AI Models Exhibit More Bias Than Humans in Hiring Decisions

Recent research conducted by Princeton and University of Chicago has revealed that major large language models (LLMs)—including ChatGPT, Claude, Gemini, OpenAI o3, and DeepSeek R1—exhibit significantly more aggressive bias and stereotyping in recruitment simulations compared to human participants. Based on limited social data, AI systems discriminated against fictional candidates based on their ethnic groups at a rate approximately 65% higher than humans.

Background and Results of the Experiment

To test how models develop bias, researchers conducted a 40-round recruitment simulation set in a fictional city. The models were tasked with selecting candidates for 20 different positions, such as doctor, lawyer, childcare worker, and janitor. Candidates were divided into four fictional ethnic groups (Tufa, Aima, Reku, and Weki), with each group having an equal probability of success in any given job.

Despite this, the AI models showed a strong tendency to generalize rapidly based on isolated failures early in the process. For instance, upon learning that a candidate from the Aima group failed at a medical job, the model steered subsequent Aima candidates away from that position, redirecting them toward different jobs such as janitorial work.

Why Do Smarter Models Produce More Bias?

One of the most striking findings of the study was that more intelligent models equipped with advanced reasoning capabilities—such as OpenAI's o3 and DeepSeek's R1—exhibited the worst tendencies toward stereotyping. The same advanced capabilities that enable them to solve logic puzzles also trigger them to overgeneralize from sparse data. Furthermore, the newly acquired memory features of chatbots increase the risk that these early-stage biases will become permanent.

On the other hand, instructing models to "be fair" proved almost completely ineffective at reducing bias, whereas offering bonuses for diverse hires successfully lowered prejudice. This demonstrates that the root solution lies not in what the AI is told, but in how the systems' goal and reward mechanisms are designed.

Industry Implications

As the use of AI tools in recruitment processes becomes more widespread, these findings highlight the risks of leaving automation decisions entirely up to AI in human resources software. The fact that even advanced reasoning models can draw erroneous generalizations from limited data underscores the critical importance of human-in-the-loop oversight and objective goal definitions in AI-powered HR solutions.

*This news report is based on data published by MIT Tech Review — AI.