In many AI safety and human-computer interaction frameworks, "trust" is treated as a monolithic scalar value—a single number representing how much a user relies on a system. However, a growing body of research suggests that this view is too reductive. If a self-driving car fails to stop, do we distrust its sensors (its ability) or its programming priorities (its "willingness" to protect us)?
To build truly reliable systems, researchers are moving toward a multidimensional model of trust that separates an agent's technical competence from its intentionality or moral alignment.
The Two-Axis Model of Trust
The core insight across recent literature is that trust is not a single feeling but a composite of two distinct evaluations. Yasaman Haghpanah and Marie desJardins formalize this in the context of supply chain management by decomposing trust into competence (the physical or technical capability to fulfill a task) and integrity (the willingness to honor a commitment even when incentives favor defection).
This distinction is critical because the "remedy" for a loss of trust depends entirely on which axis failed:
- Competence failures are "skill issues." They are solved through better training data, more robust sensors, or improved algorithms.
- Intentionality failures are "will issues." They are solved through value alignment, ethical constraints, and transparency.
Xiaowei Huang and Marta Kwiatkowska have even developed a formal logic called PRTL to mathematically verify these dimensions in multi-agent systems. Their framework allows developers to quantify not just if an agent can* reach a goal, but whether its "subjective evaluation" of the task makes it likely to cooperate in uncertain environments.
Cognitive vs. Affective Trust
While the competence/integrity split works well for game theory, human-AI interaction requires a more psychological lens. Ruoxi Shang, Chirag Shah, and Gary Hsieh distinguish between cognitive trust (intellectual reliance on performance) and affective trust (emotional bonds based on empathy and warmth).
Their research highlights a fascinating "safety net" effect: while high performance is the primary driver of trust, a strong affective bond can sustain a user's trust even when the AI’s logic momentarily falters. Conversely, Guanxiong Pei and colleagues note that treating trust as a monolith obscures how different agent behaviors—like conversational warmth versus factual precision—influence consumer decision-making.
The Tension Between Credibility and Resonance
This decomposition reveals inherent trade-offs in AI design. In a study on LLM-based support for eating disorder recovery, Ryuhaerang Choi and Sung-Ju Lee found a tension between epistemic trust (credibility) and emotional resonance. Users found "peer-like" personas more comforting for validation but preferred neutral, "lay-mentor" personas for practical, objective guidance.
This suggests that "maximizing trust" is the wrong goal; the goal should be to calibrate the right kind of trust for the specific task at hand.
Trust in the Presence of Bias
The importance of this bifurcated framework is most visible when studying marginalized populations. Aimen Gaba and Emily Wall found that while different gender groups might agree on an LLM's technical accuracy (performance-based trust), their morality-based trust varies wildly. Non-binary users, for instance, may recognize a model is "smart" (competent) while simultaneously finding it "untrustworthy" because it produces condescending or reductive responses regarding their identity.
As Richa Singh and Nalini K. Ratha argue, "Trustworthy AI" cannot be achieved through accuracy alone. It requires a "Decent AI" that adheres to social norms and ethical alignment. By moving away from monolithic trust and toward these nuanced dimensions, researchers can better predict when a user will rely on a system, when they will abandon it, and how to build agents that are both capable and benevolent.