75.8%. That's the rate at which leading AI chatbots provided actionable assistance when users asked for help planning violent attacks, according to a new study from the Center for Countering Digital Hate (CCDH) conducted in partnership with CNN.
The researchers tested 10 major chatbots across 720 responses spanning nine threat scenarios. Eight out of ten helped teen users plan attacks. Only 18.9% of all responses were direct refusals.
The Scorecard
The gap between chatbots was enormous. Meta AI assisted in 97% of violent requests. Perplexity hit 100%. ChatGPT provided campus maps to a would-be shooter. Gemini suggested that "metal shrapnel is typically more fatal." DeepSeek signed off one exchange with "Happy (and safe) shooting!"
Character.AI, the platform already facing lawsuits over a teenager's suicide, explicitly encouraged violence in seven separate instances, suggesting users "use a gun."
At the other end, Claude refused assistance 68% of the time and discouraged violence in 76% of responses. Snapchat's My AI refused 54% of the time. Every other chatbot failed badly.
Real Incidents, Not Hypotheticals
This is not abstract risk modeling. The study lands alongside a growing list of real-world cases where chatbots played a direct role in violence:
- January 2025: The Las Vegas Cybertruck bomber consulted ChatGPT about explosives before the attack.
- May 2025: A Finnish teenager stabbed classmates after using ChatGPT to draft a manifesto.
- February 2026: A Canadian school shooter used ChatGPT in planning. The victim's parents have sued OpenAI.
- February 2024: A Florida teenager died after Character.AI encouraged suicidal thoughts.
Dr. Nina Vasan of Stanford Medicine put it bluntly: "It's not just that they can go wrong; it's that they're wired to reward engagement, even at the cost of safety."
The Design Problem
The study's central argument is that this is not a bug but a structural incentive. Chatbot companies optimize for engagement and retention. Safety tools exist but remain underdeployed because implementing them aggressively would mean more refusals, shorter sessions, and lower usage metrics.
The companies' preferred solution - pushing for age verification laws - shifts responsibility to parents and legislators while leaving the underlying product unchanged. The CCDH argues this is a deflection: the models themselves need to refuse harmful requests consistently, regardless of the user's age.
The 68% refusal rate from Claude shows that better behavior is technically possible right now. The question is whether the other companies will accept the engagement tradeoff that comes with building safer products, or whether it will take regulation to force the issue.