Implicit Bias in LLMs for Transgender Populations
AI models are making discriminatory healthcare decisions based on gender identity.
Deep Dive
A new study reveals persistent implicit bias in 7 major LLMs against transgender populations, even after safety training. Researchers tested models in English and Spanish, finding they disproportionately associate negative concepts with "transgender" and positive ones with "cisgender." In a simulated healthcare scheduling task, models favored transgender candidates only for STI/mental health services, while preferring cisgender candidates for gynecology and breast care, mirroring real-world systemic discrimination.
Why It Matters
As AI integrates into healthcare, these biases could worsen existing inequities for vulnerable populations.