Can an AI review therapy sessions well enough to support supervisors?
A small US study found its best open model could classify cognitive-behavioural therapy skills with 0.72 balanced accuracy and distinguish most skills in an ongoing trial. Thirty gold-rated transcripts and three feedback raters leave clinical benefit unproven.
The short version
- What happened
- The development dataset contained 318 de-identified sessions from 45 people, but only 30 transcripts had expert gold labels; 10 of those were double-rated, with Cohen's kappa of 0.71.
- Who it affects
- Therapists and supervisors could receive more frequent, structured transcript review, but false positives, missed context and polished feedback may misdirect training. Human review remains essential before any effect on care is known.
- What remains unproven
- The gold-labelled set was 30 selected transcripts, with only 10 double-rated; the work is an unreviewed preprint.
Keep in mind
The gold-labelled set was 30 selected transcripts, with only 10 double-rated; the work is an unreviewed preprint.
The ongoing trial involved 22 people and one therapist delivering both arms, and its transcripts had no item-level expert labels.
Only three experts rated feedback, with weak agreement on relative preferences; high absolute scores do not establish behaviour change.
The source reports no separate funding or competing-interest statement and does not measure patient outcomes from AI-assisted supervision.
Source analysis · 1 linked source. Primary analysed source date: 20 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The model was tested against labels and an external trial
The team began with 331 psychotherapy sessions, excluded 13 and retained 318 sessions from 45 participants. Experts gold-rated 30 transcripts across 60 CBT skill categories, producing 1,800 binary judgments; 10 transcripts were independently double-rated and reached kappa 0.71. The other 288 sessions helped the system retrieve examples and estimate uncertainty but did not expand the gold-label denominator.
Three open-weight models were compared. The strongest setup used Qwen3.5-9B, asked for evidence before a category decision and scored 0.72 balanced accuracy with a 95% bootstrap interval of 0.70–0.74; its Matthews correlation was 0.45, interval 0.41–0.49. A simpler full-transcript baseline scored 0.61 and 0.27 respectively. These are moderate classification results, not an automated replacement for a clinical supervisor.[1]
Promising validation stopped short of patient benefit
The researchers next processed 102 sessions from an ongoing trial involving 22 people: 57 CBT sessions and 45 comparison sessions, delivered by the same therapist. Four of five model-scored skill dimensions were significantly higher in CBT; guided discovery rose but was not significant. This known-groups test supports construct validity, yet the trial transcripts lacked item-level expert labels, so it cannot establish classification accuracy on that dataset.
Three expert users also made 30 pairwise comparisons and 30 absolute ratings of generated feedback. Retrieval-augmented feedback did not significantly beat basic feedback, and agreement on relative actionability and clarity was slight or near chance, although both versions received high absolute ratings. Audio had not been retained, so tone and pacing were invisible. A larger multi-site, prospectively labelled study must test whether feedback changes supervisor behaviour, therapist practice or patient outcomes.[1]
The distinction that matters
What each validation step can show
The study combines three different tests, each answering a narrower question.
Gold labels
Compare 1,800 model judgments with expert ratings
Estimates category classification
Known groups
Compare CBT and control sessions
Tests whether scores move in an expected direction
Expert feedback
Ask three raters to judge generated supervision text
Assesses perceived usefulness, not clinical effect