AI · 8 min read · April 17, 2026
LLMs show human-like trust bias toward people, with demographic blind spots
Study of 43,200 experiments reveals language models develop trust patterns similar to humans, including susceptibility to age, religion, and gender bias in financial decisions.
Language models form trust in humans using competence, benevolence, and integrity cues, but exhibit demographic biases similar to human decision-makers.
- — LLMs assess human trustworthiness through three dimensions: competence, benevolence, integrity.
- — Trust formation in models mirrors human behavioral patterns across most tested scenarios.
- — Demographic variables (age, religion, gender) skew LLM trust judgments, especially in finance.
- — Different model architectures show varying sensitivity to trustworthiness and demographic signals.
- — Biases emerge more consistently in newer models and common benchmark scenarios.
- — Trust-sensitive applications (loans, hiring) require explicit monitoring of AI-to-human trust dynamics.
- — Trustworthiness alone does not always predict LLM trust; context and model type matter.
Frequently asked
- The study measures implicit trust through model outputs in simulated scenarios—whether an LLM recommends approving a loan or hiring a candidate based on human input. Whether this constitutes genuine trust or statistical correlation is philosophical; the practical concern is that models behave *as if* they trust, and that behavior is biased by demographics. For decision-making purposes, the distinction matters less than the measurable bias.