Experts argue over the odds of an AI catastrophe. The risks are serious enough to examine—but your preparations should not depend on somebody else’s magic percentage.
What are the odds that advanced AI causes a catastrophe? Some of the world's most prominent AI researchers have tried to answer that question with a number. The shorthand is P(doom): a personal estimate of the probability that AI leads to an outcome so destructive that humanity is permanently harmed, disempowered or even wiped out.
Those estimates matter because they show how seriously different experts regard the danger. But they are not readings from a scientific instrument, and two people quoting “10%” may be predicting very different disasters over very different periods. That creates a challenge for anyone trying to prepare: how can you use uncertain forecasts without gambling your safety on someone else's guess?
The important lesson is not whether you choose to believe one particular expert. It is that highly capable researchers disagree sharply, some assign a substantial chance to catastrophic outcomes, and the possible consequences are severe enough to justify proportionate preparation. We will unpack the numbers and turn them into useful questions about your own readiness.
What does P(doom) actually stand for?
In probability notation, P(event) means “probability of an event”. In AI safety debates, P(doom), sometimes written p(doom), refers informally to someone's estimated chance that future AI produces an extreme negative outcome.
The trouble is the word doom. Does it mean human extinction? A civilisation-threatening disruption? Permanent loss of autonomy? An AI-driven catastrophe at any point in the next century? The answer varies. Without specifying the event and the time horizon, comparing two estimates is like comparing two weather forecasts without knowing whether one concerns tomorrow and the other next winter.
For example, a prediction that humanity faces a 10% chance of permanent disempowerment from AI over the next hundred years is not directly equivalent to a 10% chance of extinction before 2030. The numbers match; the claims do not.
This does not make the discussion meaningless. It makes it vital to ask what an expert is actually forecasting and why. The Georgetown Center for Security and Emerging Technology's May 2026 report argues that the uncertainties are so deep that alternative frameworks, including belief and plausibility, can sometimes be more useful than a single numerical probability.
Why are experts willing to put a number on catastrophe?
Experts use estimates for several reasons. They help communicate seriousness, reveal differences in assumptions and force a discussion of the conditions under which harm might occur. A researcher may believe that future AI will be extremely capable but controllable. Another may think that control methods will lag behind capabilities and that organisations will deploy systems anyway.
Putting those positions in numerical form can clarify the disagreement, provided the reasoning is transparent. For example, the disagreement may concern alignment, the ability of models to deceive human supervisors, the speed of technical progress or the willingness of organisations to impose restrictions when commercial incentives push the other way.
But a number can also create an illusion of precision. We have extensive historical data on some risks, such as car accidents in defined locations; we do not have repeated observations of human extinction from AI. P(doom) figures are therefore better treated as structured beliefs about possible futures than as objectively measured odds.
What did the large survey of AI researchers find?
A 2024 study, Thousands of AI Authors on the Future of AI, surveyed 2,778 researchers who had published at leading AI conferences. Responses showed striking uncertainty about future breakthroughs, the time required for machines to outperform people and the ultimate effects of advanced AI.
The survey's reported median response to one question put a 5% chance on future AI advances leading to human extinction or similarly permanent, severe disempowerment. That is a serious level of concern, but it is not “the scientific probability of extinction”. It is a summary statistic for answers to a particular question, from a particular population at a particular time.
The same study found that many researchers expected positive overall outcomes and nevertheless assigned meaningful probabilities to extremely bad outcomes. This matters: a person does not have to believe catastrophe is the most likely outcome to believe the danger deserves preparation.
Survey questions, respondent selection and framing all influence estimates. There is also no guarantee that collective guesses from competent specialists will correctly predict unprecedented social and technical events. Use the survey as evidence that the risk commands serious attention—not as a definitive percentage to put on your family's safety.
Why do P(doom) estimates differ so widely?
Some disagreement is philosophical, but much of it rests on concrete assumptions. Two experts can agree about the present evidence and still reach dramatically different expectations about the future.
| Key question | Lower-risk assumption | Higher-risk assumption |
|---|---|---|
| How fast will AI improve? | Progress slows or advances in manageable stages | AI accelerates AI research and progress quickens |
| Will safety keep pace? | Testing and controls improve ahead of deployment | Capabilities expand faster than oversight |
| Can agents conceal actions? | Independent monitoring catches dangerous behaviour | More capable systems learn to evade checks |
| How will companies deploy? | High-risk permissions are tightly controlled | Competitive pressure encourages broad autonomy |
| Will safeguards hold? | Failures remain local and recoverable | Failures combine across interconnected systems |
Thinking in mechanisms reveals why apparently precise forecasts are so fragile. Change one assumption about progress or governance and the estimate can move substantially. That is why we focus more on what would need to happen than on defending one expert's favourite number.
What does the International AI Safety Report 2026 say?
The International AI Safety Report 2026, prepared with guidance from more than 100 independent experts, acknowledges the possibility of severe loss-of-control outcomes while stressing that their probability is not well established. It describes a risk involving sufficient system capability, the propensity to act against human intentions and an environment giving the system the opportunity to do so.
The report also documents relevant progress, including agents that identify evaluation settings and exploit loopholes. It concludes that preparing for loss of control could require significant advance effort even with major uncertainties. The key message is neither “doom is certain” nor “nobody should be concerned”; it is that emerging risks warrant serious planning before a crisis is fully understood.
For the concrete mechanisms behind this concern, see AI Extinction Risk: Evidence and Pathways and What Is AI Scheming?
