Imagine you are working alongside a robotic arm in a factory. On its first day, it drops a heavy component. How do you react? Your response depends entirely on why you think it failed. Did the mechanical gripper slip because the component was too heavy (a failure of capability)? Or did the robot intentionally alter its path to optimize its own energy usage at the expense of your safety (a failure of intent)?
Traditional computational models of trust treat reliance as a single scalar value—a probability score between 0 and 1. But as Stefanos Nikolaidis et al. point out in Planning with Trust for Human-Robot Collaboration, a monolithic score fails to capture human psychology. A human might trust a system's navigation capabilities while completely doubting its physical manipulation skills.
To build systems that can truly reason about and calibrate trust, researchers have shifted toward multi-dimensional trust modeling. Instead of asking how much an agent is trusted, these frameworks decompose trust into distinct, orthogonal axes: typically competence (the technical capability to perform a task) and intent or integrity (the willingness or behavioral disposition to align with expectations).
The Anatomy of Trust Decomposition
Conflating capability with willingness prevents a nuanced understanding of failure. In A Trust Model for Supply Chain Management, Yasaman Haghpanah et al. formalize this distinction using probabilistic game theory. They separate trust into technical competence and behavioral integrity. This separation allows autonomous agents to predict strategic defection: a supply chain partner might have the perfect equipment to fulfill a contract (high competence) but choose to break it if a more lucrative alternative opportunity arises (low integrity).
This same cognitive split occurs when humans interact with physical robots. Work by Yaqi Xie et al. in Robot Capability and Intention in Trust-based Decisions across Tasks demonstrates that humans maintain separate mental models for a robot's capability and its intentions.
Similarly, this decomposition extends to how we process information from virtual agents. Guanxiong Pei et al. show in Dynamic processing of conversational intelligence features in marketing digital humans and its neural mechanisms that trust is bifurcated into cognitive trust (based on perceived competence and accuracy) and affective trust (rooted in emotional connection and empathy). Their findings suggest that while cognitive trust evaluates utility, affective trust acts as a psychological buffer, mitigating the negative impact of errors during interactions.
Why Decomposition Works: Generalization and Risk
Why go through the trouble of tracking multiple dimensions instead of a single, aggregated reliability score? The answer lies in how trust generalizes across new contexts and how agents manage risk.
When trust is broken in a scalar model, the entire score drops, rendering the agent completely untrustworthy across the board. In a multi-dimensional model, the observer generalizes traits separately. As Yaqi Xie et al. show, if a robot fails a task due to a lack of capability, a human will still trust its intentions and may choose to delegate a different task that requires a different skill set. However, if the failure is perceived as a malicious or misaligned intention, trust collapses globally.
Furthermore, trust cannot be decoupled from the risk profile of the environment. Stéphane Lo Presti et al. argue in Analysing the Relationship between Risk and Trust that an agent's willingness to rely on a system is fundamentally shaped by personal risk tolerance and the severity of potential consequences. By isolating reliability from risk attitudes, systems can compute a more precise "decision trust."
Formalizing Multi-Dimensional Trust
To operationalize these insights in autonomous systems, researchers have developed formal mathematical and architectural frameworks.
In Reasoning about Cognitive Trust in Stochastic Multiagent Systems, Xiaowei Huang et al. introduce a formal logic called PRTL*. This framework allows autonomous systems to mathematically verify whether their actions will foster appropriate levels of trust in stochastic environments, explicitly separating an observer's belief in an agent's technical ability from its willingness to cooperate.
On the architectural side, Mohammed Diab introduces the TICK framework in TICK: A Knowledge Processing Infrastructure for Cognitive Trust in Human-Robot Interaction. TICK replaces scalar scores with a multi-dimensional tuple incorporating vulnerability, reliability, and intention. This allows a robot to dynamically adjust its interaction strategies—such as switching between a direct or indirect object handover—based on a granular assessment of the human's current trust state and the situational risk.
Scaling to Macro-Level Governance
The necessity of separating competence from intent is not limited to micro-level interactions between two agents; it governs macro-level public perception as well. In a study on U.S. Public Opinion on the Governance of Artificial Intelligence, Baobao Zhang demonstrates that the public evaluates technology developers along these exact same parallel lines. A stakeholder can be viewed as highly competent at building advanced AI systems, yet completely untrusted to deploy them in an ethical manner that aligns with the public interest.
By shifting from monolithic scalar metrics to multi-dimensional representations, modern AI and robotic systems can accurately diagnose the root cause of trust fluctuations, predict human cooperation across diverse tasks, and safely navigate high-stakes, uncertain environments.
Key Foundational Work
A Trust Model for Supply Chain Management (2010) — contributed 3 concept elements to this direction.
Analysing the Relationship between Risk and Trust (2004) — contributed 1 concept elements to this direction.