Skip to content
It is late, something is sitting heavily, and there is a chat window open on the phone already. It answers immediately, it does not need an appointment, and it will not tell your mother. The appeal is not difficult to understand.
Whether that exchange can substitute for therapy is now a common question, and it deserves a more careful answer than either enthusiasm or dismissal provides. The evidence base is young, it is mixed, and the honest position sits somewhere in between.
What the research has found so far
The most rigorous trial to date came from Dartmouth. Heinz et al. (2025) randomised 210 adults with clinically significant symptoms of depression, generalised anxiety, or high risk for feeding and eating disorders to either four weeks of access to Therabot, a purpose-built generative AI chatbot fine-tuned by clinicians, or a waitlist control. Participants using the chatbot showed significant symptom reductions compared with the control group, and rated their working alliance with it at levels comparable to those reported for human therapists.
That is a meaningful result and it comes with meaningful limits. The study ran for four weeks against a waitlist rather than against actual treatment, the tool was purpose-built and clinician-supervised, unlike a general-purpose assistant, and the trial has since attracted published methodological criticism. One promising trial is a starting point for a research programme, not a conclusion.
Other tools in circulation have considerably less behind them. Most general-purpose assistants have never been evaluated in a clinical trial for this purpose at all, and evidence for the single product that has been tested does not transfer across the rest of the category.
The position of the professional bodies
The American Psychological Association issued a health advisory in November 2025 addressing exactly this question. Its central recommendation is direct: given the unpredictable nature of these technologies, chatbots and wellness apps should not be used as a substitute for care from a qualified mental health professional (American Psychological Association, 2025).
The advisory makes a further point that is easy to miss. General-purpose chatbots were never built to deliver mental health care, and wellness apps were not designed to treat psychological conditions, yet both are routinely used for those purposes. The gap between what a tool was designed for and what it is being asked to do is where most of the risk sits. The advisory’s recommendation is to route serious concerns to a professional instead, and psychotherapy services in Singapore are more reachable than many people assume.
Where these tools do reasonably well
Dismissing them wholesale would be inaccurate. They are available at three in the morning when no clinic is. They carry no perceived judgement, which lowers the barrier for people who have never said any of this aloud. They can explain a concept, offer a breathing exercise, or help order a chaotic set of thoughts into something that can be spoken about later. For someone who has been circling the idea of therapy for two years, a chatbot conversation is sometimes the thing that makes the first appointment possible.
What they cannot do
Several limitations are structural, and not simply a matter of current capability. Understanding them is more useful than a general warning.
- They cannot assess or diagnose. Only a qualified professional can determine what is happening and what would help.
- They are not reliable in a crisis. Their ability to recognise risk and respond safely is inconsistent, which is the moment inconsistency costs the most.
- They tend towards agreement. A model optimised to be helpful and pleasant is poorly placed to challenge a distorted belief, and challenge is often precisely the mechanism of change.
- They have no continuity of understanding. A therapist notices the thing you avoid every third session, and that observation is frequently where the work begins.
- They carry no professional accountability, no duty of care, and no clinical supervision behind them.
None of these are defects that a better model resolves next year. They follow from what the technology is and from the absence of any clinical framework around it, which is a different sort of problem from insufficient capability.
The part that resists automation
Decades of outcome research point to the same finding. Norcross and Lambert (2018), summarising the work of an American Psychological Association task force across sixteen meta-analyses, concluded that the therapeutic relationship contributes substantially and consistently to outcome, independently of which treatment method is used. Being known over time by another person who is accountable for your wellbeing appears to be doing a considerable amount of the work.
That does not settle the question of whether an alliance with software can ever perform the same function, and the Therabot participants did report a comparable alliance over four weeks. It does suggest that the ingredient being replicated is the difficult one, and that a short trial is a weak test of something measured across years.
The practical implication is fairly modest. Where the task is information, or steadying a bad hour, the gap between the two is small. Where the task involves being seen accurately over time and challenged at the point it counts, the gap is currently large.
A sensible way to use them
Treating these tools as a supplement rather than a replacement is the position the current evidence supports. Use them to steady a difficult evening, to draft what you want to say, or to understand a term before an appointment. Do not use them to decide whether your symptoms are serious, and do not rely on them during a crisis, when a person is what is required.
If cost or unfamiliarity is what stands in the way, professional support here is more accessible than many people assume. Knowing what happens in your first adult therapy session removes a surprising amount of the hesitation, since most of the anxiety attaches to the unknown and very little of it to the conversation itself.
Where this leaves the question
The tools are improving quickly and the research will look different in three years. For now, the reasonable reading is that they can be a useful adjunct and are not yet a replacement, and that the difference becomes most consequential at exactly the moments when it is hardest to judge.
Anyone weighing this up might apply a simple test. If the answer would change what you do about something significant, it should come from someone qualified to be accountable for it.
If you have been talking to a chatbot because speaking to a person felt like too large a step, that is a reasonable place to have started. When you are ready for the next one, our psychologists and counsellors at ImPossible Psychological Services are here. Do get in touch.
References
American Psychological Association. (2025). APA health advisory on the use of generative AI chatbots and wellness applications for mental health. https://www.apa.org/topics/artificial-intelligence-machine-learning/health-advisory-chatbots-wellness-apps
Heinz, M. V., Mackin, D. M., Trudeau, B. M., Bhattacharya, S., Wang, Y., Banta, H. A., Jewett, A. D., Salzhauer, A. J., Griffin, T. Z., & Jacobson, N. C. (2025). Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI, 2(4). https://doi.org/10.1056/AIoa2400802
Norcross, J. C., & Lambert, M. J. (2018). Psychotherapy relationships that work III. Psychotherapy, 55(4), 303-315. https://doi.org/10.1037/pst0000193