To define human–AI divergence visibility and explain why clinical AI systems should capture how clinicians and AI interact when their outputs differ.
Approach:
Conceptual framework: The article describes divergence events across structural (workflow and interface), informational (what each party knows), and behavioral (how clinicians engage with AI reasoning) axes.
Definition: Human–AI divergence visibility is visibility into the interaction when a clinician’s judgment differs from an AI output, including what each contributed.
Key Findings:
Clinicians and AI may reach different conclusions because they use different information: clinicians draw on bedside examination and communication, while AI systems process structured records and learned patterns.
Divergence can reflect asymmetric information, genuine differences in interpreting shared evidence, or an error by either party; the article therefore distinguishes divergence from disagreement.
Existing interfaces often reduce interaction to accepting, overriding, or modifying an AI output and do not document how the clinician engaged with it.
The article cites a primary-care study in which clinicians adopted 58% of harmful recommendations and 22% of beneficial recommendations from a large language model-based decision-support tool.
The article cites evidence that physicians with formal AI-output evaluation training had 14% lower diagnostic accuracy when allowed to consult large language models.
Technical explanations of AI outputs do not, by themselves, explain why clinician and AI judgments diverged; post hoc explanations may also create false confidence in erroneous outputs.
Interpretation:
The article argues that examining the interaction at divergence points can reveal the distinct contributions of clinicians and AI, whereas focusing only on which output was correct may miss why their conclusions separated.
Limitations:
The supplied context is truncated during the discussion of the three-axis framework, so it does not show the article’s full definitions or any proposed visibility metrics.
The excerpt cites external studies but does not provide their methods or other details needed to assess their findings or generalizability.
Conclusion:
The article defines human–AI divergence visibility as visibility into the interaction when a clinician’s judgment differs from an AI output, including what each contributed. It argues that examining these divergence events can help distinguish differences in available information, differences in interpreting shared evidence, and errors by either party.
Sources:
The excerpt contains numbered citations, but the corresponding bibliographic references were not supplied.
