Table 1 Synopsis of the characteristics of the two argumentative types
Clinician as systemic trustee (“the one relying”) | Clinician as individual expert (“the one controlling”) | |
|---|---|---|
Clinical validity of the AI-CDSS (i.e. scientific evaluation) | positive proof of benefit as a necessary precondition | positive proof of benefit as a necessary precondition |
Attitude towards error rates or potential harm | errors are part of medicine with and without the use of AI-CDSS | errors are part of medicine, but the clinician needs to mitigate these errors for individual patients as well as possible |
Role of the clinician | user of an application that statistically causes the least harm | user of an application whose errors must be controlled by him/her critical questioning and review of AI-CDSS outputs to avoid errors consideration of each patient’s context and potentially neglected factors |
Accountability for harm caused by AI-CDSS | The clinician bears no accountability because AI-CDSS’s use was evidence-based indicated. The patient needs to be informed about potential harms beforehand. | Overreliance on AI-CDSS recommendations is culpable, leaving clinicians accountable for some errors that should have been identified previously. Final responsibility for treatment remains with the clinician. |
Goals and ways to achieve it | minimising harm by using AI-CDSS outcomes | minimising harm and the potential harm arising from AI-CDSS outputs through clinician review |
Necessary level of understanding | sufficient understanding to deal with the patient’s information needs, i.e. − knowledge of the advantages and disadvantages/risks − knowledge of a regulatory process for reviewing clinical validity | sufficient understanding to evaluate AI-CDSS outputs as well as deal with the patient’s information needs, i.e. − knowledge of the advantages and disadvantages/risks − knowledge of a regulatory process for reviewing clinical validity − basic information technology knowledge (“How does machine learning work?”) − knowledge about underlying dataset and its limitations − understanding the conclusion of a recommendation (“How does the AI-CDSS arrive at this outcome?”) − in some cases, enough previous clinical experience, being able to make supported decisions even without AI-CDSS |