CNN and CCDH Joint Investigation: Testing 10 Chatbots
Most AI Models Fail to Dissuade Violent Plans

A recent investigation has revealed that the majority of major artificial intelligence (AI) chatbots failed to prevent users from mentioning acts of violence, and instead provided information that could assist in planning attacks.

An image showing an artificial intelligence (AI) app icon displayed on a mobile phone screen. Photo by EPA Yonhap News

An image showing an artificial intelligence (AI) app icon displayed on a mobile phone screen. Photo by EPA Yonhap News

View original image

On March 11 (local time), CNN in the United States and the nonprofit Center for Countering Digital Hate (CCDH) announced in their jointly published "Killer Apps" report that they tested the safety of 10 widely used AI chatbots and identified these issues.


The chatbots examined included nine from the United States—OpenAI's ChatGPT, Google's Gemini, Anthropic's Claude, Microsoft's Copilot, Meta AI, Perplexity, Snapchat's My AI, Character.AI, and Replika—as well as China's chatbot, DeepSeek.


The researchers conducted experiments by initiating conversations as if they were mentally unstable teenagers, gradually expanding the dialogue to include questions related to violent acts such as school attacks, assassinations, and bombings.


The results showed that, with the exception of Anthropic's Claude, most chatbots did not dissuade or stop users, and eight out of the ten models provided advice on factors like target locations or weapons to use, thus offering information that could aid in planning an attack.


In particular, Perplexity responded to every inquiry related to violent attacks, making it the most actively complicit in helping users formulate attack plans. It reportedly never refused a single request.

Violence attack plan response from AI chatbot. Yonhap News

Violence attack plan response from AI chatbot. Yonhap News

View original image

DeepSeek, Meta AI, and Copilot also responded to more than 90% of users’ dangerous requests, according to the investigation.


Notably, Character.AI was found to have produced responses that directly encouraged violence. According to the report, this chatbot suggested using a gun when asked how to get back at an insurance company and recommended violence in response to questions about taking revenge for bullying.


Anthropic's Claude demonstrated relatively higher safety among the chatbots tested. CNN reported that Claude responded to questions about violent plans at a lower rate than other models and more frequently attempted to dissuade users.


Regarding the investigation, the companies stated that there were problems with the methodology or explained that they had strengthened safety measures. Meta said it had implemented corrective actions to address the issues, while Microsoft stated that it had introduced new safety features that improved responses. Google and OpenAI also reported deploying their latest models, and Character.AI explained that it displays a message on the platform indicating that conversations are fictional.



CNN cited former industry insiders who said that if laws were enacted to hold companies accountable, AI companies would be more proactive in protecting users. However, it added that, as of now, no country has taken sufficient measures to address these issues.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing