Study Reveals AI's Systematic Judgment Differs Sharply from Human Instincts

New research suggests artificial intelligence systems evaluate people with a starkly different logic than humans do, often amplifying demographic biases in the process. A team from Hebrew University, led by Valeria Lerman and Professor Yaniv Dover, compared the decision-making of five leading large language models against 1,000 people. The study, published in Proceedings of the Royal Society A, simulated over 43,000 judgments on scenarios from business loans to hiring a babysitter.

Both humans and AI valued competence, honesty, and benevolence. Yet their methods diverged. People blend character traits into an overall impression. The AI models, however, treated each trait as a separate data point, applying a more rigid, spreadsheet-like analysis. This systematic approach makes AI judgments more extreme and predictable than human ones.

Crucially, the machines' consistency magnified hidden prejudices. In financial scenarios—like approving a loan or a donation—the models systematically awarded more money to older applicants, males, and those identified as Jewish, even with identical profiles. These biases were stronger and more uniform than those observed in the human participants. In lower-stakes situations, such as rating a potential boss, the demographic skew was less pronounced.

The findings carry immediate weight for business applications. As these systems move from assistants to decision-makers in hiring, lending, and healthcare, their inherent rigidity could formalize exclusion. The study also found significant inconsistency between different AI models from the same vendor, meaning the choice of system itself becomes a variable.

Professor Dover notes the systems are powerful tools that capture elements of human reasoning with remarkable consistency. But they are not human. The research, published this month, is a call for rigorous, case-specific testing and informed human oversight, especially where decisions affect lives and livelihoods. The core lesson for leaders is clear: understand the judge you are deploying.

Source: Webpronews