When we think about trusting a piece of technology, we usually imagine a single linear slider. On the far left is complete distrust; on the far right is blind faith. Under this monolithic view, the goal of artificial intelligence design is simple: improve the model's accuracy or explainability to push the user's slider as far to the right as possible.
However, real-world interactions with AI consistently break this simple mental model. Users frequently say they trust an AI system yet refuse to delegate tasks to it. Instructors and students often utilize generative AI tools daily while remaining deeply suspicious of their outputs.
To explain these contradictions, a growing body of research argues that trust is not a single continuum. Instead, human-AI interaction is governed by bivariate and multi-dimensional trust models—frameworks that treat trust and distrust as independent, coexisting psychological forces, each driven by entirely different cognitive mechanisms.
The Bivariate Nature of Trust and Distrust
The foundational flaw of the traditional linear model is the assumption that distrust is merely the absence of trust. In reality, trust and distrust are qualitatively different psychological phenomena that operate through distinct cognitive and neurological pathways.
As argued by Wang in Cyber Trust and Suspicion: A Human-Centric Approach, trust is a reward-based prediction characterized by hope and assurance. Distrust, conversely, is an emotion-driven fear of loss characterized by skepticism and vigilance. Because these two states trigger different parts of the human brain, they are not mutually exclusive; they can operate simultaneously.
Traditional Monolithic View:
Distrust <----------------------------------------------> Trust
Bivariate View:
High Trust | "Trust but Verify" Blind Faith
| (High Trust, High Distrust) (High Trust, Low Distrust)
|
-------------+----------------------------------------
| Complete Apathy Deep Suspicion
Low Trust | (Low Trust, Low Distrust) (Low Trust, High Distrust)
+----------------------------------------
| High Distrust
This bivariate architecture explains why users often adopt a "trust but verify" mindset. As documented by Lyu et al. in Understanding the practices, perceptions, and (dis)trust of generative AI among instructors and surveyed by Visser et al. in Trust, distrust, and appropriate reliance in (X)AI, individuals can simultaneously hold both positive and negative sentiments toward a technology. A user can trust an AI's capability to generate a draft while maintaining a high level of active distrust (vigilance) to catch potential hallucinations. This explains why student populations do not show uniform acceptance of generative tools, but rather a complex, non-linear mixture of trust, neutrality, and skepticism, as observed by Amoozadeh et al. in Trust in Generative AI among Students.
The Two Axes: Capability vs. Character
Beyond the split between trust and distrust, the concept of trust itself bifurcates into two primary dimensions: the rational/functional and the emotional/ethical.
In Understanding Trust and Willingness to Use GenAI Tools in Higher Education, Guo et al. categorize these dimensions as system-like trust (rational and functional) and human-like trust (emotional and ethical). A similar taxonomy is proposed by Tahaei et al. in Human-Centered Responsible Artificial Intelligence, who distinguish between:
- Capability Trust: Confidence in the system's technical competence and accuracy.
- Moral Trust: Alignment of the system's actions with human ethical values and social norms.
This distinction is critical because optimizing for one axis does not automatically satisfy the other. For instance, Singh et al. in Trustworthy AI point out that an AI system can be mathematically robust and ethically compliant, yet remain socially abrasive or impolite. If the system violates social conduct, users will withhold their trust, regardless of the model's high technical competence.
Conversely, these dimensions can buffer one another. Shang et al. in Trusting Your AI Agent Emotionally and Cognitively demonstrate that while cognitive trust (intellect) drives initial adoption, affective trust (perceptions of empathy and warmth) acts as a psychological safety net. When the AI's technical logic falters, high affective trust prevents the immediate collapse of user cooperation.
This dual-axis relationship is further complicated by what Lu describes in Rethinking Trust in Human-AI Collaboration in the Generative AI Era as "agent-tool oscillation." Because generative AI models mimic human language, users constantly fluctuate between treating the system as a functional tool (requiring evidence-based capability trust) and a social collaborator (requiring relationship-based moral trust), muddying how expectations are formed and maintained.
The "Say-Do" Gap and Behavioral Realities
When researchers fail to account for the multi-dimensional nature of trust, they encounter the "say-do" gap: a sharp divergence between a user's self-reported trust and their actual behavior.
In The Dynamics of Trust in XAI, Abbaspour Onari et al. highlight the necessity of separating reliance (acting on an AI's recommendation while maintaining control) from trust (completely delegating decision-making authority). They observed that even when users reported high subjective trust in an AI that significantly outperformed them, they only delegated authority in 35% of tasks.
This behavioral divergence was also captured by Papenmeier et al. in How model accuracy and explanation fidelity influence user trust in AI. By tracking how users adjusted their decisions in response to AI outputs (implicit trust) alongside questionnaires (explicit trust), the authors found that high-fidelity explanations can actually decrease behavioral trust in highly accurate models. The explanation exposes minor, irrelevant edge cases in the model's logic, which triggers the independent distrust dimension and causes users to override correct AI decisions.
Traditional technology acceptance frameworks fail to capture these nuances because they treat technology as a passive tool. As Tian et al. argue in Investigating the actual usage behavior of Thai traditional craftsmen in adopting AI, AI requires the inclusion of unique psychological variables like perceived intelligence and anthropomorphism. Furthermore, Zhang shows in U.S. Public Opinion on the Governance of Artificial Intelligence that institutional trust does not predict public support for AI development. The public perceives AI as a uniquely complex, independent force that operates outside of the organizations creating it, requiring a dedicated, multi-dimensional approach to trust management.
Designing for Calibrated Trust
Why does understanding the multi-dimensional structure of trust matter for system design? Because it changes the objective of Human-AI interaction from maximizing trust to calibrating it. Blind trust leads to dangerous overreliance on flawed systems, while unmitigated distrust leads to the rejection of useful tools.
By treating trust and distrust as independent axes, designers can intentionally foster a state of healthy skepticism—maintaining high capability trust to utilize the system's strengths, while preserving a baseline of distrust (vigilance) to verify outputs.
To achieve this calibration, Chau et al. in Interactive Visualization for Fostering Trust in ML propose using interactive visualizations. Instead of presenting AI as an opaque oracle or an overwhelming wall of code, interactive interfaces allow users to explore the boundaries of a model's logic. This bridges the gap between evidence-based trust (verifying the technical artifact) and relationship-based trust (feeling comfortable with the interaction), giving users the precise cognitive tools needed to trust, verify, and safely collaborate with artificial intelligence.
Key Foundational Work
Cyber Trust and Suspicion: A Human-Centric Approach (2013) — contributed 3 concept elements to this direction.
U.S. Public Opinion on the Governance of Artificial Intelligence (2019) — contributed 1 concept elements to this direction.
How model accuracy and explanation fidelity influence user trust in AI (2019) — contributed 1 concept elements to this direction.