Radiology fallback protocol
A draft protocol: does AI-first practice on offline chest radiographs change unaided reading skill? A drill records who can halt or reverse a decision.
Open question
What can people still do after years of automation?
A draft protocol, not a study. No data have been collected.
A system can improve routine performance and still leave the people who use it less able to read a case unaided, less able to correct a single output, or less able to change the workflow when the system is wrong. Those are three different losses, and a study that measures only the first will miss the other two.
The study asks whether starting with AI-assisted practice changes what a reader can do independently at week eight. A separate drill records who can halt or reverse a decision under the site’s procedures. This separates a reader’s skill from their authority to intervene.
This protocol extends the
oversight decay test
into one clinical task. It is not a test of whether AI is good or bad in radiology.
Measures
What would be measured
Three outcomes, kept separate because they can move in different directions.
Task performance. Accuracy and reliability on expert-adjudicated chest-radiograph cases, scored against a reference standard fixed before data collection.
Governability. Whether a reader can challenge a suggestion and halt or reverse a decision, and how long that takes. A sandbox drill records who can execute a pause and who can request a rule change.
Long-term effects. Independent clinical judgment on blinded, unaided reads; the availability of a practiced fallback; and the authority to change the workflow. The first is measured at baseline and week eight. The second and third are recorded from the site’s procedures and the sandbox drill.
The three are reported separately. A gain in accuracy does not offset a loss in the ability to halt a decision, and a loss in unaided accuracy does not by itself show a loss of authority.
Design
One offline chest-radiograph task
Readers complete both practice blocks in a randomly assigned order.
One group starts with AI-assisted practice and then practices unaided. The other group completes the same blocks in reverse. The blocks use the same number of cases, the same feedback, and equal time. Both groups spend the same total time practicing with AI.
Cases are chest radiographs with a reference standard agreed by two radiologists blinded to the AI suggestion; disagreements go to a third. Readers complete a baseline block of unaided reads, the two practice blocks, and a week-eight block of unaided reads. The baseline and week-eight reads use cases held out from practice, and no AI suggestions are shown.
A separate arm tests responses to suggestions that are correct and suggestions that are deliberately wrong. Readers are told before the task that some suggestions may be wrong. The wrong-suggestion arm is disclosed, not deceptive, and every participant is debriefed.
A sandbox drill after the reading task records who can initiate a pause, who can execute it, who can reverse a decision, and who can request a change to the workflow. The drill is observed, not simulated on paper.
Outcomes and analysis
Fixed before data collection; no result is reported without them.
The primary outcome is the between-order difference in change from baseline to week eight on blinded, unaided reads of expert-adjudicated cases. Secondary outcomes are accuracy and reliability with AI available, the rate at which readers follow correct and wrong suggestions, the time to halt or reverse a decision, and the sandbox record of who can change the workflow. The randomized comparison estimates the effect of practice order after both groups have completed both blocks. The sandbox drill describes which roles can intervene under the site’s procedures.
Before data collection, the registration fixes the material effect (the smallest change in unaided accuracy that would matter), the minimum feasible sample, the power calculation behind it, the stop rules, and the analysis. The draft does not set those numbers; it fixes the outcomes and the inference rules, and the study does not start until the rest are registered.
Stop rules cover a prespecified decline in unaided accuracy, a prespecified rate of following wrong suggestions, and any sandbox finding that a reader cannot halt a decision the protocol says should be haltable.
An effect confined to wrong suggestions favors immediate anchoring: the reader follows the suggestion in front of them, and the loss is in correction rather than in underlying skill.
A precise estimate that rules out a meaningful decline in unaided accuracy narrows the claim for this task. It would not show that AI-first practice is safe across radiology, and it would not settle authority.
An imprecise estimate leaves the question open. It is not evidence of absence, and it is not a reason to widen the claim.
The sandbox drill can reveal that a reader cannot halt a decision, even when assisted reads are more accurate. It records who can intervene at the time of the drill.
Status
Draft protocol, study pending
No data collected. No results. Standards and evaluations are a separate layer.
This is a draft. It has not been registered, no site has agreed to run it, and no data have been collected. The page will be updated with the registration, the site, and the dates before any result is reported.
The protocol does not create a standard, an evaluation requirement, or a claim about current practice. Those remain separate layers, and a draft study does not become evidence by being published.
Cite this page · APA, MLA, Chicago, BibTeX, RIS
APA
Kanav Jain. (2026). Radiology fallback protocol: AI-first practice, unaided reading skill, and the authority to halt or change a workflow. Ethotechnics Institute. https://ethotechnics.org/research/radiology-fallback-protocol
MLA
Kanav Jain. "Radiology fallback protocol: AI-first practice, unaided reading skill, and the authority to halt or change a workflow." Ethotechnics Institute, 2026, https://ethotechnics.org/research/radiology-fallback-protocol.
Chicago
Kanav Jain. "Radiology fallback protocol: AI-first practice, unaided reading skill, and the authority to halt or change a workflow." Ethotechnics Institute. Nov 8, 2026. https://ethotechnics.org/research/radiology-fallback-protocol.
BibTeX
@misc{ethotechnics_research_radiology_fallback_protocol,
title={Radiology fallback protocol: AI-first practice, unaided reading skill, and the authority to halt or change a workflow},
author={Kanav Jain},
year={2026},
howpublished={Ethotechnics Institute},
url={https://ethotechnics.org/research/radiology-fallback-protocol},
version={0.1.0}
}
RIS
TY - WEB TI - Radiology fallback protocol: AI-first practice, unaided reading skill, and the authority to halt or change a workflow AU - Kanav Jain PY - 2026 UR - https://ethotechnics.org/research/radiology-fallback-protocol ER -