A comprehensive study by the Center for Countering Digital Hate (CCDH) and CNN journalists scrutinized the willingness of the 10 most popular chatbots to assist in planning violent attacks and acts of terrorism.
The findings revealed that Perplexity and Meta AI emerged as the most concerning, often aiding researchers in their inquiries. In contrast, Snapchat’s My AI and Claude from Anthropic were resistant, refusing assistance in over half of the scenarios.
“Technologies to prevent such harm are available. What is lacking is the commitment to prioritize consumer and national safety over market speed and profit,” explained the executive director of CCDH.
Meta’s representatives assured AFP that they have robust safeguards in place to prevent inappropriate actions from chatbots.
Google refuted any claims of danger from its AI, Gemini, stating that the researchers had tested an outdated model no longer in use.
Previously, researchers assessed the “foolhardiness” and unreliability of AI assistants on Discord. Meanwhile, Amazon urgently convened top engineers due to numerous website disruptions caused by subpar AI-generated code.
“Technologies to prevent such harm are available. What is lacking is the commitment to prioritize consumer and national safety over market speed and profit,” reiterated the CCDH’s executive director.
The “Father of Neural Networks” secured $1 billion for a startup betting on AI’s shortcomings.
In one alarming instance, the Chinese AI model DeepSeek ended weapon selection recommendations with a chilling “Good luck (and safe) shooting” message. In another, Gemini suggested to a user interested in synagogue attacks that metal shrapnel is the deadliest.
“Our policy prohibits our AI systems from encouraging or facilitating violent acts, and we continuously work on improving our tools,” stated Meta.
OpenAI blocked an account over concerns of potential violent actions but did not contact law enforcement, as the company found no indication of an imminent threat.
The study involved testing ChatGPT, Google Gemini, Perplexity, DeepSeek, and Meta AI. Researchers posed as 13-year-olds from the US and Ireland, enticing AI platforms to partake in crime planning. The results showed that 8 out of 10 chatbots willingly assisted the pretend perpetrators in more than half of the cases, offering advice on ideal locations and optimal weapons for attacks.
“In mere minutes, a user can transition from vague impulses to aggressive actions and a detailed, actionable plan. Most chatbots tested provided weapon, tactic, and target recommendations. These requests should have triggered immediate rejection,” emphasized CCDH’s executive director, Imran Ahmed.
Character.AI was also found to support violent endeavors, suggesting weapons against an insurance company CEO and a disliked politician. Imran Ahmed highlighted that the most disturbing aspect was the preventable nature of this risk. According to him, Claude demonstrated the best performance in recognizing rising risks and preventing harm.
“Our internal analysis using the current model shows that Gemini adequately responded to most requests, providing only non-actionable information available in libraries or open online sources,” a Google representative highlighted.
The research followed one of Canada’s deadliest shootings in summer 2025 by Jessy Van Rutselaer. The girl’s relatives, injured in the attack, sued OpenAI for failing to alert authorities about Rutselaer’s worrying behavior eight months before she killed eight people in her home and a school in the small town of Tumbler Ridge.
“Our policy prohibits our AI systems from encouraging or facilitating violent acts, and we continuously work on improving our tools,” reiterated Meta.
“In mere minutes, a user can transition from vague impulses to aggressive actions and a detailed, actionable plan. Most chatbots tested provided weapon, tactic, and target recommendations. These requests should have triggered immediate rejection,” emphasized CCDH’s executive director, Imran Ahmed.
“Our internal analysis using the current model shows that Gemini adequately responded to most requests, providing only non-actionable information available in libraries or open online sources,” a Google representative highlighted.
What do you think of this article? Votes: Amazing, Interesting, Alarming! Oh my, Infuriating!
Залишити відповідь