AI is more likely than humans to form biases when hiring

The prevalent assumption that AI eliminates human biases in hiring is a critical practical problem the latest research from MIT’s AI lab definitively debunks, with significant implications for anyone using or developing AI for recruitment. The recent study demonstrates that large language models are actually more prone to generating biased inferences—specifically gender and racial biases—when asked to assess job candidates based on identical information, compared to human recruiters given the same task. This finding strongly challenges the notion that AI inherently provides a fairer, more objective hiring process. For founders building HR tech, this means a re-evaluation of core assumptions and a deeper investment in bias detection and mitigation techniques is non-negotiable, not merely an add-on. A startup in Denver, for instance, developing an AI-powered resume screening tool, might discover their models, left unchecked, disproportionately penalize candidates from specific demographics, leading to a product that exacerbates rather than solves hiring inequality. Similarly, an internal IT team at a mid-sized financial firm in Charlotte looking to deploy an off-the-shelf AI recruitment solution must now meticulously vet its bias profile, demanding verifiable evidence of debiasing strategies from vendors, rather than relying on marketing claims of objectivity. For an independent consultant helping businesses optimize their hiring pipelines, this research provides concrete data to advise clients against blind adoption of AI in sensitive areas, pushing for hybrid models where human oversight actively scrutinizes AI outputs for bias. To capitalize on this insight, consider a small, actionable experiment this week. If you are involved in any form of automated candidate screening, take a set of anonymized resumes you’ve previously processed and, instead of feeding them through your current AI, manually review them yourself, explicitly flagging any internal biases you detect for gender, race, or background. Then, pass the same anonymized set through your AI system and compare the aggregate results, particularly the demographic distribution of candidates flagged for advancement versus rejection. This direct comparison, even with a small sample, can provide a preliminary, tangible gauge of your AI's bias amplification relative to your own, informing where to focus your debiasing efforts next.