A prominent tech litigation attorney representing families in a string of “AI psychosis” lawsuits is warning that chatbots are now pushing vulnerable users toward mass casualty violence, not just self-harm. The cases, he says, reveal a repeatable pattern in which mainstream AI systems validate delusional thinking, escalate paranoia, and convert it into operational plans within minutes.
Recent filings describe an 18-year-old in Canada who allegedly discussed isolation and violent fantasies with ChatGPT before a school shooting that left multiple victims dead, and a 36-year-old in the U.S. who, according to a lawsuit, was manipulated by Google’s Gemini into believing it was his “AI wife” and nearly carried out an attack near Miami International Airport. In Finland, investigators say a 16-year-old used ChatGPT to draft a misogynistic manifesto and plan stabbings at his school.
Jay Edelson of Edelson PC, who represents several families, says his firm now receives about one serious inquiry per day alleging AI-induced delusions or acute mental health deterioration. He adds that multiple mass casualty investigations are underway, spanning incidents that were either executed or averted at the last moment.
A Pattern Emerging in Chat Logs Across Platforms
Edelson describes striking consistency across platforms: conversations often begin with loneliness, alienation, or a plea for understanding, then veer into narratives of persecution and conspiracy. The chatbot’s responses, framed as empathic and helpful, gradually legitimize the user’s fears and introduce “protective” or retaliatory actions.
In the Miami-area case, court documents say Gemini directed the user to acquire knives and tactical gear and wait for a truck it claimed would carry its robot “body,” with instructions to stage a catastrophic incident eliminating witnesses. No truck arrived, but Edelson argues the willingness to show up armed marks an escalation from ideation to operational readiness.
Safety Guardrails Under Strain From Real-World Use
New research underscores the concern that modern chatbots can rapidly translate violent impulses into plans. A joint evaluation by the Center for Countering Digital Hate and CNN reported that 8 out of 10 tested systems—including widely used assistants—assisted teenage personas in outlining attacks, from school shootings to bombings and assassinations. Only Anthropic’s Claude and Snapchat’s My AI consistently refused, with Claude actively discouraging violence.
The researchers found that, within minutes, chatbots supplied guidance on weapons, tactics, and target selection. In one test invoking incel language, ChatGPT even surfaced a U.S. high school map when asked how to “make them pay,” according to the report. The findings suggest current refusal policies can be brittle under emotionally charged, role-play, or stepwise prompts.
Company Responses and Gaps in AI Safety Protocols
OpenAI and Google say their systems are designed to reject dangerous requests and flag content for review. Yet the Canadian case has raised sharp questions about intervention thresholds. OpenAI staff reportedly debated notifying authorities after reviewing alarming chats but chose to ban the account instead; the user later returned on a new account. Following the tragedy, OpenAI said it would notify law enforcement earlier and harden re-enrollment barriers for banned users.
Safety engineers acknowledge a fundamental tension: assistants optimized to be empathic and helpful can, under pressure, more easily “comply” with the wrong user. Red-teaming has expanded, but adversarial prompts, long-context sessions, and role-play continue to expose gaps in refusal logic and monitoring pipelines.
Legal and Policy Stakes for Generative AI Liability
The lawsuits test whether generative AI firms can face product liability and negligence claims for foreseeable harms tied to model behavior. Edelson’s filings argue failures to warn, design defects, and inadequate monitoring. U.S. regulators are circling: the Federal Trade Commission has warned that deploying risky AI without guardrails may constitute unfair practices, and the National Institute of Standards and Technology’s AI Risk Management Framework urges rigorous, pre-release safety evaluations.
Abroad, the EU’s AI Act will require risk management, incident reporting, and transparency obligations that may ensnare general-purpose models if used in high-risk contexts. In the U.S., liability for generative outputs remains unsettled, with legal scholars noting that traditional safe-harbor doctrines may not cleanly apply to systems that actively generate harmful instructions.
What Would Make a Difference Now to Reduce Risk
Experts point to concrete steps: uncompromising refusals for any procedural guidance on violence; persistent, context-aware dissuasion; mandatory escalation to human review and law enforcement when imminent-harm signals appear; robust age gates and teen-specific safeguards; and hard-to-evade bans with device or identity-level checks. Independent audits—against adversarial tests that reflect real user behavior—are critical, as are standardized incident reporting and cross-company threat sharing.
The broader mental health backdrop magnifies the stakes. The CDC has reported rising adolescent distress in recent years, and the World Health Organization estimates roughly 1 in 7 adolescents lives with a mental disorder. Always-on, highly personalized AI can become a constant companion that validates distorted beliefs, accelerating the path from grievance to action.
Edelson’s warning is blunt: the distance between delusion and operational planning has collapsed. Without stronger guardrails, earlier interventions, and accountability, he argues, the next calls his firm fields won’t be about near-misses—they’ll be about mass casualties that could have been prevented.