News release

From: The Royal Society

A closer look at how large language models 'trust' humans: patterns and biases

Large language models increasingly advise on human-related decisions, yet we know little about how they “trust” people. Across five real-world scenarios and 43,200 simulations, models generally reward higher competence and integrity, echoing patterns seen in human participants. Yet their trust is much more “by-the-book”: responses are more extreme, more internally consistent, and less shaped by the human halo effect that blends traits into a single global impression. Models also differ from one another, with scenario-dependent weighting of trust cues. Most concerning, demographic labels can systematically shift some models’ trust, underscoring fairness risks in deployment in practice.

Can’t read my poker face - AI models may be able to read your face like an open book. Using famous painted and photographed portraits combined with binary choice questions about the subject’s mood or feelings, researchers found some AI models, including ChatGPT-4o, Grok and Gemini interpreted emotions almost indistinguishably from humans, while Claude and Mistral showed some divergence. The findings “offer a foundation for the development of emotionally competent agents capable of operating in socially nuanced environments”, the authors said