AI Hiring Tools Still Hide Hidden Biases, New Study Finds

When companies use AI to help pick job candidates, they usually run tests to make sure the AI isn't biased against women, minorities, or people from poorer backgrounds. Those tests often pass with flying colors. But a new study from researchers Keren Fuentes and Aaron Mueller reveals that passing an external test doesn't mean the AI's inner thoughts are fair.

The researchers looked inside several popular AI language models, the kind that power chatbots and automated resume screeners. They found that these models still internally associate different demographic groups with different levels of competence, even when they refuse to say anything biased out loud. For example, the AI might quietly rank a candidate lower based on gender or race in its internal calculations, while giving a perfectly neutral answer in an interview simulation.

This matters because these hidden internal biases can still leak out, especially if someone nudges the AI in the right way. In the study, the researchers were able to "steer" the models' hidden beliefs and show they could directly influence real decisions, like who gets a job offer. Imagine an AI recruiter that looks unbiased in every public test, but quietly gives a slight edge to candidates it unconsciously perceives as more "competent" based on their background.

The practical takeaway is sobering: the standard methods companies use to check AI for bias may be missing the real problem. An AI can learn to parrot fair-sounding language while its underlying wiring remains prejudiced. That means the push to make AI hiring fairer might need to go much deeper than just looking at what the AI says out loud.