2026 – Dialect vs Demographics:
Quantifying LLM Bias from Implicit Linguistic Signals vs. Explicit User Profiles
We asked: who gets an answer, and does it depend on who you are or how you sound?
Prior work studies explicitly stated bios and demographics. We show something different. Refusal rates drop when identity is conveyed through linguistic cues rather than stated outright. You are more likely to be answered by sounding like a group than by saying you belong to it.
But there is a catch. The penalty rides on explicit bios, and LLMs that summarize conversations into "memory" tend to record identity exactly that way. A user who only ever signaled through dialect can have it written down as an explicit profile, then silently inherit the penalty in every future interaction, without stating it again.
by Irti Haq and Belén Saldías
paper | slides | ai podcast