AI exposure: Psychiatrists

Psychiatrists is one of the unusual cases where capability and friction are both high. Current systems can produce a great deal of this work, with average capability across its tasks at 2.3 out of 4. Only 25.5% of the weighted task load comes out as exposed, because 32.8% of it runs into something structural. Jobs in this band tend not to disappear. They change shape, and the person stays.

What holds the line here is verification cost. Across this occupation's 12 tasks it averages 2.57 out of 3, the highest of the five friction dimensions. In plain terms, checking the output costs more than producing it. Where an undetected error is expensive, dangerous or irreversible, the economics change. Someone has to verify the work, and verifying can cost as much as doing it. This is the friction most likely to fall as tools for checking improve.

The most exposed thing this job does is Review and evaluate treatment procedures and outcomes of other psychiatrists or medical..., at 66.7%. The least is Serve on committees to promote or maintain community mental health services or delivery systems, at 0.0%. A gap of 66.7% between two parts of the same job is the reason this index publishes at task level. An occupation-wide number would have hidden both.

Within healthcare practitioner and technical occupations, this one is more exposed than most. The median across the 89 roles in the group is 18.4%, and only 10 of them score higher than this. Occupational families are not uniform, and the spread inside them is often wider than the gap between them.

What would move this score. Of 12 tasks, 2 are currently banded exposed, 7 assisted and 3 untouched. For that distribution to shift materially would take a change in who is permitted to sign the work, which is a question for regulators rather than for engineers. The score is re-computed every quarter against a fresh capability reference, and the change is published rather than quietly applied.