AI risk

AI risk, evidenced

Advanced AI must be handled with great responsibility. Here is what has actually gone wrong, who it hurt, and which rules answer it.

Concern about AI risk is now shared by researchers, boards, regulators and heads of state. This page keeps that concern honest: every number below comes from the AI Incident Database (recorded harms) and the MIT AI Risk Repository (how experts classify risk), is dated, and links to the record behind it and to the policy instruments that respond.

Recorded incidents
1,663
since 2012, AI Incident Database
Classified to a risk domain
1,496
MIT taxonomy applied by AIID
Instruments tracked
186
across 212 jurisdictions

1. Harm is rising, and its shape is changing

Recorded incidents grow year on year while the mix shifts: generative systems moved misinformation, impersonation and fraud from the margins to the centre. The timeline marks the policy milestones that followed; each bar opens the incidents of that year.

AI incidents recorded per year, with policy milestones
2016: 41 incidents 2016 41 2017: 50 incidents 2017 50 2018: 46 incidents 2018 46 2019: 44 incidents 2019 44 2020: 91 incidents 2020 91 2021: 80 incidents 2021 80 2022: 106 incidents 2022 106 2023: 174 incidents 2023 174 2024: 299 incidents 2024 299 2025: 449 incidents 2025 449 2026: 212 incidents 2026 212 2019-05-22: OECD AI Principles 2021-04-21: EU AI Act proposed 2021-11-23: UNESCO AI ethics recommendation 2023-01-26: NIST AI RMF 1.0 2023-11-01: Bletchley Declaration 2024-05-17: Council of Europe AI Convention 2024-08-01: EU AI Act in force 2025-02-02: EU prohibited practices apply 2025-09-25: Italy Law 132/2025 2026-01-01: Texas TRAIGA in force 2026-08-02: EU high-risk obligations apply

below half of peak year · half to three-quarters · top quarter · policy milestone (hover for details)

AI incidents recorded per year, with policy milestones
YearIncidents
201641
201750
201846
201944
202091
202180
2022106
2023174
2024299
2025449
2026212
  1. 2019-05 OECD AI Principles
  2. 2021-04 EU AI Act proposed
  3. 2021-11 UNESCO AI ethics recommendation
  4. 2023-01 NIST AI RMF 1.0
  5. 2023-11 Bletchley Declaration
  6. 2024-05 Council of Europe AI Convention
  7. 2024-08 EU AI Act in force
  8. 2025-02 EU prohibited practices apply
  9. 2025-09 Italy Law 132/2025
  10. 2026-01 Texas TRAIGA in force
  11. 2026-08 EU high-risk obligations apply

Incident date; the current year is partial. Milestones are adoption or application dates recorded in the policy tracker.

Share of classified incidents by domain, 2019 (%)Percent of incidents classified in that year
Share of classified incidents by domain, 2019 (%) 36 0 Privacy & Security: 14 Privacy & Security 14 AI system safety…: 32 AI system safety… 32 Discrimination and Toxic...: 36 Discrimination and Toxic... 36 Malicious Actors & Misus...: 11 Malicious Actors & Misus... 11 Misinformation: 5 Misinformation 5 Human-Computer Interacti...: 2 Human-Computer Interacti... 2
Share of classified incidents by domain, 2019 (%)
LabelValue
Privacy & Security14
AI system safety…32
Discrimination and Toxic...36
Malicious Actors & Misus...11
Misinformation5
Human-Computer Interacti...2
Share of classified incidents by domain, 2025 (%)Percent of incidents classified in that year
Share of classified incidents by domain, 2025 (%) 58 0 Discrimination and Toxic...: 6 Discrimination and Toxic... 6 Malicious Actors & Misus...: 58 Malicious Actors & Misus... 58 AI system safety…: 9 AI system safety… 9 Misinformation: 16 Misinformation 16 Human-Computer Interacti...: 4 Human-Computer Interacti... 4 Privacy & Security: 6 Privacy & Security 6 Socioeconomic & Environm...: 1 Socioeconomic & Environm... 1
Share of classified incidents by domain, 2025 (%)
LabelValue
Discrimination and Toxic...6
Malicious Actors & Misus...58
AI system safety…9
Misinformation16
Human-Computer Interacti...4
Privacy & Security6
Socioeconomic & Environm...1

2. Who is harmed, and who deploys the systems involved

Harm concentrates on identifiable groups: minors, women, the public, workers and specific communities. The organisations named as deployers repeat, which is where obligations for deployers, transparency and post-market monitoring bite.

