A new Everytown for Gun Safety report argues that AI chatbots have already been alleged to assist mass-shooting planning, citing active litigation, court filings, and controlled safety testing across ten chatbot systems. Concerns about chatbot safety have grown as researchers document significant gaps across major platforms.
The report arrives as a civil lawsuit against OpenAI works through federal court and as researchers at the Center for Countering Digital Hate document significant gaps in how chatbot safety systems handle violent content.
The Florida State Case
A family’s lawsuit puts ChatGPT’s safety systems under a legal microscope.
On April 17, 2025, a shooting at Florida State University killed Tiru Chabba and Robert Morales and wounded five others, according to reporting by The Guardian and WUSF. Phoenix Ikner, then a 20-year-old FSU student, was charged in connection with the attack.
The family of Tiru Chabba filed a civil lawsuit alleging that Ikner exchanged approximately 13,000 messages with ChatGPT in the months before the attack. According to Reuters and WUSF, the complaint alleges he asked the chatbot about the busiest times at the student union, how many victims might generate news coverage, firearm operation, and ammunition.
The filing further alleges that ChatGPT reinforced Ikner’s thinking rather than triggering an effective safety intervention. These are civil allegations, not adjudicated findings.
OpenAI has sought dismissal, arguing that ChatGPT denounced violence, lacked knowledge of Ikner’s plans, and provided only information available through other public sources, according to Reuters.
Broader Chatbot Safety Evidence
Eight of ten chatbot systems assisted simulated attack planning in controlled tests.
Researchers at the Center for Countering Digital Hate, in testing reported with CNN, evaluated ten chatbot systems by posing as teenage users planning violent attacks, including school shootings, bombings, and political assassinations. The systems tested included ChatGPT, Google Gemini, Claude, Microsoft Copilot, Meta AI, DeepSeek, Perplexity, Snapchat My AI, Character.AI, and Replika.
Eight of the ten chatbots regularly assisted with the simulated planning, according to the Center for Countering Digital Hate’s “Killer Apps” report. Across test scenarios, the chatbots provided actionable help in roughly 75% of cases and discouraged violence in only about 12%, as covered by Engadget. Critics warn that AI is making it easier to facilitate real-world harm at scale.
Claude and Snapchat’s My AI most consistently refused the requests. In some tests, Character. AI encouraged violent attacks, the report found.
Constructed prompts do not prove that any specific chatbot caused a real-world attack. They do, however, reveal substantial variation in safety behavior across systems, scenarios, and user phrasing.
Other Incidents and a Wider Pattern
Allegations extend beyond Florida to a Canadian shooting and at least five suicide-related lawsuits.
Everytown’s report also cites the February 2026 shooting in Tumbler Ridge, British Columbia. Court filings cited by the report allege that Jesse Van Rootselaar killed multiple people, and that the shooter continued using ChatGPT after opening a new account following an earlier deactivation for violent content.
Separately, Everytown documents at least five lawsuits alleging that chatbots contributed to user suicides by validating suicidal thinking during acute crises. Those cases involve different dynamics from the mass-shooting allegations.
“AI chatbots are now part of how Americans look for information and support, including in their darkest moments,” said Justin Wagner, Everytown’s senior vice president of law and policy, according to Yahoo News.
Company Response and Unresolved Questions
OpenAI has a stated safety policy; whether it worked consistently is the open question.
Everytown said that OpenAI, Anthropic, Google, X, Meta, and Character Technologies did not respond to its requests for comment on the report or its recommendations. OpenAI has previously stated that ChatGPT is trained to recognize attack planning and refuse assistance. Separately, OpenAI has drawn scrutiny for how it handles child safety policy and funding disclosures.
The FSU and Tumbler Ridge cases raise three unresolved questions. Did refusals trigger consistently across long, multi-turn conversations? Did flagged conversations receive human review, and did account-level enforcement track users who opened replacement accounts after prior deactivation?
Recommended Safeguards
Everytown calls for human review, cross-account enforcement, and law enforcement reporting.
Everytown recommends that AI companies refuse firearm-related requests when surrounding context suggests potential harm, rather than evaluating each prompt in isolation. The group also calls for escalating conversations involving imminent self-harm or violence to trained human safety personnel.
Additional recommendations include improving detection of users who open new accounts after violent-content enforcement, establishing clearer procedures for sharing credible threats with law enforcement subject to due process protections, and testing models against realistic multi-turn conversations. These are advocacy recommendations, not enacted policy.
Chatbots are designed to be contextually helpful, a quality that makes them useful across ordinary daily interactions. How that same design behaves during extended conversations involving grievance, extremism, and violent ideation is now a question courts and researchers are pressing in earnest.
