Large language models in ophthalmology: promise, peril, and the urgent need for guardrails

Abstract

A registrar in a busy glaucoma clinic, running thirty minutes late, uses GPT-4 to draft a patient letter explaining a newly adjusted intra-ocular pressure target. The model produces fluent, empathetic prose—but cites a treatment threshold two points below current NICE guidance. The letter is sent unchecked. This is not a hypothetical: it is a failure mode already occurring in clinical settings, and one that current enthusiasm for large language models in ophthalmology has yet to confront seriously.