The algorithm that screens your résumé might be more prejudiced than any human recruiter. New research reveals that large language models don’t just inherit biases from their training data—they actively develop new stereotypes through experience, often judging job applicants more harshly than people would.
We’ve known for years that AI picks up human prejudices from the internet. But this takes the problem a step further. As companies like OpenAI and Anthropic race to build agentic models that remember user interactions, they’re also handing these systems the raw material to form their own biased associations. The timing couldn’t be worse, with AI screening tools already deployed across hiring pipelines before most candidates ever speak to a person.
Michelle Kim’s reporting digs into the mechanism behind this. The models aren’t just parroting biased language—they’re learning from patterns in who gets hired and rejected, then applying those patterns in ways that can systematically disadvantage certain groups. One researcher quoted in the piece frames it bluntly: the more these systems interact with real-world hiring data, the more opportunities they have to bake in discriminatory shortcuts.
What complicates this further is the broader push for AI memory. Tech companies see persistent, personalized models as the next frontier. But every detail a system retains about previous candidates becomes potential ammunition for stereotyping the next one. The question isn’t whether AI can screen résumés faster than humans—it clearly can. The question is whether we’re willing to automate discrimination at scale in exchange for that speed.
