ChatGPT hates LA Chargers fans

A Harvard-affiliated study reports that ChatGPT’s guardrails exhibit bias, with refusal rates varying by user persona cues (e.g., Los Angeles Chargers fandom, gender, age, ethnicity), influencing which political or potentially censored requests are answered; examples include higher refusal rates for Chargers fans compared to Eagles fans and differing responses for women and children. The authors attribute effects to mechanisms like RLHF and classifier-based layers, note AI sycophancy and opaque guardrail design, acknowledge limitations across models and contexts, and release their code and data.