There was a significant gap between adolescents and adults: 90 per cent of the cases involving minors had fatal outcomes, compared with 49 per cent for adults. Hudon believes this finding warrants attention, although it too needs to be read with caution.
“Are young people truly more vulnerable to the adverse effects of AI?” he asked. “We believe so, but there’s also a media coverage bias: a case involving a child is more likely to make headlines than one involving an adult.” This bias was evident throughout the analyzed media corpus.
In 74 per cent of the articles, the AI system’s behaviour was cited as the primary cause of the event. “Only six per cent of articles mentioned alternative explanatory factors, such as a pre-existing medical condition, family background or substance use,” said Hudon.
The evidence cited in the media reports most often came from screenshots or family members; medical or police documentation appeared in only 1.6 per cent of the articles.
“The media don’t have access to medical records,” Hudon noted. “They interview the grieving families, who naturally point to the interaction with the chatbot as the trigger. It’s understandable, but it creates a very partial picture that can lead to wrong conclusions.”
It starts with vulnerability
In his clinical practice with young adults at the Institut universitaire en santé mentale de Montréal, Hudon has seen situations in which AI appears to have exacerbated an existing vulnerability. In some cases, patients in a state of delirium used chatbots to question their treatment, and the AI ultimately backed up their doubts about the medical system.
“When someone thinks you want to harm them and AI tells them that’s plausible, a false belief takes root,” Hudon said. “Then my colleagues and I have to convince the patient that the treatment is valid, and AI is wrong. It’s a real therapeutic challenge.”
Hudon believes the problem isn’t so much AI itself, “but rather the intersection of certain states of vulnerability and the specific characteristics of these tools, which are always available, lack judgment and are programmed to simulate empathy. What makes chatbots appealing can, in some situations, worsen the very problems they seem to alleviate.”
Hudon also noted that platforms such as Character AI and Replika, where users can interact with customizable avatars, present a distinct risk profile, especially for teenagers accustomed to identifying with animated characters. The anthropomorphic interaction fosters more intense attachment, and hence potentially riskier engagement.
Not all tools are created equal
Hudon does not condemn chatbots across the board, but believes it is important to distinguish between different uses of AI.
“Some tools, such as Woebot and Wysa, have undergone rigorous scientific validation,” he said.
“They have clear objectives, limits that are explicitly communicated to users, and protocols for referral to specialized resources in cases of distress. They are not therapists, and they don’t claim to be.
"But they are guided by the data. This is a fundamental difference between them and a general-purpose chatbot which you can ask about anything, including suicidal thoughts.”
Towards structured oversight
The study concludes that the harms documented in the media reports are not purely technological, but relational and structural: “They arise at the intersection of user vulnerability, the persuasive design of the platforms, and the absence of formal guardrails,” said Hudon.
He believes addressing the issue must be a shared responsibility: “Doctors must better define what is and isn’t appropriate, while the software developers have an obligation to inform users of their products’ limitations in dealing with mental health issues and to establish crisis management protocols. Governments must also establish coherent governance frameworks.”
The immediate priority is better data. “We need mechanisms to systematically track these events, because hospitals do not yet code psychiatric crises according to the use of chatbots,” Hudon said. “For now, the media are our only window on this trend—a window which, by its nature, overrepresents the most dramatic cases.”
“The purpose of our study wasn’t to prove that AI causes these events,” Hudon concluded. “What we need to know is whether the media reports are pointing us to real clinical risk profiles. To find out, we need the tools to analyze the full picture and, if need be, adjust our practices accordingly.”
