iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
There is no settled scientific probability that AI will cause human extinction. The best way to assess a claim about an “AI apocalypse” is to pin down the harm and time horizon, inspect the evidence for each step in the proposed scenario, and keep observed system behavior separate from forecasts and speculation. Expert surveys can show what respondents believe; they do not measure how often catastrophe occurs or establish its true probability.
First define what “AI apocalypse” means
Apocalypse is not a precise outcome. A claim may refer to human extinction, permanent and severe loss of human control, a global catastrophe that is nevertheless recoverable, or a different form of harm. These outcomes are not interchangeable. A forecast about severe disempowerment, for example, should not be reported as a forecast of extinction unless the source explicitly says so.
Define the time window too. “A 5% chance” is incomplete without saying what event is being forecast and whether the period is a few years, a century, or an indefinite future. Estimates with different outcomes or horizons cannot be compared as if they answer the same question.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →What do expert surveys say about extinction risk?
AI Impacts reported results in 2024 from a 2023 survey of 2,778 researchers who had published in top-tier AI venues. The survey asked several differently worded questions about future AI advances and catastrophic outcomes. For one formulation, 655 respondents forecast the chance that future AI advances would cause human extinction or similarly permanent and severe disempowerment within 100 years.
#1 Best Overall
| Survey result | What it describes |
|---|---|
| 5% median; 14.4% mean | Respondents’ forecasts for the specific 100-year question above (n=655), as reported by AI Impacts in 2024. These are elicited judgments, not measured event frequencies or an agreed probability. |
| 41.2%–51.4% | Across three differently framed questions, the share of respondents who assigned at least a 10% chance to extinction or severe disempowerment outcomes, as reported by AI Impacts in 2024. The range reflects different question formulations. |
Do not compress these findings into “AI has a 5% chance of extinction.” The 5% is the median response to one particular question, combining extinction with a specified kind of severe disempowerment and using a 100-year horizon. The mean is different because it is a different way of summarizing the distribution of answers. The 41.2%–51.4% range describes the proportion crossing a threshold across three formulations; it is not another estimate of the chance of catastrophe.
The survey is evidence about the views of its respondents. It shows both that some researchers assign substantial probability to severe outcomes and that judgments vary considerably. It does not establish that any one estimate is objectively correct. Wording, the population asked, and the way answers are aggregated all matter.
What current evidence says about loss of control
The 2025 International AI Safety Report describes potential loss-of-control outcomes ranging from human marginalisation to extinction. It calls the probability of loss of control contested and says that disagreement likely reflects the difficulty of interpreting and extrapolating from available evidence. The report does not give a single settled probability for AI-driven extinction.
It identifies important evidence gaps: empirical studies of capabilities and trends; threat analysis; the extent of misalignment in current systems; how alignment may change as capabilities increase; which capabilities would be necessary for loss of control; and realistic models of how systems might undermine human control. Passive loss-of-control scenarios are particularly understudied. The report also says that pathways from active or passive loss of control to catastrophe have so far been laid out only in broad strokes.
That distinction matters when evaluating a specific scenario. A description of a possible chain of events is not evidence that every link in the chain has occurred, is likely to occur, or will lead to the stated outcome. Ask which steps have been observed, which have been tested under controlled conditions, which are modeled, and which are assumptions about future systems.
Loss of control is not the only route to catastrophic harm discussed in the report. Malicious use and systemic risks could cause severe consequences without an AI system independently escaping human control. Conversely, uncertainty about loss-of-control probability is not evidence that the probability is zero.
Rank #3
Separate observations, forecasts, and scenarios
These three kinds of claims answer different questions. A strong assessment makes clear which kind of evidence supports each part of its argument.
Recommended Free Tools
Observed behavior and evaluations
Observations and controlled evaluations can establish what a system did in particular settings, under particular conditions. They can help test claims about present capabilities or misalignment. They do not, by themselves, establish how future systems will behave, how capabilities will develop, or whether a particular behavior will produce catastrophe.
Expert surveys and forecasting
Expert elicitation records judgments from a defined group in response to stated questions. A forecasting tournament can also collect probability judgments, often from different kinds of forecasters and for outcomes with specific resolution dates. Both methods make judgments more explicit, but their results remain tied to the wording, participants, aggregation method, outcome definition, and horizon.
