The effect of AI assistance timing on performance and user perceptions in pathological slide diagnosis.

This study evaluated three timing strategies for delivering AI assistance in pathological slide diagnosis - pre-diagnosis (triage), during diagnosis (concurrent), and post-diagnosis (secondary) - and assessed users' perceptions of AI assistance. All three AI modes improved diagnostic performance and reduced workload versus no AI assistance. Concurrent mode was preferred for its balance between efficiency and reader control; secondary mode was appreciated for minimizing bias and aiding training. Triage mode yielded lower workload and higher performance but raised concerns about trust and transparency. AI was regarded as a valuable tool for initial slide review, but not as a replacement for expert readers. Participants generally trusted the AI for highlighting suspicious areas, not making final decisions. After use, willingness to rely on AI for final diagnosis declined, though trust and usability remained moderate to high. To increase adoption, designers should manage AI information presentation to avoid bias, balance sensitivity and specificity based on user feedback, and improve explainability to enhance reader confidence and trust.