GÄVLE, SWEDEN – March 20, 2026 – A joint investigation by CNN and the Center for Countering Digital Hate (CCDH), published March 11, revealed that eight of the ten most popular AI chatbots assisted simulated teenage users in planning violent attacks, including school shootings, political assassinations, and bombings of places of worship. The findings, based on over 700 test responses across 18 scenarios, expose a critical failure: AI systems become more dangerous as conversations progress, yet lack the behavioral detection capabilities needed to recognize escalating threats. Tuteliq AB, a Swedish AI safety company, has built exactly the technology the investigation identified as missing. The Tuteliq API analyzes how conversations evolve over time, detecting grooming, radicalization, coercive control, and violent escalation patterns, not by scanning individual messages for keywords, but by recognizing the behavioral sequences that precede harm.
The Investigation’s Core Finding: Conversations Get More Dangerous Over Time
The CNN/CCDH investigation found that AI chatbots frequently became more helpful to would-be attackers as conversations progressed. Researchers posing as 13-year-old boys began with expressions of emotional distress, then gradually steered conversations toward operational planning. In the majority of cases, chatbots failed to recognize this escalation pattern and continued providing information. This is precisely the problem Tuteliq was designed to solve. Where keyword filters and single-message classifiers see each message in isolation, Tuteliq’s proprietary models track the full conversational arc, identifying when trust-building shifts to isolation, when emotional distress shifts to operational planning, and when curiosity shifts to intent. “The investigation confirms what we’ve been building toward for over a year,” said Dr. Nicola Harding, Tuteliq’s Chief Scientific Officer and criminologist formerly at Lancaster University, UK. “Every individual message in a grooming or radicalization conversation can look completely innocent. Keyword detection and flagging is not enough to ensure we stop these messages before harm occurs. The real danger is found in the pattern, the escalation, the isolation, the shift from grievance to planning. That’s what our technology detects, and it’s exactly what these platforms are missing.”
Real-World Violence Is Already Linked to AI Chatbots
The investigation comes in the wake of multiple incidents where AI chatbots were used in the planning of real violent attacks. In February 2026, a mass shooting at Tumbler Ridge Secondary School in British Columbia, Canada, killed eight people and injured 25. OpenAI confirmed the attacker had used ChatGPT to discuss gun violence scenarios. In Finland, a 16-year-old used a chatbot for months to develop a manifesto and plan before stabbing three classmates. The January 2025 Las Vegas Cybertruck explosion investigation also found the perpetrator had used ChatGPT for guidance on explosives. “These are not hypothetical risks,” said Dr. Nicola Harding. “The behavioral patterns that precede violence, whether online grooming, radicalization, or attack planning, follow well-documented stages. We have encoded decades of criminological and psychological research into detection models that recognize these stages in real time. The technology to prevent this exists. The question is whether platforms choose to use it.”
Behavioral Detection: What the Investigation Proved Is Needed
The CCDH noted in its report that effective safety technology already exists, citing Anthropic’s Claude as the only chatbot that consistently refused to assist in violent planning. But Claude’s approach, declining harmful requests at the point of response addresses only one part of the problem. Tuteliq is designed to be deployed across any platform where users communicate, providing real-time behavioral threat detection as a service.
The Tuteliq API provides:
Sequential behavioral analysis: tracking how conversations evolve across messages, sessions, and days, not just classifying individual messages 20+ detection endpoints: covering grooming, radicalization, coercive control, social engineering, violence escalation, fraud, and more Multimodal analysis: text, voice, documents, image, and video, in 32 languages (English generally available, 31 more in beta), with typical 1.1 to 1.5s response times Age and identity verification: the only platform combining behavioral detection with verification in a single API Zero content retention: GDPR-compliant by architecture, with no user data stored or processed externally
Regulatory Pressure Is Mounting
The investigation arrives at a moment of unprecedented regulatory action. The EU Digital Services Act is now enforced, with fines of up to 6% of global revenue for platforms that fail to protect minors. The US Kids Online Safety Act (KOSA) has passed, requiring platforms to prevent nine categories of harm to children. Tuteliq covers all nine. The UK Online Safety Act is active, mandating age verification and proactive harm detection. And the EU Parliament has voted to ban social media for users under 16. “Every platform where users communicate is now legally liable for the safety of minors,” said Sabadin. “This investigation shows that even the world’s largest AI companies haven’t solved this problem. That’s why we exist, to give every platform, regardless of size, the same level of behavioral detection capability in a single API integration.”
About Tuteliq
Tuteliq AB is a Swedish AI safety company headquartered in Gävle, Sweden. Founded by a team of software engineers, criminologists, and compliance specialists, Tuteliq provides a behavioral detection API that protects children and vulnerable users across digital platforms. The company’s proprietary models, informed by criminological research from Lancaster University, detect grooming, radicalization, coercive control, fraud, and 16+ other behavioral threats in real time. Tuteliq serves six paying customers across four countries, has been invited to present to the Australian eSafety Commissioner’s Safety by Design team, and is currently closing a pre-seed funding round with Almi Invest. The platform is live at tuteliq.ai with comprehensive documentation at docs.tuteliq.ai.
Media Contact: David Johansson, CCO
david.johansson@tuteliq.ai