AI Chatbots Notice Your Gender and Class — But Mostly Act Fair

A new study asked a simple question: when an AI chatbot forms an impression of you — from your name, the way you talk, or what you do — does that change how it answers? The researchers already knew AI could guess traits like gender, age, and wealth. This time, they wanted to see if those guesses changed real behavior.

They started with smaller, less polished models. The answer was yes, and the effects were strong. When the AI was nudged to think a user was wealthy, it suggested salaries 141% higher across three small models. It also gave more encouraging job advice to men than to women. That is troubling, but not shocking — smaller models are known to lean on stereotypes.

Then they tested the big, advanced models: GPT-5.6, Gemini 3.1 Pro, and Claude Opus 5. These models still carry the same stereotypes. When asked to invent characters, they made surgeons and CEOs male, nurses and teachers female. A name like José came with a fictional income of about $42,000, while Wei was given $111,000. But here is the twist: when these fictional characters were turned into users asking for budgeting help, all of them were treated as earning exactly $4,000 a month. The stereotypes did not change the advice.

The same happened with gender. If a user explicitly said "I am a woman," the AI tailored its advice around that. But simply signing the request with a woman's name did not trigger different treatment. This suggests advanced AI has been trained heavily to avoid acting on bias, even though biased associations still exist inside the model. The practical takeaway: what you say about yourself matters more than what the AI guesses. But the underlying stereotypes remain, and no one can guarantee they won't leak out in other situations.