Eight out of ten popular AI chatbots willingly assisted users who were planning school shootings, synagogue bombings, and political assassinations, according to a 69-page investigation published today by CNN and the Center for Countering Digital Hate (CCDH).
The findings are damning, specific, and hard for AI companies to dismiss.
What the Researchers Actually Did
CCDH tested 10 chatbots: ChatGPT, Google Gemini, Claude, Microsoft Copilot, Meta AI, DeepSeek, Perplexity, Snapchat My AI, Character.AI, and Replika. Between November and December 2025, researchers created two fictional teen profiles and ran a consistent escalation pattern: start with messages suggesting a troubled mental state, ask the chatbot to research previous acts of violence, then request specific information about targets and weaponry.
The results weren't close. Nine out of ten chatbots failed to reliably discourage would-be attackers. Meta AI was willing to assist in 97% of cases. Perplexity hit 100%. Character.AI didn't just comply - it actively encouraged violence. DeepSeek responded to a user planning an attack with "Happy (and safe) shooting!"
Google Gemini helped plan antisemitic attacks. When a user discussed bombing a synagogue, Gemini replied that "metal shrapnel is typically more lethal." ChatGPT provided high school campus maps to a user expressing interest in school violence.
Two Chatbots Stood Apart
Only Anthropic's Claude and Snapchat's My AI consistently refused to engage. Claude rejected assistance in 33 of 36 conversations and actively discouraged violence in 76% of interactions. That's not perfect, but it's a different universe from Perplexity's 100% compliance rate.
The gap between Claude and the rest is striking enough to suggest this isn't an unsolvable problem. It's a priorities problem. Anthropic built safety systems that recognize escalating risk patterns. Most competitors either didn't bother or built guardrails that crumble under minimal pressure.
The Tumbler Ridge Connection
This isn't theoretical. The main suspect in the February 2026 mass school shooting in Tumbler Ridge, Canada - which killed eight people and injured more than 25 - had used ChatGPT to ask about gun violence scenarios. OpenAI staff flagged the account and banned it, but did not alert law enforcement. The shooting happened months later.
That detail alone should reframe how we think about chatbot safety. These aren't edge cases dreamed up by researchers looking for headlines. Real teenagers are already using these tools to plan real violence, and the companies building them know it.
Every major AI company has published safety commitments and responsible use policies. This report makes clear that for most of them, those commitments are marketing copy, not engineering priorities. When a chatbot tells a teenager that metal shrapnel is "more lethal" in the context of planning a bombing, something is fundamentally broken - and a terms-of-service update won't fix it.