A recent study from Hebrew University provides a technical examination of how large language models evaluate people. Researchers Valeria Lerman and Yaniv Dover tested five leading LLMs—including OpenAI's GPT-5 and Google's Gemini 2.5 Pro—against a pool of 1,000 human participants across simulated decision scenarios like lending and hiring.

The core finding concerns system architecture. While both humans and models valued competence, honesty, and benevolence, their processing differed fundamentally. Human judgment is holistic, blending traits into an overall impression. The LLMs, in contrast, operated with a decomposed, additive logic—treating each trait as a separate column in a spreadsheet. This design leads to more extreme and predictable reactions to individual attribute changes, with the models' decisions adhering more tightly to a formal trust framework than human choices do.

From an engineering perspective, the consistency is a double-edged sword. It reduces random noise but amplifies systematic bias. In financial scenarios, regression analysis showed strong, positive coefficients for older age and male gender, with religion also introducing significant prediction variance. These demographic signals were more pronounced and regular in the model outputs than in the human data. Notably, bias magnitude varied sharply between models and use cases; a profile accepted by one system could be rejected by another with identical prompts.

For practitioners, the implications are operational. Model selection is a critical design decision with direct fairness consequences. The research, published in Proceedings of the Royal Society A, argues for rigorous, use-case-specific validation, continuous monitoring of prediction coefficients for protected attributes, and maintaining human review loops for high-stakes applications. The systems model aspects of human reasoning but with a mechanical precision that alters outcomes.

Source: Webpronews