Deskilling
Unaided performance after exposure.
What deskilling means
Unaided performance after exposure.
Why it matters. Relying on AI can weaken unaided skill.
Radiologists miss more findings unaided after months of AI use.
How to measure it
Gather local scans
Use scans from your own scanners and protocols.
Adjudicate labels
Radiologists agree the reference reading.
Test across sites and groups
Check scanners, sites and patient groups separately.
Check calibration
Compare predicted and observed rates.
Performance at the vendor's test site often doesn't hold at yours.
Local sensitivity and specificity with subgroup tables, external-site results and calibration.
Report local performance with radiologist-adjudicated labels, external sites, subgroup results and calibration, against a cleared comparator.
Reviewed by the Precision research team.
How strong is the evidence
E1 · Peer-reviewed, regulator, audit or court source. Backed by a peer-reviewed paper, a regulator, an audit or a court record. Usually read through a summary, not the full text.
What the sources say
Colonoscopy ADR 28.4% to 22.4% (observational) (Warwick).
- Real-world outcomes. Needs outcome data from live use, such as time saved, burnout or patient results.
- Warwickwrap.warwick.ac.uk
Found something out of date? Write to hello@precision.example.
Agents that measure this
Quick answers
What is deskilling in healthcare AI?
Unaided performance after exposure. It sits in the imaging and diagnostics group.
Why does deskilling matter?
Relying on AI can weaken unaided skill. Radiologists miss more findings unaided after months of AI use.
How is deskilling measured?
Gather local scans, then Adjudicate labels, then Test across sites and groups, then Check calibration. Report local performance with radiologist-adjudicated labels, external sites, subgroup results and calibration, against a cleared comparator.
What should I watch out for with deskilling?
Performance at the vendor's test site often doesn't hold at yours. A good report gives: local sensitivity and specificity with subgroup tables, external-site results and calibration.
Can deskilling be automated?
Only in part. It needs real-world outcomes. It depends on data from real use, not a test set.