# AI is more likely than humans to form biases when hiring
Artificial intelligence systems screening job applications may introduce biases beyond those in their training data, according to new research. Large language models used in hiring don't just inherit human prejudices from their training datasets. They develop additional biases independently during the recruitment process itself.
The research challenges the assumption that AI removes human subjectivity from hiring decisions. Many companies deploy LLMs to filter résumés and identify candidates, betting that algorithms eliminate favoritism. The reality proves more complicated.
Researchers found that LLMs can generate biases through their decision-making patterns, even when trained on relatively neutral data. These models show measurable discrimination against certain demographics in hiring contexts. The biases emerge not just from what the AI learned during training but from how the system processes information and makes choices.
This matters because AI hiring tools now screen millions of job applications annually. LinkedIn, Amazon, and other platforms use algorithmic screening at scale. Candidates never see how the system evaluated them, making bias detection difficult.
The findings suggest several problems. First, companies can't simply audit training data to catch discriminatory patterns. Second, removing bias from training data alone won't solve the problem. Third, the "black box" nature of LLM decision-making makes it hard to identify where discrimination happens.
The research doesn't recommend abandoning AI in hiring entirely. Rather, it calls for transparency requirements and ongoing bias testing throughout a model's deployment. Companies should conduct fairness audits specific to hiring outcomes, not just general model performance. They should also keep humans in the loop, particularly for decisions that eliminate candidates from consideration.
The stakes are high. Biased hiring AI can systematically disadvantage entire groups of workers, locking them out of job markets at scale. Unlike human hiring managers, whose biases affect dozens of candidates, an algorithmic system can discriminate against thousands.