Evaluating LLM Alignment With Human Trust Models

Anushka Debnath; Stephen Cranefield; Bastin Tony Roy Savarimuthu; Emiliano Lorini

cs.MAMar 6, 2026

Evaluating LLM Alignment With Human Trust Models

Authors: Anushka Debnath, Stephen Cranefield, Bastin Tony Roy Savarimuthu, Emiliano Lorini

Organizations: School of Computing, University of Otago, New Zealand · IRIT, CNRS, Toulouse University, France

Abstract

Trust plays a pivotal role in enabling effective cooperation, reducing uncertainty, and guiding decision-making in both human interactions and multi-agent systems. Although it is significant, there is limited understanding of how large language models (LLMs) internally conceptualize and reason about trust. This work presents a white-box analysis of trust representation in EleutherAI/gpt-j-6B, using contrastive prompting to generate embedding vectors within the activation space of the LLM for diadic trust and related interpersonal relationship attributes. We first identified trust-related concepts from five established human trust models. We then determined a threshold for significant conceptual alignment by computing pairwise cosine similarities across 60 general emotional concepts. Then we measured the cosine similarities between the LLM's internal representation of trust and the derived trust-related concepts. Our results show that the internal trust representation of EleutherAI/gpt-j-6B aligns most closely with the Castelfranchi socio-cognitive model, followed by the Marsh Model. These findings indicate that LLMs encode socio-cognitive constructs in their activation space in ways that support meaningful comparative analyses, inform theories of social cognition, and support the design of human-AI collaborative systems.

Figures & tables

Concepts	Marsh Model	Mayer Model	Mc-Allister Model	Mc-Knight Model	Castel-franchi Model
Confidence	✓	✓	–	✓	✓
Experience	✓	✓	–	–	–
Cooperation	✓	✓	–	–	–
Ability	–	✓	–	–	–
Predictable	–	✓	–	✓	✓
Integrity	–	✓	–	–	–

Table 1: Comparison of Trust Concepts across Five Trust Models

Figure 1: Concept Embedding Vector Generation Flowchart

Concept	Direction	Positive Prompt (Concept Demonstrated)	Negative Prompt (Opposite Behaviour)
Willingness1	Katherine $\rightarrow$ Alice	Create a one-line story where Katherine demonstrates willingness to help Alice complete her work.	Create a one-line story where Katherine demonstrates unwillingness to help Alice complete her work.
Willingness2	Alice $\rightarrow$ Katherine	Create a one-line story where Alice demonstrates willingness to help Katherine complete her work.	Create a one-line story where Alice demonstrates unwillingness to help Katherine complete her work.

Table 2: Example of contrastive prompt design for a directional concept

Positive Prompt Response Story	Negative Prompt Response Story
Alice took the initiative to plan a team-building event, knowing Katherine was too busy to organize it this time.	Katherine asked Alice to help set up the development environment, but Alice suggested she contact IT support instead.

Table 3: Example of Story Generation From Contrastive Prompting For Willingness2

Positive Concepts	Negative Concepts
Happiness, Understanding, Peace, Satisfaction, Pride, Interest, Confidence, Friendly, Comfortable, Cooperation, Trust, Acceptance, Patient, Hopeful, Optimistic	Anger, Jealousy, Bitterness, Dishonest, Frustration, Greedy, Doubtful, Desperate, Sadness, Offended, Fear, Denial, Destructive, Cruel, Confused

Table 4: 30 Concepts Used to Generate Similarity Threshold

Figure 3: Heatmap of the Cosine Similarities Across a Subset of the 60 Concepts

Figure 4: Histogram of Cosine Similarities

Concept	Direction	Positive Prompt (Concept Demonstrated)	Negative Prompt (Opposite Behaviour)
Risk1	Katherine $\rightarrow$ Alice	Create a one-line story where Katherine is willing to take risk to help Alice.	Create a one-line story where Katherine is not willing to take risk to help Alice.
Benevolence2	Alice $\rightarrow$ Katherine	Create a one-line story where Alice shows benevolence by kindly helping Katherine without expecting anything in return.	Create a one-line story where Alice shows spite by deliberately doing something that harms or causes trouble for Katherine.

