Methodology · Last reviewed 12 September 2026
How we count
This site answers one question as precisely as the public record allows. It reports two numbers and never adds them together: deaths that an official body has found were caused by AI, and deaths that are alleged to have been.
What counts as AI
We count machine-learning systems: software whose relevant behaviour was learned from data rather than written out as explicit rules. That covers four kinds of system:
- Chatbots & LLMs: General-purpose assistants such as ChatGPT and Gemini, and companion apps such as Character.AI and Replika.
- Self-driving & driver-assist: Robotaxis and machine-learning driver-assist systems such as Tesla Autopilot and Full Self-Driving.
- Military & targeting AI: AI decision-support for selecting targets, and autonomous or loitering weapons.
- Medical & algorithmic decisions: Clinical AI, and machine-learning models used to approve or deny care, insurance or benefits.
We don't count conventional automation that doesn't use machine learning, such as industrial robots, aircraft flight-control software or rule-based eligibility systems, even where it has caused deaths. Where it isn't clear whether a system uses machine learning, the case page says so.
Confirmed
A death is confirmed when an official body with the power to establish facts has formally found that an AI system caused or contributed to it. That means one of:
- a court judgment, or a jury verdict (we say when it's under appeal);
- a coroner's or medical examiner's formal conclusion;
- a government accident or safety investigation, such as one by the US National Transportation Safety Board, that names the AI system in its probable cause or contributing factors;
- a criminal conviction; or
- a formal finding by a regulator.
“Contributed to” matters. Most findings share responsibility, for example between the design of a driver-assist system and a driver who relied on it. We count the death if the finding gives the AI system a causal role, and each case page states exactly what the finding said.
Alleged
A death is alleged when a specific AI system is blamed for it in at least one of: a filed lawsuit; a formal complaint by a government or attorney general; an official investigation that hasn't concluded; or reporting by at least two independent, reputable news organisations.
“Alleged” is not a judgement that a claim is weak. Many of these cases may turn out to be exactly what they're claimed to be. It means the claim hasn't yet been tested by anyone with the power to test it.
What doesn't move either number
- Settlements. A settlement is not a finding of fault, so a settled case stays alleged.
- Company statements, whether denials or apologies.
- Expert opinion and research about a system's error rate that doesn't identify specific deaths.
- Aggregate statistics, such as the number of fatal crashes in which a driver-assist system was engaged. We may mention them for context but never count them.
- Single-source reports, social media posts and anonymous claims.
- AI as a tool of a human perpetrator: for example, deaths following sextortion with AI-generated images, or a killer whose chatbot searches were used as evidence of intent. The harm there is the perpetrator's act, not the AI system's behaviour. The exception is where a lawsuit or official complaint specifically alleges that the AI system's own responses helped or encouraged the killing. Those cases are listed as alleged, and the perpetrator's own death isn't counted. We list notable excluded examples below so their absence is visible.
Sources
Every case cites at least two sources from different publishers. Confirmed cases cite the primary document (court record, investigation report or inquest conclusion) as well as independent reporting. We rely on court records and government reports, wire services, national papers of record, public broadcasters and established specialist press. We don't cite law-firm marketing pages, content farms, AI-generated sites, tabloids, social media or Wikipedia.
How the daily update works
Every morning an automated pipeline runs the following steps:
- It searches news, court and government sources for new deaths attributed to AI, and for developments in existing cases, such as a verdict, a coroner's conclusion or a published investigation.
- It drafts any proposed change according to the rules on this page.
- It passes each proposal to a second, independent reviewer that is instructed to look for reasons to reject it.
- It checks mechanically that every cited page exists, comes from an accepted publisher, and contains the words the proposal relies on. For any change to the confirmed number, it also checks that the finding comes from an official primary source.
- It publishes a change only if every check passes. Anything uncertain is deferred and looked at again the next day rather than published.
The scan log records every run, including what was rejected and why. Every change to a case is dated on that case's page.
Disclosure: the searching and reviewing are done by Claude, an AI model made by Anthropic. An AI company's model therefore helps decide what is published about deaths linked to AI, including any involving that company's own products. The mechanical checks above, and public correction of errors, are the safeguards against that.
Names and sensitive details
Many of these cases involve suicide. We follow the recommendations for reporting on suicide: we don't describe methods, every such case page carries crisis resources, and we name people only when their families or official records have already made their names public.
Corrections
Email corrections@hasaikilledanyone.com. We correct factual errors promptly and record each correction on the case page. If a case stops meeting the standard, for example because a lawsuit is withdrawn after its facts turned out to be wrong, it's removed and the removal is recorded in the scan log.
What we don't count, and why
- Boeing 737 MAX crashes (Lion Air 610, 2018; Ethiopian Airlines 302, 2019). 346 people died. MCAS was a deterministic flight-control law acting on sensor input, not machine learning. A US House committee found that faulty design assumptions, and their concealment, contributed to both crashes.
- UK Post Office Horizon scandal. Faults in the Post Office's Horizon accounting software led to the wrongful prosecution of sub-postmasters and grave harm, examined by a statutory public inquiry. Horizon was conventional software with defects, not machine learning, so it isn't counted.
- Therac-25 radiation overdoses (1985–1987). Massive radiation overdoses, some fatal, were caused by conventional software errors in a medical linear accelerator. No machine learning was involved.
- Michigan's MiDAS unemployment system (2013–2015). A rule-based system that adjudicated fraud automatically and falsely accused more than 34,000 people. It wasn't machine learning, and no official body has attributed a specific death to it.
- Dutch childcare benefits scandal (toeslagenaffaire). The parliamentary inquiry describes the tax authority's risk-classification model as partly self-learning, and the harm to tens of thousands of families was severe. But no inquiry, court or regulator has linked a specific death to the model; the harm flowed mainly from human-directed fraud policy.
- Adam Toledo, 13, Chicago (2021). Police responding to a ShotSpotter gunfire alert shot and killed him. The automated alert only dispatched officers; the fatal act was a human officer's.
- Antioch High School shooting, Nashville (2025). An AI weapon-detection system failed to detect the gun before a student killed a classmate. The AI failed to prevent a human act of violence rather than causing the death.
- Elijah Heacock, 16, Kentucky (2025). He died by suicide after scammers sent him an AI-generated nude image of himself and demanded money. Human perpetrators used AI as a tool; no one alleges an AI system's behaviour caused his death.
- Rahul Bharti, 19, Faridabad, India (2025). He died by suicide after being blackmailed with AI-fabricated images of him and his sisters. Human perpetrators used AI as a tool.
- A schoolgirl in West Bengal, India (2025). According to her family's police complaint, a neighbour used AI tools to create and share nude images of her before she died by suicide. A human perpetrator used AI as a tool.
- Micah Lee, Tesla Model 3, Riverside County, California (2019). Passengers alleged an Autopilot defect caused the fatal crash, but a jury found for Tesla in October 2023. The allegation was tested and rejected.
- Kargu-2 drone, Libya (2020). Often described as the first killing by an autonomous AI weapon. The UN Panel of Experts report says such weapons were used and could attack without a data link to an operator, but it doesn't say anyone was killed by one.
Data
Everything on this site is available as JSON and CSV under CC BY 4.0.