Two separate comparisons
Does correcting some AI errors establish that oversight preserves the performance people achieve without assistance? Nadine Schlicker and colleagues’ heart-sound classification experiment makes these separate success criteria especially tangible. The result I find most instructive is that participants corrected some erroneous recommendations while combined performance still fell below unassisted human performance when the system became unreliable. Counting only the correct answers added to the system could overlook correct decisions people lose while using it. When assessing an oversight claim for a clinical tool, I would want both comparisons: what people add to the model, and how the model changes the decisions people would otherwise make.[1]
The preregistered experiment randomly assigned 48 participants to two groups of 24. The sample included final-year medical students, residents and experienced physicians. Their task was to classify 40 heart-sound recordings as normal or abnormal. One group received the recommendation before making an initial decision. The other made an unassisted decision and rated confidence before seeing the recommendation and being allowed to revise the answer. The second arrangement therefore permits a comparison before and after advice. Its endpoint needs to remain explicit: the classification of a recording was measured, without patient interaction, additional clinical information or treatment outcomes.[1]
Recommendation accuracy was set to 90 per cent, 60 per cent, 50 per cent and 0 per cent across four blocks. Participants corrected some erroneous advice as reliability declined. Yet combined human–AI accuracy fell from 81.45 per cent in the first block to 58.33 per cent in the last. Requiring an independent decision first did not improve performance overall. I would not extend that finding into a claim that clinicians cannot exercise oversight under any conditions. The narrower result is useful enough: intervening against incorrect recommendations did not ensure preservation of unassisted performance in this experiment. An oversight evaluation should report those two outcomes separately rather than allowing one to stand in for the other.[1]
Deterioration order and measurement limits
The sequence of deterioration creates another limit on interpretation. Reliability always declined in the same order, and the final block deliberately mismatched the audio recording and visual display. The entire decline in that block cannot therefore be attributed to inaccurate recommendations alone. Order and time effects remain possible, and the contribution of the visual mismatch was not separately isolated. Random assignment helps compare receiving advice before or after an initial decision within this task. It does not isolate every form of malfunction from the others. I would therefore avoid using the final difference between groups as a general ranking of workflow safety in clinical practice.[1]
Only 12.5 per cent of participants spontaneously reported declining reliability; that rose to 37.5 per cent when they were explicitly asked. Confidence in their own decisions declined earlier than trust in the system. This divergence makes separate measures of self-confidence and system trust valuable. The timing alone cannot establish why participants behaved this way. Initial information about the system, task organization and the presentation of deterioration are possible factors for further investigation. In a subsequent study I would also track which initial decisions change, and whether each revision turns a correct answer into an error or an error into a correct answer. That analysis could expose decision losses concealed by an overall trust score.[1]
The practical question I take from this experiment is which comparator defines successful oversight. In a new evaluation I would compare unassisted and combined decisions on the same task distribution, and vary the deterioration sequence to test order effects separately. Correction of bad advice, reversal of initially correct decisions and time spent deciding would be separate outcomes. These are proposals for extending measurement, not clinical remedies established by this study. The step from laboratory classification to patient benefit requires additional research. Beyond adding a human approval stage, I would look for an experiment capable of showing under which conditions that approval preserves the quality of the person’s initial judgment.[1]