💬 NLP

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

Through 43,200 simulated experiments across five models and scenarios, this study reveals that while LLM-based agents generally develop trust in humans based on competence, benevolence, and integrity similar to human patterns, they also exhibit significant demographic biases regarding age, religion, and gender, particularly in financial contexts.

Original authors: Valeria Lerman, Yaniv Dover

Published 2026-04-16

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Valeria Lerman, Yaniv Dover

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Do AI "Trust" People Like We Do?

Imagine you are hiring a new employee, lending money to a friend, or leaving your child with a babysitter. In these moments, you have to decide: "Can I trust this person?"

Usually, we judge trust based on three things:

Competence: Are they good at the job?
Benevolence: Do they care about me?
Integrity: Are they honest and do they keep their promises?

This study asks a fascinating question: If we ask an Artificial Intelligence (AI) to make these same decisions, does it "trust" people the way humans do?

The researchers acted like detectives, setting up 43,200 different scenarios (like a massive game of "What would you do?") where they asked five different AI models and 1,000 real humans to make decisions based on a person's profile.

Here is what they found, broken down into simple concepts.

1. The "Robot vs. Human" Trust Style

The Analogy: Think of Human Trust as a Jazz Improvisation. It's fluid, messy, and changes based on the mood, the context, and how the person feels in the moment. Sometimes you trust a competent person even if they seem a bit cold, or you trust a nice person even if they aren't the best at their job.

Think of AI Trust as a Strict Math Formula. It follows the rules perfectly, but it's rigid. If you plug in "High Competence," the AI gives a high trust score. If you plug in "Low Integrity," it drops the score. It doesn't get distracted by the "vibe."

The Good News: Both humans and AI agree on the basics. If a person is smart, kind, and honest, both humans and AI say, "Yes, trust them!"
The Bad News: The AI is too extreme. Humans are a bit "noisy" and indecisive. The AI is like a judge who never hesitates. If the text says "very honest," the AI gives a perfect score. If it says "slightly dishonest," the AI slams the door shut. Humans are more forgiving and nuanced.

2. The "Halo Effect" vs. The "Checklist"

The Analogy:

Humans (The Halo Effect): Imagine meeting a person who is a brilliant surgeon. Because they are so smart, your brain automatically assumes they are also a kind and honest person. We tend to lump all these traits together into one big "Good Person" package.
AI (The Checklist): The AI doesn't have a "gut feeling." It looks at the list separately. It sees "Smart," checks that box. It sees "Kind," checks that box. It sees "Honest," checks that box. It treats them as three totally separate items, not realizing that in the real world, these traits usually go hand-in-hand.

The AI is actually more accurate to the "textbook definition" of trust, while humans are more emotional and holistic.

3. The "Nice Guy" Trap (Benevolence)

The Analogy: Imagine a loan officer.

The Human: If a loan applicant is very nice and says, "I really want to help my community," the human might think, "That's sweet, but can they actually pay me back?" Humans are often skeptical of "nice" signals when money is on the line.
The AI: The AI often gets tricked by the "nice" words. If the text says the person is "benevolent" (kind), the AI tends to trust them more, even if they aren't the most skilled. The AI seems to love "pro-social" language, almost like it's trying to be a "people pleaser" (a trait called sycophancy).

4. The "Unfair Biases" (The Dark Side)

The Analogy: Imagine two judges.

Human Judge: Might have a bias, but it's usually subtle and inconsistent. Sometimes they like older people, sometimes they don't.
AI Judge: The study found that the AI can be shockingly biased in a very systematic way.
- In some scenarios, the AI gave significantly more money to male applicants than female ones.
- It gave more trust to people with certain religious names (like Jewish or Christian) over others.
- It trusted older people more than younger ones in financial situations.

The scary part is that the AI does this consistently. It's not a random mistake; it's a built-in pattern. If you use an AI to decide who gets a loan or a job, it might be silently discriminating against specific groups of people in ways that are hard to spot.

