Methodology
What counts as an incident
An incident is a specific, real-world event in which the behaviour or use of an AI system caused or contributed to actual harm to people, an organisation or the public. The AI must play a real part in what happened: the harm would probably not have happened in the same way without it. It does not have to be the only cause.
- Included: harms reported by reliable sources, including alleged harms (for example in a lawsuit or a police statement), always attributed to whoever makes the claim; crimes in which AI was used as a tool (for example cloned voices in a scam); AI agents whose own actions caused a loss; and, from research tests of AI systems, only events in which people or systems outside the test were actually affected.
- Excluded: warnings, predictions and risk assessments; vulnerabilities that were not exploited; tests and demonstrations without real-world victims; opinion pieces; incidents known only from social-media posts.
- AI agent incidents are those in which an AI system that acts with tools on its own (running code, browsing, buying, sending, deleting) carried out the action that caused the harm.
- Repeated look-alike cases (for example many court sanctions for AI-fabricated citations) are included when a court or regulator imposed a sanction or found harm.
Where reports come from
- News feeds of established media outlets and official bodies, in several languages.
- The GDELT Project, whose open data lists news articles worldwide.
- Official sources: regulators, courts, police and government agencies, and the US National Highway Traffic Safety Administration's crash-report data (used for statistics only).
- Other public incident databases — the AI Incident Database, the AIAAIC Repository and the OECD AI Incidents Monitor — are used only as leads to find original reports; we do not copy their text or classifications.
Our crawler identifies itself as hurtbybot-bot, respects robots.txt and each site's terms of use, and waits between requests. See our crawler.
Evidence rules
- Every incident needs at least two independent reliable sources whose full text we checked, or one official source (a court document, or a regulator, police or government statement). Reports from the same publishing group, or copies of the same news-agency story, count as one.
- Reliable sources are established news organisations, official bodies and the companies involved (for their own statements). Social-media posts, forums and content farms are not used as sources.
- Some publishers do not allow automated access to their articles. We never fetch or process their pages; when another database points to one of their reports about the same incident, we may list it under "Further reading" without using it as a source.
- Every fact on a page comes from the linked sources. The page shows when each source was published and when we retrieved it.
How pages are written
Summaries are written by an AI model using only the source texts, under fixed rules:
- Every sentence that links a company or product to a harm says who states it ("according to…", "the lawsuit alleges…", "police said…", "the court found…", "the company said…"). This also applies when a court, a regulator or the company has confirmed part of what happened: the confirmation label on a page says what was confirmed and by whom, and it does not turn other claims into facts.
- If a company or authority responded — including a denial or a refusal to comment — the response is included.
- Private individuals are never named, even when the sources name them. Minors and their families are never named or described in identifying detail.
- Suicide and self-harm are reported following international safe-reporting guidelines: no details of method, no sensational language, and help-line information on the page.
- Numbers must appear in a source. Dates are shown from structured fields, not retyped in the text.
- No quotations, images or video from the sources are republished.
Automatic checks before publication
- Rules check — required fields, valid dates and categories, numbers traceable to sources, no quotations, no case numbers or legal citations in the text, no names of private individuals, no sensational wording.
- Independent fact-check — a model from a different AI provider compares the draft with the source texts and flags anything unsupported, misattributed, stated as fact when it is only alleged, or classified wrongly.
- Duplicate check — new reports are matched against existing incidents so that one event has one page.
- Site check — before each update goes live, the whole site is checked for empty pages, broken links, missing translations and other errors.
A draft that fails is not published. It is retried in a later round and dropped after three failures. We prefer to publish less rather than publish something wrong.
Classification
Severity is the highest level whose description is met, using only harms and figures stated in the sources; when the sources do not give the information a level needs, that level is not used.
