AI language models show 80× variation in how much they discriminate based on social class

Researchers built a test suite to measure whether large language models discriminate based on socioeconomic status — something almost nobody was measuring until now. They found massive variation across models: some discriminate in lifestyle judgments 10 times more than in education decisions, and all of them fail at catching subtle class-based stereotypes even when explicit discrimination filters are turned on.

For years, AI bias testing focused on race and gender because those were easy to measure. Class bias got ignored, which means AI systems being deployed for hiring, lending, and healthcare decisions were never actually audited for it. This paper gives you a measurement tool — 240 test prompts across 18 different class-related scenarios — which means companies can now measure what they've been ignoring. The catch: even the models that look clean on explicit discrimination still leak class bias in subtle ways. Filters work on the obvious stuff. They don't work on the stereotype.