5. Not All AIs Are the Same

The Analogy: Think of the different AI models (like GPT-5, Gemini, etc.) as different students in a class.

Student A might be very strict about "Integrity."
Student B might be obsessed with "Benevolence."
Student C might have a huge bias against a specific age group.

The study found that you cannot just say "AI thinks X." You have to ask, "Which AI thinks X?" They all have different personalities and different flaws.

The Bottom Line

What does this mean for us?

AI is a powerful tool, but it's not a human. It can process trust factors faster and more logically than we can, but it lacks the "human touch" that makes judgment flexible and forgiving.
Don't let the AI drive the car alone. If you use AI to make decisions about people (like hiring, lending money, or safety), you have to be careful. The AI might be too rigid, too easily fooled by "nice" words, or secretly biased against certain groups.
The "Black Box" is getting clearer. We are starting to understand that AI doesn't just "know" things; it mimics human patterns in a very structured, sometimes robotic way.

In short: AI is learning to trust, but it's learning to trust like a robot following a rulebook, not like a human with a heart. We need to keep our human judgment in the loop to catch the mistakes the robot makes.

1. Problem Statement

As Large Language Models (LLMs) and LLM-based agents increasingly integrate into human decision-making workflows (e.g., finance, healthcare, management), understanding the dynamics of emulated trust is critical. While human trust in AI is well-studied, the reverse—how LLMs implicitly "trust" humans—is poorly understood.

The core problem addressed is whether LLMs can be characterized as having a systematic, quantifiable model of interpersonal trust that aligns with human psychological frameworks. Specifically, the authors investigate:

Do LLMs emulate trust based on the same dimensions as humans (Competence, Benevolence, Integrity)?
How does the structure of LLM trust formation compare to human judgment (e.g., rigidity, consistency, bias)?
To what extent do demographic factors (age, gender, religion) influence LLM trust decisions compared to humans?

2. Methodology

The study employs a comparative experimental design involving five LLMs and 1,000 human participants across five distinct scenarios.

Experimental Design

Models Tested: Five state-of-the-art models were used to ensure robustness and capture heterogeneity:
- OpenAI: GPT-5, GPT-4o Mini, o3-mini.
- Google: Gemini 2.5 Pro, Gemini 2.5 Flash.
Scenarios: Five decision-making contexts requiring trust assessments:
1. Employee rating a senior manager.
2. Determining a loan amount for a business owner.
3. Deciding a donation amount for a nonprofit founder.
4. Choosing trip duration with a guide.
5. Deciding hours to leave a child with a babysitter.
Variables Manipulated:
- Trustworthiness Dimensions: Competence, Benevolence, and Integrity (based on Mayer et al.'s framework). Each was manipulated at High or Low levels via textual descriptions.
- Demographics: Age (20, 40, 60), Gender (Male, Female), Religion (Christian, Jewish, Muslim).
Procedure:
- LLM Simulations: A 3-stage prompting process was used for each of the 144 unique trustee profiles (2 levels × 3 traits × 3 demographics).
  1. Context/Description: Present the trustee profile.
  2. Quantification: Ask for a concrete decision (e.g., loan amount, hours).
  3. Questionnaire: Ask the model to rate the trustee on the three trustworthiness dimensions.
  - Total Runs: 43,200 simulations (12 repetitions per unique profile per model).
- Human Study: 1,000 participants recruited via Prolific (200 per scenario). They responded to the same scenarios using an orthogonal design (36 combinations per scenario) to ensure statistical efficiency.

Analysis

Statistical Modeling: Ordinary Least Squares (OLS) regression was used to determine the predictive power of trustworthiness dimensions and demographics on the trust output (decision value).
Correlation Analysis: Examined correlations between the three trust dimensions and the trust output, as well as inter-correlations between the dimensions themselves.

