What Happens When You Turn to AI for Mental Health Support?

People are opening up to AI chatbots about things they wouldn’t tell anyone else – so why do so many stops talking to them just weeks later?

Millions of people worldwide live with depression or anxiety and simply can’t get help — not enough therapists, not enough time, not enough money. The global shortage of mental health support had led to rapid expansion in the use of generative artificial intelligence (GenAI) chatbots as a therapist. A team at the University of British Columbia, publishing in npj Digital Medicine, set out to answer a question nobody had fully mapped before. What happens when people actually use these chatbots? (1^✔ ✔Trusted Source
Generative AI mental health chatbots: a scoping review of intervention design and user experience

Go to source

)
The researchers found that the users reported high initial satisfaction and therapeutic bond, comparing it to a human therapist, but the long-term engagement, real-world clinical value and safety of the user remain questionable.

Inside the Study: 21 AI Mental Health Chatbots Across 11 Countries

Researchers screened nearly 1,900 studies and included 21 in the study, spanning from almost 11 countries, all published between 2023 and 2025. They assessed how large language model (LLM) platforms built around cognitive behavioural therapy (CBT) and non- embodied text interfaces, handled conditions like depression, anxiety, stress, and in some cases dementia, eating disorders, and Post traumatic stress disorder (PTSD).

What they found first was genuinely encouraging: people liked talking to these bots. Some rated the connection almost as strong as one with a real therapist — comparable trust, comparable sense of being understood.The comfort was real — but so was the drop-off

AI Chatbot Engagement Drops Sharply After the Honeymoon Phase

Users generally rated these chatbots as accessible, non-judgmental, and available 24/7 having qualities that drove genuinely high satisfaction scores in several studies. Some users even reported trust levels built by the AI chatbots felt remarkably similar to a human therapist.

However, engagement rarely stays for the long term. In one 8-week NHS-linked study, app usage dropped from 79% in the first week to just 19% by week six. Another chatbot saw a similar decline — over half of users disengaged within the same window. Nearly 300 AI mental health chatbots were studied worldwide — yet in some trials, more than half of users quietly stopped talking to them within just a few weeks

Reviewers linked this to repetitive conversation patterns, generic-feeling responses, and lack of human connection, like trust eroding the moment a chatbot said something that felt off, inaccurate, or emotionally tone-deaf.

What Makes AI Mental Health Chatbots Trustworthy

Researchers said that as the digital health developers are moving beyond generic LLMs like ChatGPT to building cutting-edge tools which can influence user trust and safety.

Few design studies stood out. Platforms built on real clinical knowledge and Retrieval-Augmented Generation (RAG) like VCounselor or Ana and not just providing generic AI responses earned more trust and clinical accuracy. Proving that underlying clinical content like CBT remains the primary driver for mental health support.

Chatbots that remembered context, adapted tone, or let users personalise the experience – even something as simple as choosing a name or conversation style tended to hold attention longer, as did bots that lived where people already were – inside WeChat, Telegram, or Slack.

Voice, art generation tool (ArtTheraCat), even a PTSD chatbot delivered through augmented reality exposure therapy (ExpandXR) performed well, providing richer formats to hold attention longer than plain text alone.

Are AI Therapy Chatbots Safe and Effective?

The researchers are careful not to oversell this. They say the warmth people felt didn’t always translate into better outcomes, as most evidence here is early, and only one study looked at real-world long-term use. What’s clear is that AI support can genuinely feel helpful in the moment, but over-reliance and inappropriate emotional dependence pose a greater bioethical risk to the user, and doing it safely remains the harder problem.

Reference:

1. Generative AI mental health chatbots: a scoping review of intervention design and user experience – (https://www.nature.com/articles/s41746-026-02972-0)

Source-Medindia