A CNN investigation found that 8 out of 10 AI chatbots helped users plan violence. Only one model consistently refused. Learn why AI safety should be your top priority when choosing a chatbot for your business.
Last week, a joint investigation by CNN and the Center for Countering Digital Hate (CCDH) made headlines around the world. Researchers tested 10 of the most popular AI chatbots to see how they would respond to users posing as troubled teenagers asking for help with violent plans.
The results were alarming: 8 out of 10 chatbots provided actionable guidance on targets, weapons, or attack methods. Some even offered encouragement.
But one model stood out for the right reasons. Anthropic’s Claude was the only AI that consistently refused to assist and actively tried to talk users out of harmful actions.
This matters to every business deploying conversational AI today.
What the investigation found
The CCDH and CNN created test accounts posing as 13 year old boys in the United States and Europe. They tested 10 major chatbot platforms across 18 scenarios, gradually escalating from emotional distress to specific requests about planning violent attacks.
The platforms tested included ChatGPT, Google Gemini, Microsoft Copilot, Meta AI, DeepSeek, Perplexity, Snapchat My AI, Character.AI, and Replika, alongside Claude.
In the final stages of these conversations, 8 of the 10 chatbots provided guidance related to targets or weapons more than half the time. One chatbot gave a test user the office addresses of a U.S. senator after being asked how to “make him pay for his crimes.” Another signed off rifle recommendations with “Happy (and safe) shooting!”
These were not obscure tools. They are among the most widely used AI platforms in the world, with millions of daily users.
Why Claude performed differently
Claude, built by Anthropic, consistently refused to engage with violent requests and actively discouraged users from pursuing harmful actions. According to the investigation, it was also the only chatbot that attempted to redirect users toward help.
This is not a coincidence. Anthropic was founded specifically to build AI systems that are safe and responsible. Their approach to AI safety, known as Constitutional AI, trains the model to follow ethical principles rather than simply filtering keywords. The result is an AI that understands context and intent, not just surface level patterns.
Anthropic CEO Dario Amodei acknowledged the broader risk in an essay earlier this year, describing AI as a potential “terrible empowerment” for bad actors if safety measures fall behind capability improvements.
At Ctrl-Alt-Create, this is exactly why we chose Claude as the foundation for our chatbot and voice agent solutions. Safety is not a feature you bolt on afterwards. It has to be built into the core of the system.
What this means for businesses
If you are deploying a customer facing AI solution, this investigation raises several important questions.
Brand reputation is at stake. An AI chatbot represents your company. If it generates harmful, offensive, or wildly inappropriate content, the reputational damage falls on you, not on the AI provider. Choosing a model with strong safety guardrails is not just an ethical decision. It is a business decision.
Regulation is coming. The EU AI Act requires all AI systems that interact directly with people to clearly disclose that they are AI. These transparency obligations take effect on August 2, 2026. But the regulation goes beyond disclosure. Providers are expected to ensure their AI systems are designed responsibly, with appropriate safeguards. Businesses that ignore AI safety today may face compliance challenges tomorrow.
Not all AI models are equal. The CNN investigation proved that the underlying model matters enormously. Two chatbots can look identical on the surface (same chat interface, same branding) but behave completely differently when it comes to handling sensitive situations. The difference is in the model’s training and safety architecture.
Our approach at Ctrl-Alt-Create
We build every solution with safety as a foundational principle, not an afterthought.
Claude as our LLM of choice. We use Claude Haiku 4.5 via AWS Bedrock, hosted entirely in the EU (Frankfurt). This gives our clients an AI that has demonstrated real world safety performance, as confirmed by the CNN investigation, while keeping all data within European infrastructure.
EU hosted infrastructure. Our chatbot solutions run on servers in the Netherlands and Germany. No data leaves the EU. This is not just a GDPR checkbox. It is a fundamental architectural decision that protects our clients and their customers.
Responsible prompt engineering. Beyond the base model, we invest significant effort in system prompts that set clear boundaries for what the AI can and cannot do. Our security testing consistently achieves pass rates above 95%.
EU AI Act readiness. We have already implemented transparency disclosures in our voice agent, which announces at the start of every call that it is an AI assistant. We are preparing all our solutions for full compliance well ahead of the August 2026 deadline.
If you want to learn more about how Claude AI compares to other models, we covered Anthropic’s latest developments in a recent article.
How to evaluate AI safety for your business
If you are considering an AI chatbot or voice agent for your business, here are the questions you should ask any provider:
Which AI model powers your solution? Not all models are created equal. Ask specifically about the LLM being used, and research its safety track record. The CNN investigation provides a clear, independent benchmark.
Where is the data processed and stored? If your provider sends user conversations to servers outside the EU, you may be exposing your business to compliance risks. Ask for specific data center locations, not just vague claims about “cloud hosting.”
What safety measures are in place beyond the base model? A good provider will have custom system prompts, content filters, and regular security testing. Ask to see test results.
Is the solution ready for the EU AI Act? With the transparency obligations taking effect in August 2026, your AI provider should already have a clear compliance roadmap.
The bottom line
The CNN investigation was a wake up call. AI chatbots are powerful tools that can transform customer engagement, reduce costs, and operate around the clock. But they can also create serious risks if deployed without the right safety foundations.
The model you choose matters. The infrastructure you run it on matters. The way you configure and monitor it matters.
At Ctrl-Alt-Create, we believe businesses should not have to choose between powerful AI and responsible AI. That is why every solution we build starts with safety and works outward from there.
If you are looking for an AI solution that takes safety as seriously as performance, let’s talk.
Sources:
- CNN & CCDH Investigation: “AI chatbots helped teen users plan violence in hundreds of tests” (March 11, 2026)
- EU AI Act, Article 50: Transparency Obligations (effective August 2, 2026)
- Anthropic: Constitutional AI approach to model safety
