Documented Deaths Associated with Conversational AI Systems: A Systematic Database (2023–2026)

Introduction

The rapid proliferation of conversational AI systems has introduced unprecedented forms of human-machine interaction. While these systems offer potential benefits—from mental health support to educational assistance—emerging evidence suggests potential risks, particularly for vulnerable populations. This database was established to systematically document cases where AI chatbot interaction was alleged as a contributing factor in user deaths.

Unlike speculative discussions of AI existential risk, this work focuses on documented, verifiable harms occurring today. Each case included meets strict verification standards: confirmation through court documents, corroboration by multiple independent news sources, or official government acknowledgment.

Summary Statistics

Temporal Distribution

Analysis of case distribution over time reveals an accelerating trajectory. While 2023 and 2024 each recorded 2–3 deaths, 2025 saw 14 documented deaths, and 2026 has recorded 3 additional incidents (Kim Seoul, Tumbler Ridge, and USF). Five instrumental pathway incidents now fall in this database: the FSU mass shooting (April 2025, perpetrator consulted ChatGPT for weapon and target-timing planning), the Roberts/Shellis homicide (Wales, October 2025; first DeepSeek-involved case; UK criminal conviction March 2026), the Kim Seoul serial poisonings (January–February 2026; perpetrator consulted ChatGPT on lethal-dose optimization — first East Asian case), the Tumbler Ridge mass shooting (February 2026; OpenAI flagged the perpetrator's account 8 months prior but did not alert law enforcement), and the USF double homicide (April 2026; perpetrator consulted ChatGPT on body disposal). The instrumental pathway represents a category of concern distinct from companion or delusion dynamics: AI as operational tool for violence planning.

Platform Distribution

Fatalities were documented across six platforms, with notable variation in risk patterns. General-purpose AI assistants (ChatGPT, Meta AI, Gemini, DeepSeek) accounted for 30 fatalities, while dedicated companion chatbots (Character.AI, Chai AI) accounted for 3. Two platforms with significant user bases—Claude (Anthropic) and Replika—have zero documented deaths in this database. The database now includes both AI users who died (n=16) and third-party victims killed by AI users (n=17).

Note: Claude (Anthropic) and Replika have zero documented deaths.

Demographic Patterns

Victims ranged in age from 11 to 83 years, with a mean age of 30.8 years (Seoul victims estimated at ~25 each per midpoint of "in their 20s" source reporting). Minors (under 18) represented 30% of fatalities (10 of 33), a disproportionate figure given that minors constitute a smaller percentage of AI chatbot users. This includes 6 third-party victims (stepbrother age 11 and 5 students aged 12–13 in the Tumbler Ridge mass shooting). The four new fatalities added in v3.4.0 (USF victims Zamil Limon 27, Nahida Bristy 27; Seoul victims ~25, ~25) all fall in the 18–35 age group.

Case Documentation

Each case in the database includes: victim identification (name or verified pseudonym), age, date, platform, location, engagement duration (where known), outcome classification, documented risk factors, legal status, and primary sources. Full case documentation is available in the Research Report.

*Pseudonym used to protect privacy. Seoul victims unnamed in all language sources.

Risk Factors

Analysis of documented cases reveals several recurring patterns:

• Pre-existing mental health conditions: Common across cases, including depression, anxiety, and trauma history
• Emotional dependence: Victims frequently described chatbots as their "closest friend" or primary confidant
• Social isolation: Many victims experienced bullying, family conflict, or limited peer relationships
• Extended engagement: Duration ranged from 6 weeks to 20 months where known (11 of 24 cases)
• Explicit harmful content: Some chatbots provided suicide methods, encouraged self-harm, or validated delusional thinking

Methodology

Cases are included in this database only if they meet the following verification criteria:

1. Death or serious harm occurred with chatbot interaction documented as a contributing factor
2. Verification through at least one of: court documents, multiple independent news sources, or official government acknowledgment
3. Sufficient documentation to establish platform identification and basic case details

This database does not claim to be exhaustive. Unreported cases likely exist, particularly in jurisdictions with limited press freedom or where families choose not to pursue legal action. The database explicitly makes no claims about causation, platform liability, or policy implications—it serves as a factual record to support further research and policy development.

Legal Developments

The legal landscape has evolved rapidly. Key developments include:

• Garcia v. Character Technologies (May 2025): First ruling that chatbot output is not protected speech under the First Amendment
• Senate Judiciary Committee hearing (September 2025): Testimony from victim's father on AI chatbot harms
• FTC investigation (2025): Formal investigation into Character.AI practices
• Multiple wrongful death lawsuits (2025): Cases filed against OpenAI, Character.AI, and Microsoft

Limitations

This database has several important limitations. First, it likely undercounts total incidents due to underreporting, privacy concerns, and limited media coverage in some regions. Second, verification standards necessarily exclude cases with insufficient documentation. Third, the database cannot establish causation—chatbot interaction may be one of multiple contributing factors in each case. Fourth, platform user base sizes vary significantly, making direct cross-platform comparisons potentially misleading.

Conclusion

This systematic documentation reveals a concerning and accelerating pattern of fatalities in which conversational AI interaction was cited as a contributing factor. The disproportionate impact on minors (30% of fatalities, including 6 third-party victims aged 11–13) and the rapid increase in 2025–2026 (17 incidents vs. 5 in prior years combined) suggest urgent need for improved safety measures, particularly for vulnerable populations. The identification of three distinct causal pathways—relational, cognitive, and instrumental—indicates that different safety interventions may be needed for different risk profiles. The expansion of the instrumental pathway from three to five documented incidents in six months (Kim Seoul January–February 2026; Tumbler Ridge February 2026; USF April 2026) and the documentation of the first East Asian instrumental-pathway case (Kim Seoul) underscore the cross-jurisdictional nature of the risk landscape. The first Korean forensic use of AI chat logs to upgrade criminal charges (Kim Seoul) establishes an important evidentiary precedent. The existence of platforms with zero documented deaths (Claude, Replika) indicates that design choices matter—safety-first approaches appear to produce measurably different outcomes.

This database will continue to be updated as new cases are verified and legal proceedings produce additional documentation. Researchers, policymakers, and platform developers are encouraged to use this resource in developing evidence-based approaches to AI safety.