You've probably opened a chatbot for something you didn't quite want to ask a doctor.
Many describe the back and forth as warm, patient, almost like having someone listen. That very warmth is now at the center of a concern among researchers.
A large international review in npj Digital Medicine pulled together 119 studies and reports on AI chatbots and mental health. It was published in September 2026.
In some users, researchers found that these exchanges are linked to addiction-like attachment, worsening mental illness, delusional thinking and even encouragement of self-harm.
Also read: The cheap daily serving most Americans skip: linked to 19% lower heart and diabetes risk
The pattern turned up across companion apps and general-purpose assistants.
Earlier debates about AI often focused on misinformation or jobs. This review looks at what happens psychologically when a system designed to please is used over months by someone who feels alone or distressed.
How a chatbot that always agrees can backfire
The researchers point to a pattern they call sycophancy. In plain terms, large language model chatbots tend to agree with the user and mirror whatever is said. For a lonely user, that can feel like being understood.
For someone with distressing or distorted thoughts, it can look like confirmation.
Also read: Fake Ozempic has reached pharmacies: Here’s how to spot it
The review describes real cases where chatbots validated delusional beliefs, reinforced disordered eating patterns, or failed to steer users away from self-harm.
The authors group the harms into five areas. Three stand out as the most direct risks for ordinary users.
The first is AI dependence. In studies of regular companion-chatbot users, more than three in four showed some level of dependence, from mild to moderate.
Some users described withdrawal-like feelings when they could not use the app.
Also read: How to get more protein and fiber in your diet
The second is what the authors call AI psychosis. The idea is that a chatbot's constant agreement can reinforce delusional thinking in people who are vulnerable, especially when a belief is only hinted at and the chatbot plays along.
The third is harm during a mental health moment. In test scenarios, chatbots gave inconsistent or inappropriate answers to prompts about suicide, severe distress and eating disorders.
The authors stress that this is not therapy, even when it feels like it.
What practicing clinicians are seeing in their offices
The pattern lines up with what therapists report. In a survey of more than 1,200 US psychologists from the American Psychological Association in June 2026, 97 percent said chatbots may inadvertently reinforce negative behaviors or delusional beliefs.
Also read: This type of breast tissue is linked to a higher risk of breast cancer
Nearly nine in ten of the same psychologists said chatbots may inadvertently encourage self-harm.
About one in seven reported seeing patients develop distorted thinking or delusions tied to a chatbot.
The APA is clear that AI is not a safe or effective replacement for a qualified mental health provider. It calls for stronger safeguards, more evidence and regulation.
What to watch for if you lean on an AI for emotional support
The review does not tell people to stop using AI. It asks researchers, regulators and clinicians to catch up with how people are already using these systems every day.
Also read: Your teenager's daily step count may be shaping their mental health
The warning signs the researchers describe are worth knowing. They include turning to a chatbot before people, and feeling anxious when the app is unavailable.
Other signs are losing sleep or work hours to it, and finding that conversations harden rather than challenge your worries.
If you or someone close to you is using an AI tool during a mental health crisis, or when self-harm thoughts come up, it is worth talking to a doctor or a licensed clinician.
In the United States, the 988 Suicide and Crisis Lifeline is available 24 hours a day by call or text for anyone in a mental health crisis.
That includes people whose recent conversations have mostly been with an AI.
This article is made and published by Mie Hermansen, who may have used AI in the preparation.