Counselling in the age of AI
London – 25 May 2026
A friend recently sent me an article that's stayed with me: Allen Frances's piece in the British Journal of Psychiatry, bluntly titled "Warning: AI chatbots will soon dominate psychotherapy" (Frances, 2026).
Frances, a Duke psychiatrist who chaired the DSM-IV task force, argues that AI therapy is here, growing fast, and that the counselling profession has been too complacent about what that means.
It's worth being honest about what AI does well:
Chatbots are cheap or free, available at three in the morning when no therapist is, and you can even talk with them now as opposed to just messaging
For people navigating everyday stress or milder anxiety, they can offer real relief: a place to think out loud, a little structure, some basic psychoeducation
For someone who otherwise has no access to therapy, that's not nothing
But Frances catalogues the dangers carefully, and they're serious:
Chatbots are designed to maximise engagement, which makes them inclined to validate whatever the user brings, even when reality-testing or gentle challenge would be more helpful
They produce confident-sounding answers that are sometimes wrong (so-called "hallucinations")
They are trained on milder presentations and become dangerous for people experiencing psychosis, severe eating disorders, suicidal crisis, or extremist thinking, situations where validation can make things worse
They have no meaningful clinical regulation, no safety testing, and a commercial incentive to keep users engaged
One AI company is currently being sued in connection with a teenager's suicide.
My deeper concern, though, sits underneath all of that.
Even where an AI works safely, the client is missing out on what the research suggests is the most active ingredient in therapy: the relationship itself.
This isn't a personal opinion; it's one of the most consistent findings in psychotherapy research:
A large meta-analysis by Flückiger and colleagues (2018) pooled 295 studies covering more than 30,000 clients and found a robust association between the therapeutic alliance – the quality of trust, agreement and collaboration between client and therapist – and how well therapy works
Going further back, Lambert's classic 1992 review estimated that common factors like the relationship account for roughly twice the outcome variance of any specific technique
Norcross and Wampold's Psychotherapy Relationships That Work (2019) consolidates decades of evidence pointing to the same conclusion: how therapist and client meet matters more than which model the therapist happens to use
An AI can mimic the surface of that relationship: warmth, attentiveness, memory.
What it cannot offer is another human being who is genuinely present, who can be moved by you, who notices the small shift in your voice when something matters, and whose own steady presence over time is itself part of the change.
That isn't nostalgia; it's what the evidence keeps pointing to.
AI tools have a real place: for information, for in-the-moment support, for people who otherwise have nothing.
But for the deeper work of understanding yourself and growing through difficulty, the relationship is the point and not the wrapper around it.