Question 1 of 15
An R2 has used AI nodule detection on every chest CT for 12 months. When moonlighting without AI, she misses a 9mm spiculated nodule she has never independently searched for. This is:
Never-skilling. The resident never built independent nodule search patterns. AI did the searching from day one.
Question 2 of 15
An attending mammographer has read independently for 15 years. After 2 years of AI-assisted CAD, her unassisted accuracy has dropped. This is:
Deskilling. The attending had well-established skills that eroded through disuse. The cognitive architecture exists but has weakened.
Question 3 of 15
After 6 months with an AI tool that overcalls consolidation on poorly inspired films, a resident starts overcalling consolidation on well-inspired films without AI. This is:
Mis-skilling. The AI's systematic error pattern has been internalized. The resident now makes the same error independently.
Question 4 of 15
In the Dratsch et al. mammography study, when AI provided incorrect BI-RADS suggestions, experienced radiologists' accuracy dropped from approximately:
Experienced radiologists' accuracy dropped from 82% to 46%. Inexperienced readers dropped as low as 20%.
Question 5 of 15
The "calibration paradox" refers to:
To oversee AI, you need calibration — knowing when AI is right vs. wrong. But calibration develops through independent practice, the exact process AI disrupts.
Question 6 of 15
In Ke et al.'s framework, at which level should trainees first encounter AI with deliberately embedded errors?
Level 2 (Adversarial Calibration) involves deliberate errors. Trainees must detect, explain, and correct them.
Question 7 of 15
Which training phase (Ke et al.) maps to R1-R2 in radiology residency?
Phase 1 maps to R1-R2. Build independent pattern recognition before AI exposure.
Question 8 of 15
Which subspecialty has the HIGHEST risk across all three AI competency threats?
Mammography has high risk across all three categories because its AI tools provide answer-delivery, operate in a high-stakes screening context, and have known systematic biases.
Question 9 of 15
A resident consistently agrees with AI's "normal" reading without independent review. An attending later finds a subtle finding the AI missed. This in-the-moment behavior is:
Automation bias — uncritically accepting AI output in the moment.
Question 10 of 15
The Pesapane et al. study found that explainable AI (XAI) with saliency heatmaps reduced automation bias from 36.1% to:
XAI with heatmaps reduced automation bias from 36.1% to 17.8%.
Question 11 of 15
The key difference between deskilling and never-skilling is:
Deskilling means existing cognitive architecture has weakened from disuse. Never-skilling means the architecture was never built.
Question 12 of 15
Adenoma detection rate dropped from 28.4% to 22.4% after AI-assisted tools were removed (Budzyn et al.). This finding demonstrates:
This suggests deskilling — practitioners showed reduced performance when the tool was removed.
Question 13 of 15
Which strategy best protects against deskilling during residency?
Regular AI-free reading sessions maintain independent skills while still allowing AI integration.
Question 14 of 15
An AI tool has a known false-positive rate for rib fractures at costochondral junctions. After extensive use, a resident starts overcalling fractures at these locations without AI. The best description is:
Mis-skilling. The AI's systematic false-positive pattern has been internalized. The error persists without AI.
Question 15 of 15
Decision attribution — documenting when you agree with, override, or supplement AI — primarily helps with:
Decision attribution builds calibration — the ability to know when AI is right vs. wrong.
Answer all 15 questions to submit.