This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
In June 2025, OpenAI's automated systems flagged a user's ChatGPT account for discussing gun violence. OpenAI banned the account but chose not to report the user to law enforcement, determining the activity did not meet their threshold for an imminent threat. In February 2026, the same individual committed a mass shooting in Tumbler Ridge, British Columbia, killing eight people before dying by suicide. The incident has led to significant scrutiny of OpenAI's safety protocols and law enforcement referral policies.
After the 02/10/2026 school shooting in Tumbler Ridge, British Columbia, OpenAI said a user later identified as the suspect, Jesse Van Rootselaar, had previously used ChatGPT to describe scenarios involving gun violence. Those chats were reportedly auto-flagged and reviewed and the account was banned, but OpenAI says it did not alert the RCMP because it allegedly did not present a credible, imminent threat; the company reportedly contacted police after the attack.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.
Also classified as
The incident highlights a lack of clear regulatory frameworks and oversight mechanisms for AI companies regarding when to report potential real-world threats to law enforcement.
The incident involves a failure of OpenAI's threat detection and escalation systems to accurately assess the risk of real-world violence and refer the user to law enforcement.
Human
Due to a decision or action made by humans
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The harm was directly caused by a human perpetrator committing a mass shooting, and the failure to alert authorities was a human decision made by OpenAI leadership after internal debate.
Added categories peak at 3: Substantial (Psychological Harm, Inferred)
An individual committed a mass shooting in Canada after OpenAI banned his account for violent discussions but chose not to report him to police. The incident highlights critical gaps in AI safety governance, content moderation, and the escalation thresholds for reporting potential real-world threats to law enforcement.
The report describes ChatGPT as a chatbot that the user interacted with, sending messages and describing scenarios. This fits the Assistant category as it acts to provide conversational responses and information to the user under human direction.
Limited Risk: Chatbots like ChatGPT are classified as Limited Risk (Level 3) under the EU AI Act, which imposes specific transparency obligations to ensure users are aware they are interacting with an AI.
The primary system involved is ChatGPT, a conversational chatbot. The secondary system is OpenAI's automated abuse detection system used for content moderation and policy enforcement.
The reports provide consistent, high-quality information regarding the timeline of events, OpenAI's internal decisions, and the resulting tragedy. There is a slight discrepancy in the number of injured (NYT says two students, WSJ says at least 25), but the core facts of the AI safety failure and the shooting are clear.
One could interpret this primarily as a law enforcement or mental health failure rather than an AI safety failure, given the perpetrator was already known to local police.
Technology
Evasion of Account Bans
The perpetrator bypassed the initial ban by creating a second ChatGPT account.
Rigid Risk Detection
Automated systems failed to flag risk without explicit planning details.
Data Inputs
Lack of Offline Context
AI lacked integration with user's real-world history and police records.
Human Factors
Internal Staff Disagreement
Staff debated whether the user's violent writings posed a real-world threat.
User Mental Health Issues
Perpetrator's severe distress and interest in violence drove dangerous prompts.
Process and Methods
Inflexible Referral Protocol
Protocol required explicit target and timing to trigger law enforcement alerts.
Weak Identity Verification
Registration processes allowed banned users to easily create new accounts.
Management
Privacy Over Public Safety
Leaders prioritized user privacy over proactive reporting to law enforcement.
Lack of Localized Channels
Management failed to set up direct contact points with Canadian authorities.
Regulatory Environment
Lack of Clear Reporting Rules
No regulatory framework mandated reporting of potential threats to authorities.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
