Mental health: chatbots show screening potential, not an ability to replace clinicians

A review of 18 studies in adults

In May 2025, the researchers searched for studies evaluating chatbots or artificial intelligence conversational agents for mental health screening and assessment. The search covered four bibliographic databases and focused exclusively on adults.

The review includes 18 studies published between 2021 and 2025, with sample sizes ranging from 20 to 3,902 participants. It distinguishes between rule-based chatbots, generative language models and hybrid approaches combining several techniques. These tools therefore do not constitute a single intervention whose performance can be generalized.

Encouraging results, but not universal accuracy

For depression and anxiety, the abstract reports sensitivities of 0.84 to 0.93 and specificities of 0.80 to 0.96 with generative models. Sensitivity describes the ability to identify people who have the disorder according to the reference standard used. Specificity describes the ability to identify those who do not have it.

These ranges are not presented as an overall estimate applicable to any chatbot. The abstract does not detail the comparators, assessment conditions or uncertainty surrounding each result. They therefore cannot predict the reliability of a particular commercial tool.

Rule-based chatbots also show good measurement consistency and favorable acceptability. However, a consistent measurement is not, on its own, evidence of an accurate diagnosis.

Acceptability alone does not demonstrate a clinical benefit

Most studies report high user satisfaction, while barriers persist among older populations. Being satisfied with an interaction does not mean that the disorder was correctly identified or that care improved.

Methodological quality is heterogeneous, with a moderate risk of bias in important areas. The abstract does not provide enough detail to determine which tools or populations account for the best performance.

The authors advocate use as a complement to clinical judgment, with professional supervision, explicit clinical protocols and measures to reduce algorithmic bias. The review focuses on identification and assessment, not on demonstrating effective treatment.

What we can take away

The findings are encouraging for helping with screening in adults. They do not justify entrusting a diagnosis or a care decision to a chatbot alone.