Rank #4
A forecasting tournament paper, for example, describes anonymous probability judgments by subject-matter experts and superforecasters on existential risks. Those estimates should not be put beside an expert-survey figure without first checking that the questions, populations, outcomes, horizons, and aggregation methods are comparable. If a forecast has a resolution date and a clearly defined outcome, it may later be possible to score it; that does not make it interchangeable with a survey asking a different question.
Scenario analysis and causal arguments
A scenario explains how harm might happen. It can reveal dependencies, vulnerabilities, and assumptions, but a coherent narrative is not a probability estimate. For every claimed transition—from a capability to a behavior, from that behavior to loss of control, or from loss of control to a specified harm—ask what supports the link and what could prevent it.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →A practical checklist for evaluating a claim
- Specify the outcome. Write down whether the claim concerns extinction, permanent loss of human control, a large but recoverable catastrophe, or another harm. Keep the outcome consistent throughout the argument.
- Set the horizon. Record the period covered by the claim. Do not compare a forecast for the next few years with one for the next century as though the numbers had the same meaning.
- Identify the assumed system. Ask what kind of AI is being discussed and what capabilities the scenario requires. Separate demonstrated capabilities from projections, and ask what evidence supports the projected trajectory.
- Map the causal pathway. List the steps between the assumed system and the claimed harm. Label each step as observed, evaluated, modeled, or speculative. Look for missing dependencies and for mechanisms that could interrupt the chain.
- Name the evidence type. Is the claim based on observed behavior, controlled evaluations, incident data, an expert survey, a forecast tournament, scenario analysis, or argument? Ask what that method can establish—and what it cannot.
- Examine uncertainty and alternatives. What evidence would change the claimant’s view? Have plausible counter-scenarios and mitigating factors been considered? Are ordinary, beneficial, and catastrophic possibilities treated consistently rather than selecting only evidence that supports one outcome?
- Separate probability, severity, and action. “How likely?” is different from “How severe?” and both are different from “What preparation is worthwhile?” A low-probability, high-severity argument should make its assumptions visible and compare the costs of preparation with the costs of preparing unnecessarily.
How to compare two estimates fairly
Before interpreting a difference between two numbers, compare the questions behind them. If the definitions do not align, the apparent disagreement may come from method or framing rather than a direct difference in beliefs about the same event.
- Outcome: extinction, severe disempowerment, or another harm?
- Horizon: What period does the forecast cover?
- Question wording: What exactly were respondents asked?
- Forecasters: Who took part, and what relevant expertise did they have?
- Aggregation: Is the figure a median, mean, threshold share, or another summary?
- Causal support: What evidence supports each link in the proposed pathway?
- Uncertainty: How are uncertainty, counter-scenarios, and possible mitigations handled?
- Resolution: Is the outcome defined in a way that can later be checked or scored?
The AI Impacts survey and forecasting tournaments have different designs. Their probability figures should not be placed side by side without accounting for those differences. A methods review published in 2020 likewise argues for a more critical approach to quantifying existential hazards and for awareness of the range of methods available. A number is most informative when accompanied by its definition, method, assumptions, and limitations.
How to interpret disagreement without false balance
Disagreement is important evidence about the state of judgment, not proof that all positions are equally well supported. The 2025 International AI Safety Report links disagreement about loss-of-control probability to limits in the available evidence and the difficulty of extrapolation. That is a reason to examine the underlying claims carefully, not to assume either that catastrophe is imminent or that it cannot happen.
When a claim is uncertain, inspect the quality of its evidence and the plausibility of its assumptions. The same standard should apply to reassuring claims: an absence of established probability is not a demonstration of safety. Independent evaluation can be valuable, particularly because companies may face incentives that affect how they assess their own systems.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Further reading
- The 2025 International AI Safety Report provides an international assessment of loss-of-control risks, evidence gaps, and other pathways to catastrophic harm.
- AI Impacts’ 2024 publication of its 2023 researcher survey provides the question-specific results and explains how respondents’ judgments varied with wording.
- A 2020 literature review examines methods for quantifying existential hazards and the limitations of quantified claims.
- Forecasting Research Institute research is relevant to differences between domain experts and generalist forecasters.
- If Anyone Builds It, Everyone Dies by Eliezer Yudkowsky and Nate Soares presents a forceful argument about AI risk; it is one position in the debate, not a neutral textbook.
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.
