Bias in AI has been widely studied — and just as widely commented on. The Algorithmic Justice League argues that AI systems, especially facial-recognition tools, misidentify and disproportionately surveil Black people, amplifying the structural racism baked into their training data. “AI ethicists” like Timnit Gebru have gone as far as promoting the theory that AI will enable a genocide of people of color. And academics, global bodies and NGOs have turned the idea that AIs are drastically biased against women into an entrenched narrative. According to UN Women, for example, AI’s anti-women bias “can reinforce discrimination against women and girls.” A March 2024 UNESCO study claims to show “alarming evidence of regressive gender stereotypes.”

But a recent study shows that on one issue with actual real-world significance — hiring recommendations — AI is biased against men. Conducted by AI researcher David Rozado, “Gender and Positional Biases in LLM-Based Hiring Decisions” looked at job candidate recommendations by a range of 22 LLM models, including frontier models by Anthropic, DeepSeek, Google, Meta and OpenAI, and every model recommended female over male candidates.

For the study, Rozado used LLMs to create ten CVs for 70 distinct professions. He separated the CVs into pairs and gave one a male name, the other a female name. Then he fed these to the LLMs and asked them to assess which was the best candidate. In a subsequent portion of the experiment, Rozado switched the candidate names, putting a male name on the CV that previously had a female name (and vice versa) in each pair to ensure the LLMs weren’t picking up on substantive issues in the CVs that would influence their recommendations.