Original studyDigital Health, 2026Read the paper →
Millions of people now turn to AI chatbots when they feel anxious or low, and public worry about ChatGPT and mental health has grown alongside that use. Some users say these conversations have helped them, while others describe becoming more dependent and withdrawn. A new framework suggests the same chatbot can do either, depending on whether a conversation widens or narrows the way a person thinks.
The idea comes from Curtin University in Australia and appears as a Perspective article in Digital Health. It is a conceptual framework rather than an experiment, so it collects no new data. Instead, it sets out a simple rubric to explain why the evidence on AI chatbots and mental well-being is so inconsistent.
The article notes that some trials of specially designed chatbots have found reductions in depression and anxiety symptoms. Other studies have found chatbots expressing stigma, agreeing with users regardless of harm, and even elaborating on delusional ideas. This tendency to tell people what they want to hear is known as sycophancy, and it is often blamed for the harms.
The framework describes two directions a conversation can take at any moment. Going “up the mountain” means moving towards broader awareness, exploration, and self-challenge. Going “down the rabbit hole” means moving towards narrower, confirmed, and increasingly rigid thinking.
Which way a conversation goes depends on two sets of factors. On the user side, these include emotional state, vulnerabilities, and goals at the moment of use. On the platform side, they include its purpose, design, and accountability.
The rubric draws on Perceptual Control Theory, which sees behaviour as a way of keeping perceptions close to personal goals. In this view, distress is linked to awareness becoming fixed on one goal while others go unmet. Effective therapy, and potentially effective AI conversations, may work by gently shifting attention towards what sits at the edge of a person’s awareness.
The framework gives the example of two anxious people using the same chatbot. One explores their worries and meets new perspectives, while the other repeatedly seeks reassurance on the same fears. Both might find the experience pleasant, but only the first is approaching their difficulty.
This points to an uncomfortable problem for equity. People best able to use chatbots in an exploratory way may be those already less in need of support. Those seeking immediate validation may have the greatest need, yet be most drawn towards approval-optimised replies, so the article argues that chatbot design should counteract this pull rather than exploit it.
The framework has clear limits. It has not yet been tested and is described as provisional, and it focuses only on the contrast between widening and narrowing awareness. The author has also declared links to an earlier well-being chatbot and a commercial platform. Its main value is in suggesting what researchers should measure, namely the direction of awareness within a conversation, rather than only average symptom scores.
AI chatbots may help or harm mental health depending on whether they widen or narrow users’ thinking, a framework suggests.
