Study shows AI can develop its own hiring biases

There’s something of a cliche in the AI world that if you put garbage data into the system, you’ll get garbage results. But a new study from researchers at Princeton University and the University of Chicago has found that some AI models are developing novel social biases all on their own. The researchers tested 15 models including Open AI, Anthropic, Google and Meta, in a simulated hiring process and found that the models began sorting demographic groups differently based on hiring patterns. So if a company hires more white men, for example, the model will learn that this company prioritizes white men and deliver more of them to the hiring team, even if all the data was equal at the beginning of the process.