AI Can Transcribe — But Can It Digest?
Yesterday I read someone talking about the power of AI can be harnessed in medicine. He said something that was quite interesting: “AI scribes are a distraction”. Recently, an article published in NEJM concluded that AI scribes have only a modest improvement in productivity. Based on my experience as a clinician, I must agree with both assessments. It does not take much time for me to write the consultation notes. What takes more time is the endless box ticking and form filling that follows the consultation, even in a public healthcare setup such as the NHS. Unless these processes are automated, AI scribes by themselves will have limited impact on the productivity of clinicians.
On a related note, I had another very interesting conversation with a therapist about the role of AI scribes. I have prototyped a simple automated AI scribe pipeline and he had very kindly provided me with an audio note for testing. He offered very sophisticated feedback that kind of feeds into the above paragraph.
His view was that AI scribes are very good for capturing what happened during the session. He was not worried about the accuracy of the transcription, technical competence nor note-writing per se. However, something deeper bother him. Psychotherapy notes are not just records, they are part of the thinking and digestion process of the therapist. Good therapy does not just live in what is said during the session. It lives in what is understood later. If I understood him correctly, he felt insight is non-linear and meaning evolves over time. A therapist’s inner process is part of the clinical material. Unfortunately, with the current models of AI scribes, the summary is frozen at the end of the session, leading to premature closure, over-simplification and false certainty.
That led me to think that what he is actually wanting is a second order AI. He wants a tool that elevates AI from transcription and summarisation tool to a reflective companion that helps the therapist’s evolving formulation.
So, what does this mean for HearMeNow?
In version 1, HearMeNow is doing accurate transcription, structured summary and saves times. This is crucial from the point of cognitive load, admin fatigue and documentation burden.
He has pointed me to a clinical intelligence layer (something I have been thinking about since I conceptualised HearMeNow). Ultimately, what I would like to do with HearMeNow is this:
- Initial AI summary- Version 1
- Therapist reflection later- Version 2
- After supervision- Version 3
- After rupture/repair- Version 4.
Eventually, I would want AI to continuously synthesise the material from all versions of each session to become a longitudinal meaning making tool than a static note-taking tool. This idea feels closer to the spirit of psychotherapy than a static note.
I am very grateful for his feedback. His feedback has genuinely shaped how I think about the next phases of HearMeNow.