Improving care5 min read

Bias and fairness in AI

AI trained on unrepresentative data can underperform for certain populations; building fair models is one of the field's most urgent tasks.

An AI model is only as good as its training data. If the mammograms used come overwhelmingly from one population, performance can degrade for women of other ages, breast densities, ethnicities or economic backgrounds.

This is not a theoretical concern — several studies have found that AI performance varies across groups. The fix is deliberate: diverse training data, subgroup testing before deployment, transparent reporting by population and continuous monitoring in the real world.

How fairness is protected

  • Training and validation on diverse, representative data
  • Performance reported separately for each subgroup
  • Regulatory review of bias before approval
  • Continuous monitoring after clinical deployment

If AI works well for some and not others, it can widen health inequalities. Fairness is a safety requirement, not an afterthought.

In brief

AI performance can vary by age, ethnicity and breast density.
Diverse training data is the first line of defence.
Subgroup reporting makes bias visible before deployment.

Keep learning

Risk prediction with AIAI HubClinical validation and evidenceAI HubSafety and data privacyAI Hub