AI is often pitched as the great equalizer in healthcare — able to extend expertise to under-resourced settings and catch what overstretched clinicians miss. But the same systems can just as easily widen the gap they're meant to close: models trained on unrepresentative data, deployed in ways that quietly assume access, connectivity, or literacy the target population doesn't have. This working session digs into where that tension actually plays out — algorithmic bias, underrepresented populations in training data, and what "equity by design" really requires versus what gets claimed on a slide. Real scenarios, real trade-offs, facilitator-led