User Trust in AI Systems

• User trust in AI systems is a multidimensional construct defined by the willingness to accept vulnerability amid uncertainty, integrating competence, affective, and procedural dimensions.
• Measurement of trust combines self-report instruments, behavioral proxies, and dynamic observations to accurately capture and calibrate user reliance.
• Active trust management leverages adaptive interventions such as layered explanations and outcome feedback to optimize system reliability and ethical alignment.

User trust in AI systems is a multidimensional construct that governs whether, how, and under what conditions individuals and organizations are willing to rely on automated decision-making or recommendations. Across domains such as finance, healthcare, security, and consumer services, the calibration of user trust—its formation, measurement, manipulation, and repair—drives the uptake, effectiveness, and ethical alignment of AI in human–machine teams. This article lays out the contemporary scientific landscape on user trust in AI systems, integrating theoretical models, formal definitions, empirical findings, measurement methodologies, and actionable guidelines.

1. Theoretical Foundations and Definitions

The scientific consensus distinguishes sharply between trust (the user’s attitude or willingness to be vulnerable to the system’s actions) and trustworthiness (the system’s ability, integrity, and benevolence, often evidenced through technical and organizational attributes) (Duenser et al., 2023). Trust is neither reducible to system performance nor to user compliance; it is fundamentally relational, involving acceptance of uncertainty, vulnerability, and risk (Sun et al., 11 Oct 2025). Prevailing models decompose trust into rational (competence, reliability), relational (affective, partner-like engagement), and procedural (adherence to ethical standards, accountability) components.

Key formalizations include:

• Mayer et al.: Trust is “the willingness of a party to be vulnerable to the actions of another party based on the expectation that the other will perform a particular action important to the trustor” (Bach et al., 2023).
• Lee & See: “Trust is the attitude that an agent will help achieve an individual’s goals in situations characterized by uncertainty and vulnerability” (Sun et al., 11 Oct 2025).
• In generative AI, validated constructs are: Affective Trust, Competence Trust, Benevolence & Integrity, and Perceived Risk, as captured by the HAITS scale (Sun et al., 11 Oct 2025).
• Appropriate reliance (calibrated trust) is the alignment between user reliance and actual system correctness; disuse and overtrust are failure modes (Visser et al., 2023).

Multidimensional models (e.g., MATCH (Liao et al., 2022)) further articulate Ability (system performance), Intention/Benevolence, Process Integrity, and the communicative mechanisms (affordances, cues, heuristics) through which these attributes are revealed and interpreted by users.

2. Antecedents and Influencing Factors

User trust is shaped by interacting antecedents across three axes:

• User characteristics: Trust propensity, prior experience, domain expertise, motivation, personality traits (e.g., high neuroticism, low openness), self-confidence, and socio-cultural background (Bach et al., 2023, Amugongo et al., 18 Aug 2025, Sun et al., 11 Oct 2025). Afro-relational perspectives, for example, embed trust in communal, bidirectional accountability rather than the system-centric constructs prevalent in WEIRD societies (Amugongo et al., 18 Aug 2025).
• System attributes: Perceived reliability, accuracy, transparency, fairness, explainability, uncertainty quantification, anthropomorphic cues, and adaptability of the AI (McGrath et al., 2024). The extent and modality of transparency—basic feature importance, contextual explanations, interactive counterfactuals—differentially impact both trust and user cognitive load (Sunny, 17 Oct 2025).
• Contextual/environmental factors: Task risk, regulatory regimes, organizational trust, domain specificity, and the broader socio-technical ecosystem (regulations, governance structures, community norms) (Duenser et al., 2023, Alalawi et al., 2024).

Empirical studies confirm that user characteristics often dominate, with factors such as education, culture, and application domain explaining significant variance in trust scores (Amugongo et al., 18 Aug 2025, Bach et al., 2023).

3. Measurement and Quantification of User Trust

Measurement approaches fall into three major categories:

• Self-report instruments: Multi-item Likert scales (e.g., HAITS (Sun et al., 11 Oct 2025), Trust in Automation, Jian 2000, bi-factor scales for trust/distrust (Visser et al., 2023)). Composite trust scores are typically formed as the mean or weighted sum of item ratings: $T = \frac{1}{N}\sum_{i=1}^N t_i$ (Sunny, 17 Oct 2025). High reliability (Cronbach’s $\alpha > 0.88$) is typical for well-constructed scales (Sunny, 17 Oct 2025).
• Behavioral proxies: Reliance/agreement rates, “weight of advice” (WoA), task accuracy, override rates, and performance–trust discrepancies (Ahn et al., 2021, McGrath et al., 2024). Behavioral indices offer objective insight but may diverge from self-reported attitudes.
• Dynamic and high-resolution observation: Real-time analytics of trust over conversational turns (e.g., VizTrust tracks competence, benevolence, integrity, predictability per utterance, then aggregates as $T^t = \frac{1}{4}(c^t + i^t + b^t + p^t)$) (Wang et al., 10 Mar 2025).

Calibration metrics—e.g., calibration error ($$\mathrm{CalError} = \frac{1}{N}\sum_{i=1}^N |\mathrm{Trust}_i - \omega_i|$$), agreement/switch rates, and performance anchoring—distinguish mere trust from appropriate, risk-aligned reliance (Visser et al., 2023).

4. Effects of Explanations and Feedback on Trust

Explainability is widely assumed to raise trust, but meta-analysis and experimental results demonstrate only a moderate average correlation ($\bar{r} \approx 0.20$) between explainability and user trust (Atf et al., 16 Apr 2025). Interactive and contextual explanations yield higher trust than basic feature-importance lists, with interactivity supporting engagement and “agency” (Sunny, 17 Oct 2025). However, outcome feedback—informing users of the system’s actual accuracy post hoc—often has a more robust, reliable impact on trust formation than interpretability per se (Ahn et al., 2021).