Type of harm
| Fraud & scams | Money or assets lost to deception that used AI (voice clones, deepfake video calls, AI-written phishing, fake investment ads with AI likenesses). |
|---|---|
| Deepfake abuse | Non-consensual synthetic images, video or audio of real people used to humiliate, sexualise, extort or impersonate them (not primarily for money). |
| Psychological harm | Mental-health harm linked to interaction with an AI system: encouragement of self-harm, emotional dependency, manipulation, distressing outputs. |
| Physical injury or death | Bodily injury, death or property damage involving a physical AI system or AI-driven decision (vehicles, robots, medical systems). |
| Wrongful accusation | A person wrongly identified, arrested, accused, flagged or penalised because of an AI output (face recognition, fraud scores, AI detectors). |
| Discrimination | Unfair treatment of people based on protected characteristics produced or amplified by an AI system. |
| Wrongful denial | Benefits, credit, insurance, care, employment or housing wrongly denied, cut or delayed by an automated decision system. |
| Privacy & data leaks | Personal or confidential data exposed, leaked or misused through an AI system or AI product. |
| Defamation & false claims | An AI system generated false, damaging statements about an identifiable person or organisation. |
| Misinformation | AI-generated false content that spread and caused measurable harm (public confusion, panic, election interference), not about one identifiable person. |
| Agent mis-action | An AI agent took an action on its own that caused loss: deleting data, unauthorised purchases or transfers, sending messages, changing systems. |
| Fabricated citations | AI-fabricated facts or legal citations submitted in court, official or professional work, leading to sanctions, errors or wasted proceedings. |
| Security breach | AI systems exploited or used to compromise systems: prompt injection, AI-assisted intrusion, malicious AI tooling. |
| Harmful content | AI output that exposed people (especially minors) to sexual, violent or dangerous content or instructions. |
AI system
| Chatbot / assistant | Conversational AI (general assistants, companion apps, customer-service bots that only talk). |
|---|---|
| AI agent | An AI system that takes actions with tools on its own: coding agents, browser/computer-use agents, agents that buy, send, delete or change things. |
| Generative image / video / voice | Systems that create synthetic images, video, audio or voice clones. |
| Recommender / ranking | Feeds, ranking, ad targeting or content moderation algorithms. |
| Autonomous vehicle / driver assist | Self-driving cars, robotaxis, driver-assistance systems, autonomous drones or vessels. |
| Face recognition / biometrics | Facial recognition, voice or other biometric identification and emotion recognition. |
| Automated decision system | Scoring or eligibility systems used in credit, insurance, welfare, hiring, policing, healthcare or education. |
| Robot | Physical robots other than vehicles (industrial, delivery, domestic, surgical). |
| AI developer tooling | Models, APIs, coding assistants or AI platforms whose failures or exploits harmed users or operators (when not an agent acting on its own). |
| Other AI system | Any other AI system. |
Severity
| 1 — Minor | Inconvenience or embarrassment, reversible within days; loss under US$1,000 per victim; no injury (e.g. a wrong chatbot answer later corrected, a small unwanted purchase). |
|---|---|
| 2 — Moderate | Loss of US$1,000-100,000; recoverable data loss or downtime under 24 hours; minor injury; wrongful stop or detention under 24 hours; personal data of under 10,000 people exposed; a non-sexual deepfake harming one identifiable person. |
| 3 — Serious | Injury needing hospital care; loss of US$100,000-10 million; detention of 24 hours or more or a wrongful charge; unrecoverable data loss for an organisation; outage of 24 hours or more or affecting 100,000+ users; sensitive data of 10,000+ people exposed; sexual deepfakes of identifiable people; discriminatory denial of jobs, credit or benefits for a group; harm to a minor. |
| 4 — Severe | Death of a person or serious harm to health; loss over US$10 million; sensitive data of 1 million+ people exposed; serious, irreversible disruption of critical infrastructure; sexual abuse material or harm to many minors; a court or regulator finding of a fundamental-rights infringement. |
| 5 — Catastrophic | Multiple deaths, or disruption of a community or society beyond its capacity to cope. |
How established
| Reported | Described by reliable sources, but not confirmed by an involved party and not established by a court or regulator. Includes lawsuits filed, complaints, arrests, charges, investigations, consent orders and settlements without an admission of wrongdoing. |
|---|---|
| Confirmed by a party | The company that built or used the AI system, or the harmed organisation, publicly confirmed the facts of what happened (even if it disputes blame), or police/authorities confirmed that the event occurred. |
| Regulator or court finding | A court judgment or ruling with findings of fact, a final regulator decision with findings, a criminal conviction or guilty plea, or a tribunal decision establishes what happened. Settlements, consent orders and charges do not count. |
Translations
Pages are first written in English and then translated by an AI model, using a glossary. Every translation is checked automatically — the same numbers and names, the same number of timeline items, no untranslated text — and then reviewed by a model from a different AI provider for changed meaning, lost attributions or denials, and unsafe wording. A language version is published only if it passes both. Chinese summaries are shortened to 80–150 characters; they may leave out secondary details but keep who says what.
The "Protect yourself" guides are checked the same way, and in addition every recommendation and fact in the English guide is checked against the official source it links to; a guide is shown only while that check passes. The check is repeated every 30 days.
Quality statistics
Counts from our automated pipeline over the last 30 days.
| Incidents published | 2 |
|---|---|
| Incident updates published | 0 |
| Drafts held back by the automatic checks | 8 |
| Candidate incidents dropped after repeated failed checks | 3 |
| Correction or removal requests handled | 1 |
| Entries corrected, taken offline or removed after a request | 0 |
Updates
New reports are collected every hour. Every hour, headlines that name a company or product of a tracked incident together with a development (a lawsuit, ruling, fine, recall or response) are checked against that incident, and once a day we run additional searches; when a new source is confirmed, the timeline is updated. Each page shows when it was added and last updated.
Limitations
- Our counts reflect what reliable sources have reported and what our pipeline has found and verified. They are not the true number of AI harms in the world and should not be read as a ranking of companies or products.
- Media coverage is uneven across countries, languages and types of harm.
- Crisis lines shown on pages about suicide or self-harm are checked every day against the official page of each service. Where that page cannot be reached from our checker, the number is confirmed from a news report instead, which may not show a later change; such entries are marked on the page.
- Automated systems make mistakes. If you find one, please report it.