
GPT-3.5 Plays Favorites With Resumes, Bloomberg Finds, Because Of Course It Does
Ourtake
Hiring bots learned bias from the humans who wrote their training data, then got a job anyway.
What happened
A Bloomberg investigation tested OpenAI's GPT-3.5 by ranking equally qualified resumes with names associated with different races and genders. The model favored some demographic groups more often, to a degree that would fail benchmarks for job discrimination. OpenAI said most business clients fine-tune the model to reduce bias.