Is existential risk the same as global catastrophic risk?

No. A global catastrophe can be devastating yet recoverable; existential risk concerns outcomes that end or permanently curtail humanity's future potential.

Do experts agree on the probability of AI-related existential outcomes?

No. There is substantial disagreement and uncertainty about capabilities, behavior, deployment, and safeguards.

Does discussing the risk mean predicting disaster?

No. Considering a scenario and evaluating possible prevention are different from asserting that it will happen.

AI Glossary / Governance, ethics and safety / Intermediate

Existential risk

Definition

Existential risk is the possibility of an outcome that destroys humanity or permanently and drastically limits humanity's future potential. The concept concerns the scale and lasting character of the possible harm, not a particular probability or date. In AI discussions, it refers to hypothesized pathways by which advanced systems could contribute to such outcomes; those pathways and their likelihood remain subjects of substantial uncertainty and disagreement.

What it means

An existential threat in this technical sense is different from a threat to one business's survival. A severe recession, a major service outage, and even a global catastrophe are not automatically existential if humanity can recover without permanent loss of its long-term potential. The category directs attention to outcomes that cannot be repaired through ordinary recovery. It does not imply that such outcomes are inevitable, that any current model demonstrates them, or that all AI risks belong in this category. Present harms and extreme future possibilities require distinct evidence and can both deserve attention.

How it works

Risk analysis describes a possible pathway, identifies the capabilities and conditions it would require, and examines evidence for each step. In AI, discussions may consider misuse, loss of effective human control, or interactions with other powerful technologies and institutions. A scenario is not a forecast simply because it can be described. Analysts should separate demonstrated capabilities from extrapolations, identify barriers and countermeasures, and state uncertainty about deployment conditions. Extreme consequences make preparation relevant, but uncertainty also makes precise probability claims difficult to justify. Good analysis keeps those two points visible together.

Three practical examples

A research group studies whether proposed safeguards remain effective as AI systems gain more autonomy, clearly distinguishing controlled tests from claims about global outcomes.
A policymaking exercise examines a hypothetical catastrophic failure pathway and identifies assumptions that would need evidence before treating it as a credible forecast.
A technology organization sets escalation rules for unexpected high-consequence capabilities while continuing to address immediate reliability and misuse concerns. These are risk-governance activities, not examples of an existential event having occurred.

Think of it this way

Think of protecting the only remaining copy of an irreplaceable archive. Losing a working copy is serious but recoverable; losing every recoverable copy is a fundamentally different consequence. The analogy is limited because humanity's future is not a single object and its value cannot be reduced to a file. It illustrates the importance of irreversibility without assigning a probability to a particular AI scenario.

Worked example

Consider a fictional risk-review worksheet with three statements. First: an AI service could suffer a week-long outage. Second: a failure across several important services could produce widespread disruption, followed by recovery. Third: a hypothetical sequence could permanently prevent humanity from recovering control over its future. The first is an operational risk, the second may be a catastrophic risk depending on scale, and the third is a candidate existential-risk scenario because of its permanence and scope. Classification does not establish plausibility. For the third statement, reviewers must specify the causal steps, necessary capabilities, deployment access, failure of countermeasures, and reasons recovery would be impossible. If those elements are unsupported, the worksheet should say so. A useful worked analysis clarifies the hypothesis instead of attaching an invented percentage to it.

Illustrative case study

Fictional case: Cedar Research creates a review process for a planned autonomous research assistant. Some participants raise extreme long-term concerns, while others focus on near-term errors. The team separates observed behavior, plausible deployment hazards, and speculative future pathways. It assigns evidence-gathering tasks and defines limits on the proposed system's authority without describing the pilot as an existential threat. The invented case illustrates disciplined discussion across different risk scales, not a claim that the organization has measured or eliminated existential risk.

Common misconceptions and mistakes

Existential risk does not mean any risk that feels frightening or important. It also does not mean extinction alone under every definition: permanent destruction of humanity's future potential can be included. A high-severity scenario is not automatically a high-probability scenario. Conversely, uncertainty is not proof of zero risk. A system displaying an unwanted behavior in a test does not by itself establish a pathway to humanity-wide catastrophe. Another mistake is treating the debate as a reason to disregard existing harms or to present speculative probabilities as measured facts.

Business use

For most organizations, connect this concept to the actual capabilities and authority of the systems being developed or deployed. Use credible evidence, clear escalation responsibilities, and proportionate controls for consequential actions. When a project could materially affect broader safety, seek relevant expertise and examine assumptions about control, monitoring, and reversibility. Keep operational risk management active rather than replacing it with abstract debate. Communicate what has been observed, what is hypothesized, and what remains unknown. Avoid using dramatic terminology as marketing or as a substitute for a specific, reviewable risk argument.

Check your understanding

A hypothetical AI failure causes severe global disruption, but humanity recovers without a permanent loss of its long-term potential. Is it necessarily an existential catastrophe?

Answer: No. It may be a global catastrophe, but the existential category requires the relevant permanent and humanity-wide consequence.

Related terms

Uncertainty
Resilience
global catastrophic risk
extinction risk
AI alignment
loss of control
misuse
scenario analysis
irreversible harm

Learn next

Take one high-consequence AI scenario and divide it into causal steps. Mark each step as observed, inferred, or speculative, and identify what evidence would change the assessment. Keep severity, likelihood, and recoverability as separate questions.

Sources

Nick Bostrom, Existential Risks: Analyzing Human Extinction Scenarios and Related Hazards, original conceptual paper: https://nickbostrom.com/papers/existential-risks/ International AI Safety Report 2026, loss-of-control evidence and uncertainty: https://internationalaisafetyreport.org/sites/default/files/2026-02/international-ai-safety-report-2026.pdf

Previous: Existential graphs Next: Exogenous variable