Explanation effectiveness is strongly moderated by user expertise and context:

• Experts benefit from deeper, technical explanations; novices prefer concrete, example-based rationales (Sunny, 17 Oct 2025, Visser et al., 2023).
• High-stakes domains (e.g., healthcare, justice) intensify the user demand for transparent, compositional, and error-aware rationales (Atf et al., 16 Apr 2025).
• In high-risk settings with fundamentally biased or untrustworthy AI (e.g., predictive policing), explanations may increase perceived trust without improving appropriate reliance, and can sometimes induce unwarranted confidence or confirmation bias (Mehrotra et al., 15 Apr 2025).

Notably, explanation presentation form (textual, visual, hybrid) influences subjective but not always appropriate trust; hybrid explanations may increase experts’ subjective trust but have no effect on decision accuracy (Mehrotra et al., 15 Apr 2025).

5. Trust Calibration, Dynamics, and Active Management

Trust is not static—it evolves over time through repeated interaction, affected by error exposure (especially confidently incorrect outputs), miscalibration, and experiences of repair (Dhuliawala et al., 2023). Persistent overconfidence by the AI system produces sharp, long-lasting decrements in user trust, often generalizing across unrelated tasks (Dhuliawala et al., 2023). Real-time, multi-method tracking is required for effective calibration.

Active trust management frameworks (e.g., CHAI-T) call for:

• Phased models of trust evolution: formation, calibration, maintenance, and repair (McGrath et al., 2024).
• Adaptive interventions based on real-time user trust signals, e.g., providing supporting explanations to users with low trust and counter-explanations (“why might this recommendation be wrong?”) to high-trust users, yielding up to 38% reductions in inappropriate reliance and up to 20% gains in accuracy (Srinivasan et al., 18 Feb 2025).
• Instrumentation of trust–performance discrepancy and dynamic recalibration through explicit feedback or “cognitive forcing functions” (Srinivasan et al., 18 Feb 2025, McGrath et al., 2024).

Measurement and adjustment at the socio-technical system level—incorporating all stakeholder interactions, regulatory affordances, and trust-enabling constructs (explanation utility, interactive recourse, team cognition)—is essential for sustainable, appropriate trust (Benk et al., 2022).

6. Socio-Technical and Ethical Dimensions

Trust in AI is embedded in a broader socio-technical regime. Regulatory incentives, procedural trust, transparency of governance, and organizational reputation significantly shape both perceived trustworthiness and user trust (Duenser et al., 2023, Alalawi et al., 2024). Evolutionary game-theoretic modeling reveals that mechanisms such as regulator reputation and user-conditioned trust can establish stable equilibria where creators are incentivized to deliver trustworthy systems, regulators enforce standards, and users adopt AI with justified confidence—contingent on regulatory cost structures and system risk (Alalawi et al., 2024).

Culturally situated perspectives (e.g., Afro-relational models) expand the trust discourse beyond technical or individual frames to relational, bidirectional, moral, and community-centered axes, requiring co-design, shared governance, and trust cues adapted to local values (Amugongo et al., 18 Aug 2025).

Trustworthiness cues must be screened for warrantedness (truthfulness, relevance, calibration) and costliness (“expensive” cues, such as certification or audited documentation, are harder to fake and thus more credible) (Liao et al., 2022).

7. Design and Evaluation Guidelines

Synthesizing across domains, the following design tenets, grounded in empirical evidence, are recommended for fostering calibrated user trust:

• Provide layered, context-sensitive explanations: concise rationales for novices, technical drill-down for experts (Sunny, 17 Oct 2025, Atf et al., 16 Apr 2025).
• Optimize for interactivity: enable limited “what-if” explorations, interactive dashboards, and user-driven critique (Sunny, 17 Oct 2025, Benk et al., 2022).
• Integrate outcome feedback where feasible to reinforce trust calibration and learning (Ahn et al., 2021, Dhuliawala et al., 2023).
• Monitor reliance and trust dynamically, using behavioral and attitudinal measures, and intervene adaptively to correct over/undertrust (Srinivasan et al., 18 Feb 2025, McGrath et al., 2024, Wang et al., 10 Mar 2025).
• Align explanation complexity and trust interventions to domain risk and user expertise, avoiding one-size-fits-all approaches (Visser et al., 2023, Mehrotra et al., 15 Apr 2025).
• Embed procedural and organizational transparency, including disclosure of governance, audit outcomes, and failure modes (Duenser et al., 2023).
• Recognize and address cultural and communal factors; co-design governance, explanations, and feedback with representative user communities (Amugongo et al., 18 Aug 2025).
• Evaluate trust using validated, multi-method metrics (self-report, behavior, calibration error, temporal tracking) and across the full user–AI–environment ecosystem (Sun et al., 11 Oct 2025, Benk et al., 2022).
• Prioritize expensive, hard-to-fake trustworthiness cues in high-assurance domains (Liao et al., 2022).
• Shift regulatory policy toward systems that enable and protect “appropriate” rather than merely “subjective” trust, with contestability and recourse provisions (Mehrotra et al., 15 Apr 2025, Alalawi et al., 2024).

In sum, effective engineering of user trust in AI systems requires an integrated, context-aware, and dynamically managed approach—one that simultaneously engages with cognitive, affective, procedural, and socio-technical points of leverage, always grounding user reliance in warranted and transparent system capability.