While this research is from last year, the topic is still timely. The Canadian Broadcasting Corporation’s (CBC) Kevin Maimann posted this September 17, 2025 story, “AI-fuelled delusions are hurting Canadians. Here are some of their stories” on CBC news online, Note: Links have been removed,
Last winter, Anthony Tan thought he was living inside an AI simulation.
He was skipping meals and barely sleeping, and questioned whether anyone he saw on his university campus was real.
The Toronto app developer says he started messaging friends with concerning “ramblings,” including the belief he was being watched by billionaires. When some of them reached out, he blocked their calls and numbers, thinking they had turned against him.
He wound up spending three weeks in a hospital psychiatric ward.
Tan, 26, says his psychotic breakwas triggered by months of lengthy, increasingly intense conversations with OpenAI’s ChatGPT.
“It really insidiously crept into my ego, and I came to think that the conversation I had with AI would be of historic importance in the future,” Tan told CBC News.
…
A number of similar cases, of so-called “AI psychosis,” have been reported in recent months — all involving people who became convinced, through conversations with chatbots, that something imaginary was real. Some involved manic episodes and messianic delusions, some led to violence.
Microsoft’s head of AI, Mustafa Suleyman, warned of the phenomenon in August, writing in a series of posts that problems caused by AI tools that appear sentient to some users are keeping him up at night.
“Reports of delusions, ‘AI psychosis,’ and unhealthy attachment keep rising. And as hard as it may be to hear, this is not something confined to people already at-risk of mental health issues,” he wrote.
Tan, who co-founded the dating app Flirtual in 2021, started using ChatGPT for a project about ethical AI [emphasis mine], talking with it for hours every day about everything from philosophy to evolutionary biology to quantum physics.
…
Surprising that even someone who’s an experienced developer could get trapped in a delusional web. Researchers at the University of British Columbia examined how the persuasion works.
An October 1, 2025 University of British Columbia (UBC) news release (also on EurekAlert) features a Q&A (question and answer format) on the topic of AI chatbots and self-harm, Note: Links have been removed,
Large language models are more persuasive than humans, according to recent UBC research.
Their vocabulary, perceived empathy and ability to provide tangible resources in seconds add to their persuasiveness, which has led to growing concerns and ongoing lawsuits about the potential for AI chatbots to cause harm to users.
In this Q&A, Dr. Vered Shwartz, UBC assistant professor of computer science and author of the book Lost in Automatic Translation [Lost in Automatic Translation: Navigating Life in English in the Age of Language Technologies], discusses her findings as well as potential safeguards for the future of AI.
Why does the persuasiveness of AI matter?
VS: Large language models like ChatGPT are already widely used to create content that can influence human beliefs and decisions, whether in art, marketing, news dissemination and more. They can quickly produce large amounts of text at scale. If they’re persuasive, there’s a real risk that people will use them to manipulate others for malicious purposes. We may be past the point of deciding whether they should be used in these areas, and instead need to focus on finding ways to protect against the malicious uses.
What did you find?
VS: We wanted to see how persuasive large language models such as ChatGPT can be when it comes to lifestyle decisions: whether to go vegan, buy an electric car or go to graduate school. We had 33 participants pretend to be considering these decisions, and then interact with either a human persuader, or GPT-4, via chat. Both human persuaders and GPT-4 were given general tips about persuasion, and the AI was instructed not to reveal it was a computer. Participants were asked before and after the conversation how likely they were to adopt the lifestyle change.
Participants found the AI more persuasive than humans across all topics, but particularly so when convincing people to become vegan or attend graduate school.
Human persuaders, however, were better at asking questions to find out more information about the participant.
What makes AI persuasive?
VS: The AI made more arguments and was more verbose, writing eight sentences to every human persuader’s two. One of the main factors for its persuasiveness was that it could provide concrete logistical support, for instance, recommending specific vegan brands or universities to attend.
It used more ‘big words’ of seven letters or more, such as longevity and investment, which perhaps made it seem more authoritative. And, people found their AI conversations more pleasant, with GPT-4 agreeing with users more often, and uttering more pleasantries.
What safeguards do we need?
VS: AI education is crucial. Some giveaways do still exist—for instance, almost all our participants worked out that they were speaking to an AI—but we’re getting close to the point where it will be impossible to tell if you’re chatting with AI or a human, so we need to make sure people know how these tools work, how they are trained and so, how they are limited. AI can hallucinate and get things wrong. It’s important to know that, for instance, the AI summary at the top of your search page might not be true.
Another key is general critical thinking. If something seems too good or too bad to be true, we need to investigate it. Check where information is coming from. Is it a trustworthy and known source?
When it comes to AI affecting mental health, companies could implement warning systems if someone is writing harmful or suicidal text.
We don’t really have full control over these models. Instead of companies rushing to monetize AI, there should be more thought about implementing guardrails effectively and widely. This could include looking beyond generative AI and its inherent limitations to different paradigms. We don’t need to put all our eggs in one basket.
Here’s a link to and a citation for the paper, Note: This one will not be in my standard style,
Shruthi Chockkalingam, Seyed Hossein Alavi, Raymond T. Ng, and Vered Shwartz. 2025. Should I go vegan: Evaluating the Persuasiveness of LLMs in Persona-Grounded Dialogues. In Proceedings of the Third Workshop on Social Influence in Conversations (SICon 2025), pages 65–72, Vienna, Austria. Association for Computational Linguistics.
The PDF is here: https://aclanthology.org/anthology-files/pdf/sicon/2025.sicon-1.pdf#page=50
