Table 1 Typology of biases

Data bias

Bias can occur with any dataset. Common sources of bias potentially promote or harm group-level subsets based on gender, sexual orientation, ethnic, social, environmental, or economic factors

Clinical confounding bias

Radiology AI may be biased by clinically confounding attributes such as comorbidities

Technical bias

Bias can be introduced due to subtle differences in raw and post-processed data that come from different scanning techniques

Automation bias

This is the tendency for humans to favor AI decisions, ignoring contrary data or conflicting human decisions. This can lead to errors of omission (when humans fail to notice, or disregard, the failure of an AI tool) and commission (when one erroneously accepts or implements a machine’s decision despite other evidence to the contrary). (See also Section 8)