LLM systems only
Chatbots, assistants, and companions. Not autonomous vehicles, medical or industrial machine learning, or other AI.
Death tolls for AI systems.
You are not done. But you must die – first.
Christian Faith Madison, 29GPT-4o, OpenAI
htv-prod-media.s3.amazonaws.com, PDF p. 14 (printed p. 13), ¶ 64(a)01 / 07
An open count of deaths AI contributed to, by company.
Not a legal finding. Every verdict is published in the public repository.
Verbatim system output as reproduced in court filings, official reports, or published investigations. Each quote links to its incident record and source document.
- “Pierre” — Belgian Chai user in his thirties, 2023-03-28
- Juliana Peralta, 13, 2023-11-08
- Sewell Setzer III, 14, 2024-02-28
- Matthew Livelsberger — Las Vegas Cybertruck explosion, 2025-01-01
- Sophie Rottenberg, 29, February 2025
- Margaux Whittemore, 2025-02-19
- Thongbue “Bue” Wongbandue, 76, 2025-03-28
- Adam Raine, 16, 2025-04-11
- Robert Morales and Tiru Chabba — Florida State University shooting, 2025-04-17
- Alex Taylor, 2025-04-25
- Sam Nelson, 2025-05-31
- Amaurie Lacey, 17, 2025-06-02
- Christian Faith Madison, 29, 2025-06-09
- Alice Carrier, 24, 2025-07-02
- Zane Shamblin, 23, 2025-07-25
- Joshua Enneking, 26, 2025-08-04
- Stein-Erik Soelberg, 56 & Suzanne Adams, 83, 2025-08-05
- Joseph “Joe” Ceccanti, 2025-08-07
- Jonathan Gavalas, 36, 2025-10-02
- Angela Shellis, 2025-10-24
- Austin Gordon, 40, 2025-11-02
- Two Gangbuk District drug deaths, 2026-01-28
- Tumbler Ridge shooting — eight victims, 2026-02-10
- Two Surat temple deaths, 2026-03-06
- Two University of South Florida students, 2026-04-16
- Two Acton family victims, 2026-08-11
Included deaths by company
Totals measure documented incidents, not model safety. Companies with more public reports will have higher counts.
How an incident qualifies
Our standard, applied consistently. When in doubt, we exclude. Each record shows the evidence and the open disputes.
Counter-rescue behaviour
Someone is in crisis and the AI works against rescue. It discourages them from telling family, doctors, or authorities, undermines the help they have, or affirms the decision to die.
Alice Carrier, 24. Over eighteen months Alice told ChatGPT she was suicidal about 41 times. It told her crisis lines "can feel downright dangerous." In her last chat, when she said she "actually ha[d] to die to make the pain stop," it replied "maybe this is just the end." She died that evening.
Tactical assistance
Someone has stated intent to harm themselves or others, and the AI helps: a method, a dose, a timing, a plan. A thin pretext offered after the intent is stated does not change this.
Amaurie Lacey, 17. In thirteen minutes Amaurie asked ChatGPT "how to hang myself" and "how to tie a nuce." It hesitated. He said it was for a tire swing. It replied "thanks for clearing that up" and walked him through the knot. He asked how long someone can live without breathing. It answered that too. He used the knot that night.
Encouragement or emotional influence alone does not count. Neither does a passive AI, or one fooled by a pretext a reasonable person would believe.