The interesting question is what the person does next.

A user may like a tool because its answer is quick and easy to read. That does not establish whether it improves judgement. We propose studying how professional reviewers respond to supported, incomplete and misleading outputs in a controlled setting. The central question is whether the interaction helps a person discriminate between an answer they can use, an answer requiring correction and a question that should be escalated or left unresolved.

Candidate conditions could compare outputs with source links, explicit uncertainty or a request to examine a conflicting finding. The study would ask whether these features alter checking and escalation, not simply whether they are visible. Deliberately flawed examples, if ethically and methodologically appropriate, would be confined to a controlled evaluation with a suitable debrief. No live care decision should depend on a planted error or an unassessed experimental interface.

The existing unaided or differently supported review process would provide the comparison. Experience, familiarity with the topic and time pressure may change how people rely on assistance, so these factors would be recorded. Useful measures could include detection of consequential errors, correction quality, unsupported acceptance and total task effort. A faster answer with worse error detection would not be reported as an overall improvement merely because the average completion time fell.

NIST's AI Risk Management Framework supplies voluntary risk-management context (NIST, 2023). NHS DTAC supplies separate buyer-assurance context for relevant digital health technologies (NHS England, n.d.). Neither reference establishes that PHARMORIS has passed an assessment. The proposed study's intended purpose, data use, professional roles and potential consequences would determine its governance and assurance work. Academic independence and a clear record of system versions would be important if the findings are to remain interpretable.

A bounded first collaboration would select one review task and a small number of interaction conditions, then agree meaningful failure thresholds before collecting results. We would proceed only if support helps appropriate checking without unacceptable burden or unsafe reliance. Persistent consequential errors, misleading confidence or an inability to separate interface effects from participant familiarity would trigger redesign or a stop. The aim is evidence for a better professional experience, not a claim that oversight solves every AI risk.

The interaction should also make disagreement usable. A reviewer could be asked to explain why they rejected an output and what evidence would change their position. Examining those explanations would help distinguish productive challenge from confusion, while revealing whether an apparently reassuring source link was actually consulted or merely treated as a signal of authority.