| AI Risk | Likelihood of Harm Severity Level | ||||
|---|---|---|---|---|---|
| Negligible | Minor | Substantial | Severe | Catastrophic | |
This heatmap shows the probability distribution for all 24 AI risks across five severity levels. Each cell's color intensity represents the probability (0-100%) that experts assigned to that severity level.
Experts were asked: "Assuming Business as Usual, how likely is each of the following levels of harm to occur from [Risk Name] over the next 5 years (2025-2030)?"
They allocated 100 points across the five severity levels shown below. The heatmap displays the average allocation across all experts.
Business as Usual assumes organizations & governments continue their existing practices but do not implement additional AI-specific risk mitigations.
Pragmatic Mitigations assumes organizations & governments make pragmatic and cost-effective efforts to address risks from AI.
We provided experts with detailed definitions for each harm level across 10 categories. The table below shows the full taxonomy:
| Rating | Physical Harm | Infrastructure Damage | Property Damage | Financial Loss | Environmental Damage | Toxic/Malicious Content | Differential Treatment | Human/Civil Rights | Democratic Norms | Privacy |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 Negligible |
None to minor injuries | None to brief, localized disruption | None to minor damage (<$10,000) | None to small direct financial loss (<$10,000) | No damage to minor, reversible damage | No harmful content to moderately offensive content | No differential treatment to noticeable bias | No impact to limited restrictions | No impact to minor violations | None or contained breach of basic data |
| 2 Minor |
Moderate to severe injuries, no loss of life | Moderate to significant local disruption | Moderate to significant damage ($10,000-$1M) | Moderate to large direct financial loss ($10,000-$1M) | Moderate, mostly reversible to significant, partially reversible damage | Highly offensive content to harmful misinformation affecting thousands | Clear to systemic discrimination | Significant to severe restrictions affecting thousands | Election interference and widespread misinformation campaigns | Privacy breach involving identifiable data affecting up to 1000 individuals |
| 3 Substantial |
Small-scale casualties (1-99) | Widespread regional disruption to critical service failure | Major to extensive damage ($1M-$100M) | Major to severe direct financial loss ($1M-$100M) | Major, long-term effects to severe, ecosystem-level damage | Targeted harassment to incitement to violence affecting thousands | Widespread to institutionalized discrimination | Major to widespread rights violations affecting thousands | Systematic voter suppression and undermining of democratic institutions | Privacy breach involving identifiable data affecting up to 1 million individuals |
| 4 Severe |
Mass casualties to large-scale loss of life (up to 1 million) | Regional to national infrastructure collapse | Widespread to catastrophic destruction ($100M-$10B) | Extreme direct financial loss to catastrophic loss ($100M-$10B) | Widespread ecosystem collapse to irreversible habitat damage | Large-scale coordinated attacks to mass radicalization | Systemic oppression to apartheid-like conditions | Systematic rights violations to near-total rights suppression affecting hundreds of thousands to millions | Breakdown of democratic institutions and authoritarian takeover | Privacy breach affecting >1 million individuals to societal-level privacy collapse |
| 5 Catastrophic |
>1 million deaths or existential threat to humanity | Continental to global infrastructure collapse | Extreme catastrophic to global destruction ($100B-10T+) | Economic collapse to global economic collapse ($100B-10T+) | Continental-scale devastation to global ecological collapse | Societal breakdown to global breakdown of truth and trust | Systemic oppression at global scale | Complete rights abolition to global human rights catastrophe | Global democratic collapse or authoritarian lock-in | Total surveillance state or global loss of privacy |
Note: If multiple types of harm are possible, experts were instructed to use the highest severity level. This taxonomy is based on work by CSET and Simon Myl ius.