85% Racial Bias, 61% Gender Bias: ChatGPT and DeepSeek Still Fail to Eliminate Stereotypes in Medicine

A team at Flinders University, in work published in the Journal of Medical Internet Research, shows that advanced reasoning models like ChatGPT and DeepSeek significantly distort the demographic representation of common diseases. The study finds persistent racial and gender bias in how these models construct fictitious patients, raising concerns about patient safety and fairness if such systems are used for clinical decision support.