AI in the Therapy RoomAugmenting Couples Counseling

A couple describes last week's fight from memory, and the therapist works with two competing summaries. Meanwhile the whole argument, and the six days of silence after it, sit in a chat thread neither of them thought to mention. That gap is where AI is genuinely useful in couples work, and this post is about what it can do there, what the research supports, and where it should stay out.

What couples therapy is trying to do

Couples therapy helps two people understand each other, argue better, and stop repeating the same fight. Most of that depends on the therapist noticing patterns: who withdraws, who escalates, what "fine" actually means. The catch is that the therapist sees one hour a week and the couple reports the other 167 from memory.

A model that has read the whole conversation can describe it without taking sides. The therapist can then do what a model cannot: decide what the pattern means for these two people.

Three jobs AI can already do for a therapist

What the research actually shows

The evidence for AI in therapy is early, and most of it is about improving therapists' tools rather than replacing them. Gabor-Siatkowska and colleagues (2023) used ChatGPT to generate extra training data for a therapeutic dialogue system built for psychiatric patients, and improved its intent recognition by 13%. Read the study. That is a modest, practical gain of the kind you should expect: AI making the instruments sharper, with a clinician still holding them.

The model for using observation in therapy is much older. Gottman's lab predicted marital outcomes from coded recordings of couples arguing, then turned those observations into teachable skills such as soft start-ups and repair attempts. Chat analysis applies the same idea to text: code what actually happened, then teach from it.

Two cautions from adjacent research. When ChatGPT was asked to rank relationship advice, its judgments barely tracked human ones and changed between runs. And the first randomised trial of a generative-AI therapist showed meaningful symptom reductions for individuals with depression and anxiety, which is a different task from mediating between two people.

Where AI should stay out

• Interpreting meaning. A long reply gap can be a night shift or a cold shoulder; the model can time it, not explain it.
• Crisis and safety. Any sign of coercion, abuse or self-harm belongs with the clinician, immediately.
• One partner analysing the other. Uploading a chat without your partner's knowledge is surveillance, not therapy. The research on phone access in couples is clear that access should be mutual and consensual.

How to bring your own texts into a session

1. Agree together. Both of you export the chat, both of you see the results.
2. Look at three things, not thirty: who initiates, how reply times have changed, and where the tone timeline dips.
3. Bring the pattern, not the screenshots. "Our tone drops every time we plan a visit to my parents" is a therapeutic conversation. A stack of quoted messages is a trial.
4. Let the therapist decide what matters. If a pattern surprises you, that is a topic. If it is obvious, it is confirmation.

Frequently asked questions