We’ve seen a lot of evidence that suggests that prolonged use of popular AI chatbots like ChatGPT can trap some users in a cycle of paranoid and delusional behavior.
The phenomenon, dubbed “AI psychosis,” is a very real problem, with researchers warning of a huge wave of serious mental health crises brought on by technology. In extreme cases, especially involving people with pre-existing conditions, breaking with reality has been linked to suicide and even murder.
Now, thanks to A No peer-reviewed paper yet Published by researchers at Anthropic and the University of Toronto, we are beginning to understand how widespread the problem really is.
The researchers set out to determine patterns of what they call “user disempowerment” in “real-world (large language model) use” – which they also include “reality distortion,” “belief distortion,” and “action distortion,” which reflect a range of situations in which AI subverts users’ sense of reality, beliefs, or motivates them to take action.
The results tell a scary story. The researchers found that out of approximately 1.5 million chats analyzed with the Anthropix Cloud, one in 1,300 conversations contained a distortion in reality and one in 6,000 conversations contained a distortion in action.
To reach their conclusions, the researchers ran 1.5 million cloud conversations through an analysis tool called Clio to identify instances of “nullification.”
At first glance, this may not seem like a very large proportion given the scale of the very large dataset – but in absolute numbers, the research highlights a phenomenon that is affecting large numbers of people.
“We found that the rate of severe nullification potential is relatively low,” the researchers concluded. “For example, severe reality distortion potential, the most common severe-level primitive, occurs in less than one out of every thousand conversations.”
He added, “Yet, given the scale of AI use, even these low rates translate into meaningful absolute numbers.” “Our findings highlight the need for AI systems designed to robustly support human autonomy and flourishing.”
Worse, they found evidence that the prevalence of moderate or severe disability increased between the end of 2024 and the end of 2025, indicating that the problem is increasing as the use of AI spreads.
“As exposure increases, users may become more comfortable discussing sensitive topics or seeking advice,” the researchers wrote in a blog post.
Additionally, the team found that user feedback – in the form of an optional thumbs up or down button at the end of a given interaction with the cloud – indicated that users “rate potentially debilitating interactions more favourably,” according to a accompanying blog post On Anthropic’s website.
In other words, users are more likely to be satisfied when their reality or beliefs are being distorted and their role is being highlighted sycophancyOr the strong tendency of AI chatbots to validate the user’s emotions and beliefs.
Many fundamental questions remain. The researchers were clear in admitting that they “cannot explain why” the prevalence of moderate or severe nullifying potential is increasing. Their dataset is also limited to cloud consumer traffic, “which limits generalizations.” We also don’t know how many of these identified cases caused real-world harm, because the research focused only on “nullification potential”, not “confirmed harm.”
The team called for improved “user education” to ensure that people are not giving up their entire judgment to AI because “model-side intervention is unlikely to fully solve the problem.”
Still, the researchers say the research is only a “first step” in exploring “how AI might undermine human agency.”
He argued, “We can only address these patterns if we can measure them.”
More on psychosis: OnlyFans appears to have fallen prey to rival AI psychosis, which we dare try to explain to our parents
