Work

S2958 Benchmarking the Readability of Large Language Model-Generated Patient Materials in Gastroenterology and Hepatology