Most frequently named harmed partiesAlleged harmed parties as recorded by AIID
Most frequently named harmed parties 88 0 Privacy: 88 Privacy 88 Women and girls: 82 Women and girls 82 Women: 67 Women 67 Epistemic integrity: 46 Epistemic integrity 46 Biometric data subject...: 38 Biometric data subject... 38 Students: 35 Students 35 General public: 30 General public 30 Minors: 29 Minors 29 Educational communitie...: 28 Educational communitie... 28 Victims of deepfake ab...: 24 Victims of deepfake ab... 24

below half of peak · half to three-quarters · top quarter

Most frequently named harmed parties
LabelValue
Privacy88
Women and girls82
Women67
Epistemic integrity46
Biometric data subject...38
Students35
General public30
Minors29
Educational communitie...28
Victims of deepfake ab...24
Most frequently named deployersAlleged deployers; a name is an allegation in a report, not a finding
Most frequently named deployers 62 0 Scammers: 62 Scammers 62 Synthetic media creato...: 44 Synthetic media creato... 44 Deepfake creators: 42 Deepfake creators 42 Google: 42 Google 42 Tesla: 41 Tesla 41 Openai: 40 Openai 40 Facebook: 35 Facebook 35 Unknown: 24 Unknown 24 Meta: 23 Meta 23 Amazon: 22 Amazon 22

below half of peak · half to three-quarters · top quarter

Most frequently named deployers
LabelValue
Scammers62
Synthetic media creato...44
Deepfake creators42
Google42
Tesla41
Openai40
Facebook35
Unknown24
Meta23
Amazon22
Assessed harm level (CSET)Only incidents with a CSET assessment
Assessed harm level (CSET) 126 0 None: 126 None 126 AI tangible harm event: 44 AI tangible harm event 44 Unclear: 10 Unclear 10 AI tangible harm near-...: 10 AI tangible harm near-... 10 AI tangible harm issue: 10 AI tangible harm issue 10
Assessed harm level (CSET)
LabelValue
None126
AI tangible harm event44
Unclear10
AI tangible harm near-...10
AI tangible harm issue10

3. Where harm is recorded versus where rules exist

For policymakers and funders the question is coverage: do the places where incidents are recorded have binding AI rules? Only 140 incidents carry a country code, and reporting is biased toward English-language media, so read this as a prompt for enquiry rather than a ranking.

Incidents by country and the AI instruments recorded there
CountryIncidentsInstruments trackedBindingStatus
United States9142Federal: executive-branch policy (Executive Order 14179 of January 2025 and the July 2025 AI Action Plan) plus...
United Kingdom64Principles-based, regulator-led framework; no AI-specific Act in force. The government has signalled a future...
Russia421National AI strategy (2019, updated 2024) as policy; experimental legal regimes in force; AI ethics code volun...
China432Instruments on record: Interim Measures for the Management of Generative Artificial Intelligence Services (202...
Germany321EU AI Act applies; national implementing act (KI-Marktüberwachungs- und Innovationsförderungsgesetz) proposed;...
New Zealand32National AI Strategy (2025) and responsible-AI guidance as policy; Algorithm Charter for government agencies;...
Canada33Instruments on record: Pan-Canadian Artificial Intelligence Strategy (2017, strategy, adopted); Artificial Int...
France32EU AI Act applies; national AI strategy (phases since 2018) as policy; CNIL AI guidance; competent-authority d...
Vietnam221Instruments on record: National Strategy on Research, Development and Application of Artificial Intelligence t...
India241No binding cross-sector AI law. Binding obligations arise from the DPDP Act 2023 (phased commencement under th...
Ireland21EU AI Act applies; competent authorities designated in phases; national AI strategy refreshed 2024.
Australia23No binding cross-sector AI law; mandatory guardrails were consulted on in 2024 but the December 2025 National...
South Korea211Instruments on record: Framework Act on the Development of Artificial Intelligence and Establishment of a Foun...
Indonesia131Instruments on record: Strategi Nasional Kecerdasan Artifisial Indonesia 2020-2045 (2020, strategy, adopted);...
Sweden11EU AI Act applies; AI Commission roadmap (2024) under implementation as policy; national approach (2018).

4. How policy responds, domain by domain

Each domain links to the instruments whose recorded use cases address it, and to the obligations, deadlines and templates behind them. Click a domain for its subdomains, frameworks and incidents.

5. What to do with this, depending on who you are

AI CISO or compliance lead

Start from the domains your systems touch, then the obligations with dates.

Researcher

Filter, drill down and export with licence and citation attached.

Policymaker or diplomat

Compare jurisdictions and see which harms remain unaddressed.

Civil society, donor, journalist

Evidence of who is harmed and where governance is missing, with sources.

Causal taxonomy: who causes the risk, and was it intended?

Risk entries in the MIT database by responsible entity and intent. Timing: Not coded 237 · Other 430 · Post-deployment 977 · Pre-deployment 190 .

Risk entries by entity × intent
Risk entries by entity × intent
Entity \ IntentIntentionalUnintentionalOtherNot coded
Human 356 149 88 1
AI 147 320 194 1
Other 39 84 210 2
Not coded 1 1 2 238

Explore: domains and subdomains

Each band is a domain; each block a subdomain, sized by how much of the evidence it carries. Click a block for its definition, causal breakdowns, frameworks, risk entries and incidents.

Informational only, not legal advice. Verify every claim against the linked official sources and consult qualified counsel before acting.