The Numbers
31 citable stats on AI risk — expert odds, public opinion, incidents, lab grades and the pace of the frontier. Every number links to its source; hover any card to copy a ready-made citation.
Where insiders put the odds
Probability estimates of AI-caused catastrophe, from the people building it and the people forecasting it — every figure published and sourced.
Each row is a published probability estimate of AI-caused catastrophe or extinction. Click through to the primary source.
For scale: a 5% chance is one Russian-roulette chamber in twenty. No published insider estimate is zero.
What the public thinks
Polling on AI concern and appetite for regulation. The public is consistently more worried — and more pro-slowdown — than the policy debate assumes.
The incident record
What has actually gone wrong: 1,497 documented real-world incidents, classified by MIT's AI Risk Initiative.
Incidents per year
Reported AI incidents, 1983–2026. The last bar is the current year in progress.
What kind of risk — by year
Each year's incidents stacked by MIT risk repository domain, last 12 years. Hover for values.
Risk domains, all time
Total incidents per MIT risk repository domain.
Top 10 risk subdomains
The most common specific failure modes across all incidents.
How the labs score on safety
The Future of Life Institute's AI Safety Index (summer 2025) — an independent expert panel grades the frontier labs like a school report.
Overall safety grades
Scored 0–4 GPA-style across six safety domains.
- —no company scored above D on existential safety — despite all of them saying AGI could arrive within the decade
- —only 3 of 7 companies reported substantive testing for dangerous capabilities (Anthropic, OpenAI, Google DeepMind)
- —graded across six domains — risk assessment, current harms, safety frameworks, existential safety, governance & accountability, information sharing — by an independent panel of six AI scientists and governance experts
How fast the frontier is moving
Compute, cost and capability curves — from the International AI Safety Report, Stanford's AI Index and Epoch AI.
Governance & transparency
How companies and regulators are (and aren't) keeping up — Stanford AI Index, the EU AI Act and Georgetown CSET.
Jobs & the economy
Exposure estimates from the institutions markets actually listen to.
Misuse & harm to children
The sharpest end of AI harm — deepfakes, exploitation, and AI companions. Handle these numbers with care; they describe real victims.
About this data
Every stat on this page links to its source — click through and check before you put it in a script. The page draws on the MIT AI Incident Tracker, the Future of Life Institute's safety index, the International AI Safety Report, Stanford's AI Index, Epoch AI, the AI Impacts researcher survey, the Forecasting Research Institute, Pew, YouGov, the AI Policy Institute, the IMF, Goldman Sachs, Georgetown CSET and Internet Safety 101 (which aggregates IWF, NCMEC and Common Sense Media research).
The incident record comes from the MIT AI Incident Tracker, which classifies real-world reports from the AI Incident Database against the MIT AI Risk Repository taxonomy, CSET's AI harm taxonomy, and the EU AI Act's risk levels. Incidents are voluntarily reported, so the data under-counts real-world harm; treat trends as indicative, not definitive.
Incident data: MIT AI Risk Initiative & the Responsible AI Collaborative, licensed CC BY 4.0. Snapshot from 2026-08-25.