LLMs Form Structured Trust Assessments Like Humans
Researchers at the Hebrew University of Jerusalem examined how large language models evaluate people across practical scenarios and found that LLMs develop coherent, componentized representations of interpersonal trust that resemble human reasoning but behave more rigidly. In a study by Valeria Lerman and Yaniv Dover comparing five LLMs and human participants across 43,200 simulations, models consistently scored targets on competence, integrity, and benevolence. The models reproduced the same basic axes humans use, yet applied them mechanically, creating consistent assessments and, in financial scenarios, amplifying demographic effects tied to age, religion, and gender even when all other attributes were identical. The result raises practical concerns for LLM-driven decision systems and agents operating in lending, hiring, recommendations, or any context where proxy trust judgments affect outcomes.
What happened
Researchers at the Hebrew University of Jerusalem, led by Valeria Lerman and Yaniv Dover, tested how LLMs evaluate people and found models form coherent, componentized trust assessments that mirror human judgment but operate more rigidly. The paper compared five LLMs with human participants across 43,200 simulations using scenarios like lending to a small business, hiring a babysitter, rating a manager, and deciding donations.
Technical details
The study measured model and human responses along three primary trust axes:
- •competence
- •integrity
- •benevolence
Models effectively decomposed profiles, scoring each axis independently and aggregating them into an overall judgment. This produced consistent, repeatable outputs across runs, unlike humans who showed more holistic and context-dependent variation. The authors report the models applied the axes almost like separate spreadsheet columns, producing deterministic combinations of attributes into final decisions. The experiments isolated demographic variables while holding other features constant to reveal systematic differences in model outputs.
Key findings: Models aligned with humans on the importance of competence, integrity, and benevolence, but were more prescriptive and less noisy. Critically, in monetary decisions models amplified demographic effects: age, religion, and gender altered lending and donation amounts in consistent ways even when profiles were otherwise identical. That pattern points to a rigid internal mapping from demographic tokens to numerical outcomes.
Context and significance
This work reframes model behavior from mere language prediction to structured social inference. For practitioners building LLM-based agents or decision support, the results matter for fairness, governance, and system design. Rigid, componentized trust models can help with explainability and predictable policy enforcement, but they also lock in biases that humans might attenuate through contextual reasoning. The paper connects to existing work on representation bias and social stereotyping in embeddings and downstream classifiers, and it surfaces a new failure mode: systematic, model-consistent demographic skew in normative judgments.
What to watch
Validate deployed agents with scenario-based audits that probe componentized trust scores and demographic sensitivity. Consider mitigation layers that recalibrate trust outputs or inject controlled stochasticity to better match human variability.
Scoring Rationale
The study reveals a novel, actionable characterization of LLM social inference that matters for deployed agents and fairness audits. It is notable for practitioners but not paradigm-shifting, hence a mid-high 'notable' score.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problemsStep-by-step roadmaps from zero to job-ready — curated courses, salary data, and the exact learning order that gets you hired.