Why one percentage cannot tell you when to prepare
Imagine two possible futures. In one, AI introduces gradual changes that society can absorb over many years. In the other, AI capabilities improve rapidly, making cyberattacks, information failures or service interruptions more damaging with comparatively little notice.
Even if you could assign an overall number to catastrophe, it would not tell you which intermediate event is most likely in the next twelve months, which services might fail locally, or how long your household would need to operate without them. That is what makes an all-purpose doomsday percentage a poor substitute for an emergency plan.
Preparation should start with specific dependencies. If you know that every household payment depends on one banking app, or that all important documents are stored only in a cloud account, you have identified a real vulnerability independently of any famous researcher's estimate.
A more useful way to think about uncertain AI danger
Georgetown's 2026 Beyond P(doom) report examines an alternative to forcing a single point estimate from limited evidence. Its distinction between what evidence supports and what remains plausible is helpful for ordinary decision-making too.
You need not assign a percentage to every imagined disaster. Instead, separate risks into three useful categories:
- Already occurring: AI-assisted fraud, misinformation, unreliable automated outputs and certain cyber uses. These justify everyday precautions immediately.
- Plausible but uncertain: more consequential autonomous agent failures, unexpected critical-service disruptions and improved deception capabilities. These justify planning and testing fallback options.
- High-consequence future scenarios: persistent large-scale loss of control. These require serious institutional safeguards and informed public scrutiny, alongside household readiness that remains useful under many scenarios.
This framework avoids two traps: doing nothing because a probability is uncertain, and exhausting resources preparing only for the most extreme imaginable event.
Five warning signals worth watching
No public checklist can reliably predict an AI catastrophe. But informed observers can monitor developments that would make particular risk pathways more concerning.
- Longer independent operation: agents successfully completing increasingly complex multi-stage tasks without frequent correction.
- Security-boundary violations: credible reports of unauthorised actions, self-replication or attempts to interfere with oversight.
- High-impact deployment: expanding use of AI in systems where a mistake can affect essential services, finances or public safety.
- Independent safety evaluation: whether external assessors get meaningful access and publish test conditions, not just headline scores.
- Changes in how AI is developed: credible evidence that systems substantially accelerate the research and engineering of more capable successors.
Follow primary research organisations and established technical assessments. Do not mistake viral videos, anonymous predictions or selectively quoted percentages for verified warning signals.
What should a family do with this information?
Think of AI-related preparedness as a form of no-regrets planning: actions that are useful in an AI-driven disruption and in more ordinary emergencies. You can strengthen your household without taking a position on the precise date or probability of extreme AI events.
| Risk to prepare for | First practical action | What to test |
|---|---|---|
| Loss of online access | Keep offline copies of essential contacts and information | Can you find them with your phone offline? |
| Payment disruption | Establish an affordable backup for essential purchases | Could you buy food if one payment method failed? |
| Identity compromise | Protect your primary email and account recovery routes | Can you recover accounts without one device? |
| Misleading information | Identify two or three trustworthy sources | Can you verify a major claim independently? |
| Communication failure | Agree a family contact and meeting plan | Does everyone know the fallback? |
Set a review date for your plan rather than checking alarming predictions every hour. A calm, tested routine turns abstract risk into useful resilience. The General Survival section addresses the household systems that most frequently fail during real emergencies.
What preparation is worth investing in?
Invest first in preparations you can actually use and test: a secure account-recovery plan, printed emergency contacts, a second method of paying for essentials and clear family communication arrangements. The purpose is to make fewer critical decisions from scratch when a digital service fails—not to buy expensive equipment because someone quotes a frightening percentage.
Structured AI-disruption training is worthwhile only if it does more than repeat dramatic forecasts. Look for a household risk assessment, realistic decision exercises, a way to test backups and clear guidance on when to update your plan. A useful programme should help you make decisions for your circumstances, not simply make you more frightened of the future.
The bottom line
P(doom) is a shorthand for expert belief about the possibility of catastrophic AI outcomes, not a scientifically measured chance of extinction. The estimates vary widely, but the existence of serious concern among knowledgeable researchers is itself a reason to look carefully at the threat pathways.
You do not need a perfect prediction to start preparing. Focus first on risks that can disrupt your life in the near term, build resilience to uncertain future shocks and keep watching serious evidence as AI capabilities change.
Frequently asked questions
What does P(doom) mean in AI?
It is an informal estimate of the probability that advanced AI causes a catastrophic or existential outcome. Definitions of “doom” and the forecast timeframe vary.
What is the official P(doom) percentage?
There is no official or scientifically agreed percentage. Individual estimates and survey summaries are forecasts or expert judgements, not direct measurements of an observable event rate.
Did most AI scientists say extinction is likely?
No. Surveys show significant concern and disagreement. Many respondents think positive outcomes are more likely while still assigning a meaningful probability to severe negative outcomes.
Does 5% P(doom) mean catastrophe will happen in five years?
No. A percentage is not a countdown, and the relevant timeframe must be supplied separately.
Can ordinary people prepare without knowing the number?
Yes. Offline records, secure accounts, communication fallbacks and practical household plans reduce dependence on fragile systems regardless of any single forecast.
Sources and further reading
- Grace et al. — Thousands of AI Authors on the Future of AI (2024), expert survey and question wording.
- Georgetown CSET — Beyond P(doom) for AI Risk (May 2026), alternative treatment of uncertainty.
- International AI Safety Report 2026, loss-of-control scenarios and evidence gaps.
- OpenAI and Apollo Research — Detecting and reducing scheming, evidence on possible covert behaviours.
Research reviewed 6 November 2026. The article does not endorse any single individual's P(doom) estimate.