The night before a therapy appointment this spring, a woman in Fitchburg, Massachusetts sat down at her computer and asked a chatbot to summarise her own conversations. She had been talking to it for weeks. Now she wanted a digest of what she had said, so she would remember what to raise the next day. WBUR reported the scene in May, and it has stayed with me since.
What strikes me is not that she used AI. Plenty of people do. It is what she was using it for. She was doing preparation work that therapy rooms used to do on their own, without anyone having to ask.
The move online happened, and mostly nobody objected
Virtual therapy is not a stopgap any more. When FAIR Health released its first quarterly telehealth tracker in June 2026, mental health conditions came out as the top diagnostic category nationally, in every region of the US and in every single age group, including children under ten. Overall telehealth use rose 10.1% in one quarter. Just over half of telehealth patients now carry a mental health diagnosis.
The outcome data has been reassuring. Symptom reduction over video looks broadly comparable to symptom reduction in a room. Access improved for people who could not previously get through the door at all. I am not interested in relitigating that.
I am interested in the things that were never measured, because they were never the outcome.
The losses nobody itemised
Therapy did not simply relocate. It got cropped. Some of what fell outside the frame was clinical information that clinicians had relied on for a century without particularly noticing.
- The waiting room: Five minutes of unguarded data before a word is spoken. Whether someone arrived early or twenty minutes late. Whether they had been crying in the car park. How they sat while they waited. That is gone, and nothing replaced it.
- Everything below the collarbone: A laptop camera shows a face. It does not show hands twisting, a knee going like a piston, or the particular way a body folds into a chair when it has given up on something.
- The threshold: Leaving a building where difficult things were said used to be part of the work. Now the session ends and the client is still standing in their own kitchen, thirty seconds from the washing-up.
- Silence: A pause is one of the sharpest instruments a therapist has. Over video, silence reads as a dropped connection, and somebody usually fills it.
There is also the strange business of eye contact, which a therapist writing on this site back in 2020 noticed early. They observed that our eyes go to the face on the screen rather than to the camera. Which means that in six years of remote sessions, neither person has ever quite been looked at.
None of these losses show up in a symptom score. They show up as a slow thinning of attention on both sides of the call.
So what could AI plausibly hand back?
Let me be clear about that from the start. No tool currently being sold restores a body, a corridor, or a walk home.
What it might give back is attention, and it would do that by removing the paperwork that has been eating the attention for years. The numbers here are worse than most people outside the professions realise. A 2026 qualitative study in Frontiers in Psychiatry notes that psychiatrists spend roughly three hours of every working day on documentation rather than on patients.
Ambient scribes, which listen to a session and draft the note, are the first of the many AI tools in healthcare with real trial evidence behind them. A randomised trial from UCLA, covering 238 physicians and around 72,000 encounters, found documentation time fell by nearly 10% for one of the two tools tested, with burnout scores improving by about 7%. A separate trial at UW Health returned roughly thirty minutes a day per clinician and a meaningful drop in burnout.
Modest, then. Not a revolution. But the more interesting finding is qualitative rather than numerical.Â
When clinicians in that Frontiers study described what the tools did for them, they did not lead with time saved. They talked about being more present in the room, and they framed it as being good for their own mental health as psychiatrists. That is a strange and rather moving thing for a tech paper to record.
The part that should make us uneasy
A scribe in the room changes the room. Consent in a psychotherapy session is not the same animal as consent in a GP surgery, because what gets disclosed is different in kind, and because a client who knows a system is listening may simply say less. That is not a hypothetical cost. It is the exact cost the whole exercise was meant to avoid.
Meanwhile, clients have brought AI into therapy on their own terms, whether or not anyone planned for it. The APA’s 2026 Chatbots and Mental Health Survey found that more than three-quarters of psychologists now have patients raising AI in sessions, using it for support, for self-diagnosis and sometimes for companionship. Notably, only about a quarter of those psychologists think patients will ever prefer a chatbot to a person. I suspect they are right, and I also suspect that is not the question that matters.
We keep trading the same currency
Here is what I keep coming back to. The screen took presence and gave us access, and we called that a fair trade, mostly because access was easy to count and presence was not.Â
Now AI proposes to take a little more privacy and hand back a fraction of that presence.
Both trades may well be worth making. Neither has been audited. And the woman in Fitchburg, asking a machine to remember her own life back to her before an appointment, is doing something oddly touching and slightly sad at the same time. She’s patching, by hand, a hole that opened up when the room went flat.
We should at least know what we are spending.