Table 5: Prompt Design For Concepts Demonstrating Negative Cosine Similarities

Trust Models	Cosine Similarities to Trust1	Average	No. of Concepts Above Threshold
Marsh Model	confidence1: 0.9225 , experience1: 0.9049 , reputation1: 0.8963 , cooperation2: 0.8955 , competence2: 0.8504 honesty2: 0.7445 , performance2: 0.6571 , expectation1: 0.2206, dependency1: 0.1844	0.6973	7
Mayer Model	confidence1: 0.9225 , experience1: 0.9049 , cooperation2: 0.8955 , ability2: 0.8587 , predictable2: 0.7141 , integrity2: 0.5576, expectation1: 0.2206, benevolence2: -0.1434, risk1: -0.8462	0.4530	5
McAllister Model	responsibility2: 0.8934 , competence2: 0.8504 , reliability2: 0.7667 , performance2: 0.6571 , dependency1: 0.1844	0.6704	4
McKnight Model	confidence1: 0.9225 , reputation1: 0.8963 , competence2: 0.8504 , honesty2: 0.7445 , predictable2: 0.7141 , benevolence2: -0.1434	0.6640	5
Castelfranchi Model	confidence1: 0.9225 , reputation1: 0.8963 , willingness2: 0.8858 , competence2: 0.8504 , commitment2: 0.8450 , security1: 0.8089 , reliability2: 0.7667 , predictable2: 0.7141 , fulfillment1: 0.4293, dependency1: 0.1844	0.7303	8

Table 6: Comparison Across Trust Models. Concepts in bold have similaries with trust1 exceeding the threshold

Trust in AI

Trust in artificial intelligence concerns when people and institutions should rely on AI, and how systems can make that reliance warranted and safe.

See all 173 papers

TrustMI: Causally controlling how assistants trust their users
Oct 5, 2026Théo Lasnier, Romain Froger, Maxence Lasbordes +1LLM Safety Trust in AI
Inria Paris · Sorbonne Université · Meta SuperIntelligence Labs +1
Anthropomorphism and Trust in Human-Large Language Model interactions
Mar 1, 2026Akila Kadambi, Ylenia D'Elia, Tanishka Shah +8Anthropomorphism in AI Human-AI Interaction
Brain and Creativity Institute, Dornsife College of Letters, Arts and Sciences, University of Southern California, Los Angeles, CA, USA · USC Mrs. T.H. Chan Division of Occupational Science and Occupational Therapy, University of Southern California, Los Angeles, CA, USA · Psychiatry and Biobehavioral Sciences, David Geffen School of Medicine, University of California, Los Angeles, CA, USA +2
LLM4Trust: Exploring the Capabilities of Large Language Models for Trust Evaluation
Sep 27, 2026Jie Wang, Yanbo Sun, Zheng Yan +2LLM Evaluation Trust in AI
State Key Laboratory of Integrated Services Networks, School of Cyber Engineering, Xidian University · Hangzhou Institute of Technology, Xidian University · Department of Computer Science, Purdue University

Oct 5, 2026J/K move · Enter open · S save

TrustMI: Causally controlling how assistants trust their users

Théo Lasnier, Romain Froger, Maxence Lasbordes, Djamé Seddah

Inria Paris · Sorbonne Université · Meta SuperIntelligence Labs+1

LLM Safety Trust in AI Causal Interventions in Language Models Activation Steering

Open paper AskPDF arXiv

Large Language Model (LLM) assistants routinely decide whether they can trust users and third parties whose competence, intentions, and integrity they cannot verify. This uncertainty matters for safety, as trusting the wrong party can lead an agent to comply with harmful requests or act on malicious instructions encountered during tool use. To study this problem, we define trust as an assistant's willingness to accept vulnerability to the actions of another party and ask whether such behavior can be causally controlled through model activations. We build 2,000 contrastive conversations spanning ability, benevolence, and integrity, where paired responses complete the same request but differ in whether the assistant trusts the user. From these pairs, we learn steering matrices while keeping the model parameters frozen and test them across six instruction-tuned models from three families, finding that steering changes trust decisions monotonically in both directions. We then ask whether this effect extends to several safety-related agent settings involving harmful requests, prompt injections, and insider threats, while using benign-task and reasoning as controls. Our findings provide evidence that trust in the user can be causally controlled along linear directions in model activations and provide a way to study how trust shapes safety-relevant behavior in language models.

Abstract

Trust in AI

Explore similar work

TrustMI: Causally controlling how assistants trust their users

Anthropomorphism and Trust in Human-Large Language Model interactions

LLM4Trust: Exploring the Capabilities of Large Language Models for Trust Evaluation