Will AI Replace Ophthalmologists?

Teamwork, creativity, adaptability, empathy—all traits that physicians employ on a daily basis to effectively deliver patient care. One may argue that these are elements of physician-patient interaction that artificial intelligence (AI) could never replicate. However, others would contend that AI models have already demonstrated some of these features. Recent notable examples include AI mastering cooperative gameplay and generative adversarial networks creating novel artwork and melodic music.1–4 These advances were all made possible due to the recent proliferation of deep neural networks, which have also ushered a stepwise improvement in machine learning performance in ophthalmology.5–8 However, it is crucial to clarify that these and similar AI models that show creativity, teamwork, and adaptability are examples of “narrow” AI. These algorithms are typically validated in constrained testing environments and have limited generalizability. Furthermore, when evaluated outside their test environments in a more abstract fashion or presented with intentional adversarial counterfactuals, they often fail with unfortunate consequences.9