact than the underlying knowledge warrants. CSET proposes considering belief and plausibility as well as probability in questions dominated by epistemic uncertainty. These are analytical alternatives it advances, not replacements that all AI researchers have adopted.
Uncertainty cuts both ways. It does not show that catastrophic risk is negligible, and the fact that a worst-case scenario is possible does not show that it is likely. A responsible account makes both limits visible rather than using uncertainty to argue for either complacency or certainty.
Rank #3
Compare risk arguments by their assumptions, not just their estimates
A useful risk argument should make its structure inspectable. When someone offers a forecast, look for the following elements:
- Outcome: Does “doom” mean extinction, permanent human disempowerment, societal collapse, or a grave but recoverable catastrophe?
- Horizon: Is the claim about the next few years, a medium-term period, or a longer future?
- Mechanism: Is the concern deliberate misuse, an accident, loss of control, concentration of power, or indirect disruption through information systems and critical services?
- Evidence: Does the argument rely on observed incidents, capability evaluations, expert elicitation, or theoretical scenarios? These forms of evidence answer different questions.
- Mitigation: Does the estimate assume effective testing, monitoring, regulation, or international coordination—and how would it change if those measures were absent or ineffective?
- Uncertainty: Is the uncertainty mainly about outcomes in a known model, or about the outcomes and causal model themselves?
These questions also help explain why expert views can differ without one side necessarily misunderstanding the other. The international report says contributors disagree about AI capabilities, risks, and mitigations; it notes that expert judgment can inform debate but cannot replace research. Its executive summary stresses that the future is uncertain and that societal and governmental decisions will help shape the trajectory (International Scientific Report, 2025).
Free tools Windows power users keep installed
One-click scans. No signup required.
Expert disagreement is real, but surveys are not a consensus meter
Disagreement is not limited to abstract probability estimates. Experts differ over future capabilities, how plausible extreme loss-of-control scenarios are, and whether technical safeguards and governance can keep pace. A 2023 UK parliamentary report records disagreement about the realism of existential-risk arguments. It also quotes Meta vice-president of AI research Joelle Pineau warning that focusing on AGI could reduce the opportunity for “rational discussions about any other outcomes.” That is Pineau’s warning in testimony, not a conclusion reached by the committee (UK House of Commons Science, Innovation and Technology Committee, “The Governance of Artificial Intelligence: Interim Report”).
Rank #4
A 2025 preprint by Severin Field offers one snapshot of views within a particular survey sample: 111 AI professionals responded, and 66.3% of respondents were academic researchers. In that sample, 77% agreed that technical AI researchers should be concerned about catastrophic risks. Field also describes clusters of beliefs, including views of AI as a controllable tool and as a potentially uncontrollable agent (Field, “Why do Experts Disagree on Existential Risk and P(doom)?”).
Those figures describe the survey, not all AI professionals. They show that concern and disagreement exist within the sample; they do not establish a global consensus or that professional groups hold uniform views. A survey result also cannot, by itself, tell us whether a forecast is well calibrated.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Governance can address risks without a settled extinction probability
Policy does not have to wait for agreement on one P(doom) figure. Governments and organizations already face practical choices about how to identify and reduce harms across risk classes. The OECD’s 2024 report organizes its analysis around ten priority benefits, ten priority risks, and ten policy priorities. Among the risks it discusses are increasingly sophisticated cyberattacks, manipulation and disinformation, fraud, critical-system incidents, concentration of power, and exacerbated inequality and poverty. It does not resolve the existential-risk debate; it supplies a wider governance frame (OECD, “Assessing Potential Future Artificial Intelligence Risks, Benefits and Policy Imperatives”).
Recommended Free Tools
Several concrete policy levers follow from that frame:
- Set adequate risk-management procedures: identify relevant risks, assess them, and apply controls proportionate to the system and its use.
- Invest in AI safety: support work that improves understanding of capabilities, failure modes, and safeguards.
- Clarify liability: make responsibilities clearer when AI systems cause harm, rather than leaving accountability uncertain.
- Consider red lines: define uses or capabilities that should be prohibited or subject to especially strict limits.
- Coordinate internationally: seek shared understandings and workable oversight where systems and risks cross borders.
The international report likewise emphasizes that societal and governmental choices affect which AI trajectory emerges. The UK parliamentary report records proposals to learn from international security frameworks, while noting diplomatic and technical challenges in agreeing on shared understandings and inspection mechanisms. Coordination is therefore a policy problem to solve, not a safeguard that can simply be assumed.
What a useful P(doom) conversation should leave you with
A probability estimate is most useful when it opens a discussion rather than ending one. Ask what harm is being forecast, on what timeline, through what mechanism, and under which assumptions about safeguards. Then ask what evidence would change the estimate and what decisions are prudent across several plausible futures.
The international report captures why that broader approach matters: “The future of AI (artificial intelligence) is uncertain, with a wide range of trajectories appearing possible even in the near future, including both very positive and very negative outcomes.” A single number cannot represent that full range—or the role that policy and social choices play in it.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →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.