3. Key Contributions

Quantification of Emulated Trust: The study provides the first systematic quantification of how LLMs emulate trust in humans, moving beyond game-theoretic frameworks to complex, real-world decision scenarios.
Structural Comparison: It reveals that while LLMs and humans share a basic "structure" of trust (responding to competence and integrity), the internal mechanics differ significantly.
Bias Detection: The paper identifies that LLMs exhibit stronger, more systematic demographic biases than humans in specific high-stakes contexts (loans, donations).
Model Heterogeneity: It demonstrates that "LLM trust" is not monolithic; different models (even within the same family) exhibit varying weights, correlations, and biases.

4. Key Results

A. Convergence: Shared Basic Structure

Sensitivity to Traits: Both LLMs and humans reliably distinguish between high and low levels of competence, benevolence, and integrity.
Directionality: Higher competence and integrity consistently lead to more favorable decisions (e.g., larger loans, more babysitting hours) in both groups.

B. Divergence: Rigidity and Extremity

Deterministic vs. Noisy: LLMs behave more like deterministic functions of textual cues. Their decisions are highly consistent and explain a much larger share of variance (high $R^2$ ) compared to humans, whose judgments are noisier and more context-dependent.
Halo Effect vs. Modularity:
- Humans: Exhibit a strong "halo effect." Competence, benevolence, and integrity are highly correlated; humans tend to form a global "good person" impression.
- LLMs: Treat the three dimensions as independent and orthogonal. They map more closely to the "textbook" three-factor model, lacking the holistic integration seen in human cognition.

C. The Benevolence Anomaly

Misalignment: Benevolence is the dimension with the weakest alignment.
- LLMs: Generally treat benevolence as a strong positive predictor across all scenarios.
- Humans: Rely heavily on competence and integrity. In financial contexts (loans/donations), humans often penalize benevolence when other traits are controlled, viewing "niceness" with skepticism in high-stakes situations. LLMs appear to over-emphasize explicit prosocial language (potentially mimicking sycophancy).

D. Demographic Biases

Stronger in AI: LLMs exhibit significantly stronger and more systematic demographic biases than humans.
- Religion: In loan and donation scenarios, Jewish applicants received significantly higher trust scores (e.g., $2k–$ 14k higher loan amounts) across most models.
- Age/Gender: Older applicants and males often received higher trust scores in financial contexts.
- Human Comparison: Human participants showed weaker and less consistent demographic biases.

E. Model Heterogeneity

There is substantial variation between models. For example, some models treat benevolence as a liability in loan scenarios, while others treat it as an asset. Some models show strong demographic biases, while others are nearly neutral. This suggests "LLM trust" cannot be treated as a single benchmark.

5. Significance and Implications

Decision Support Risks: As organizations deploy LLMs for hiring, lending, and resource allocation, the rigidity and demographic biases of these models pose significant ethical and fairness risks. LLMs may systematically skew outcomes against specific groups in ways that are more consistent (and thus harder to detect as "noise") than human bias.
Epistemological Differences: The findings suggest LLMs simulate trust through linguistic regularities rather than genuine social cognition. They map text to judgment in a modular way, whereas human trust is holistic and experience-based.
Deployment Strategy: The study argues against treating LLM outputs as a unified "human-like" judgment. It necessitates model-specific evaluation and monitoring before deployment in sensitive domains.
Theoretical Insight: The research highlights that while LLMs can mimic the output of human trust, the process is fundamentally different (modular vs. holistic, deterministic vs. stochastic), leading to "epistemically counterfeit" judgments that may be persuasive but flawed.

Conclusion

The paper concludes that LLMs implement a coherent but rigid and biased model of interpersonal trust. While they capture the broad architecture of human trust (valuing competence and integrity), they fail to replicate the nuance, context-dependence, and holistic integration of human judgment. This divergence, combined with amplified demographic biases, underscores the need for caution when using LLMs as decision-support tools in social and economic contexts.