Adaptive Explainability Trust Framework (AXTF)

• AXTF is a framework that calibrates trust by adapting explanation timing, granularity, and content based on system performance and user state.
• It integrates multimodal implicit sensing with neuro-fuzzy modeling to assess workload, stress, and emotional valence for dynamic explanation adaptation.
• The framework emphasizes appropriate trust, ensuring explanations are tailored to reflect actual system trustworthiness and promote accountable human-AI interaction.

Searching arXiv for the cited papers to ground the article and verify metadata. Using arXiv search for the core paper and related work on trust, explainability, and adaptive explanation. Adaptive Explainability Trust Framework (AXTF) denotes a trust-oriented approach to explainable AI in which explanations are selected, timed, and structured so that human reliance better matches actual system trustworthiness. In its explicit high-stakes formulation, AXTF is a conceptual closed-loop framework that combines multimodal implicit sensing, a personalized trust inference model, and trust-sensitive adaptation of explanation features such as timing, duration, granularity, content, transparency, and mode of delivery (Fernando et al., 25 Jul 2025). In the surrounding literature, the same label is also supported as a broader design stance: explainability is treated not as an end in itself, but as a mechanism for trust calibration, uncertainty communication, robustness signaling, stakeholder adaptation, and accountable human-AI interaction (Newen et al., 10 Sep 2025, Kästner et al., 2021).

1. Conceptual foundations

AXTF is grounded in a distinction between trust and trustworthiness. Trust is treated as an attitude a stakeholder holds toward a system, whereas trustworthiness is treated as a property of the system. A formal trustworthiness condition states that a system $S$ is trustworthy to a stakeholder $H$ in a context $C$ if and only if $S$ works properly in $C$, and $H$ would be justified to believe that it works properly in $C$ if $H$ came to believe that (Kästner et al., 2021). In this framing, explainability matters because accurate explanations can help stakeholders gain the understanding needed for justified belief, but explainability is not itself equivalent to trustworthiness.

This distinction is reinforced by quantitative evidence. A meta-analysis of 90 studies reported a statistically significant but low positive correlation between explainability and trust, with a random-effects correlation of approximately $r = 0.194$ to $0.196$, and substantial heterogeneity at $H$0 (Atf et al., 16 Apr 2025). AXTF therefore aligns more naturally with trust calibration than with trust maximization. The literature repeatedly warns that explanations may raise trust, leave it unchanged, or reduce it, depending on stakeholder background, explanation form, domain, and the underlying quality of the system.

A recurrent implication is that AXTF should be understood as a framework for appropriate trust. This suggests a dual requirement: the system should be worthy of trust, and the explanation policy should help stakeholders recognize when trust is warranted, when caution is warranted, and when reliance should be withheld.

2. Architectural structure

The explicit AXTF proposal for high-stakes environments is organized as a closed-loop pipeline. It integrates physiological and behavioral signals with environmental data such as task goals, urgency, and state in order to assess workload $H$1, stress $H$2, and emotional valence $H$3; it also uses a normalized system performance score $H$4 (Fernando et al., 25 Jul 2025). These variables are mapped through a multi-objective neuro-fuzzy model to a trust state $H$5. The resulting trust estimate guides explanation adaptation.

The sensing layer is intentionally multimodal. Candidate inputs include EEG, ECG, GSR, HRV, pupillometry, gaze tracking, facial expressions, voice features, and task-switching patterns. AXTF is explicitly designed to rely on implicit feedback rather than explicit user reports, because high-pressure settings may not permit verbal queries or deliberate feedback entry. The framework is therefore non-intrusive in the specific sense that it reacts to user state through passive sensing rather than demanding additional cognitive effort from the operator (Fernando et al., 25 Jul 2025).

The adaptation layer modifies seven explanation features identified in the framework: timing, duration, granularity, content, transparency, adaptability, and mode of delivery. Timing can be proactive or reactive; duration can be shortened under high workload; granularity can shift between high-level summaries and step-by-step detail; content can remain local or expand hierarchically; transparency can emphasize either how the AI produced an output or why it selected an action; delivery can be visual, textual, auditory, or multimodal. The architectural logic is therefore not merely “attach an explanation,” but “configure an explanation policy as a function of trust-relevant human and environmental state.”

3. Adaptation, user modeling, and multilevel explainability

AXTF is closely aligned with a broader movement from static explanation artifacts toward layered and user-centered explanation systems. One three-layered HCXAI framework organizes explanation as: a foundational AI model with built-in explainability mechanisms, a human-centered explanation layer tailored to user expertise and cognitive load, and a dynamic feedback loop that refines explanations through real-time interaction (Silva et al., 14 Apr 2025). A related multilevel framework separates algorithmic and domain-informed explainability, human-centered and interactive explainability, and social explainability aimed at public accountability and natural-language accessibility (Bello et al., 6 Jun 2025). AXTF inherits the same basic separation between explanation generation, explanation delivery, and explanation refinement.

Earlier work on explainability levels anticipates this architecture. A five-level framework moves from no explanation, to explicit explanation representation, to alternative explanations, to knowledge of the explainee, and finally to interactive explanation (Atakishiyev et al., 2020). AXTF is most naturally situated at the upper end of that scale, because it assumes both explainee modeling and adaptive interaction. This suggests that AXTF is not just a presentation strategy; it is an orchestration framework that decides which explanation should be shown, to whom, at what level of detail, and under what constraints.

In strategic and expert domains, adaptation also has a domain-level dimension. For superhuman DRL-based systems, local action explanations are argued to be insufficient; expert users require domain-level explainability organized around future projections, hypothetical scenarios, risk, transparency and safety, and uncertainty, formalized as $H$6 through $H$7 in the literature (Andrulis et al., 2020). A plausible implication is that AXTF should vary not only by user expertise but also by task horizon: local views for case assessment, global views for audit, and domain-level scenario reasoning for strategic reliance.

4. Trust calibration through uncertainty, robustness, and fairness

AXTF is especially strongly supported by work that treats explainability as a mechanism for calibrating trust to actual model reliability rather than improving explanation satisfaction. In one study, trust calibration is explicitly defined as “the balance between human trust and the actual trustworthiness of the application,” and is operationalized behaviorally as whether objectively better learners are trusted more than worse ones (Newen et al., 10 Sep 2025). That study argues that explainability should communicate not only why a model behaves as it does, but also where it is uncertain and where it may lack robustness. It further distinguishes local explanations from global explanations and shows that both can be useful: some participants relied on dataset-level uncertainty structure, while others verified whether uncertain and certain labels made sense on concrete examples.

The same study is notable for showing that several trust-related outcomes are not equivalent. Participants rated learners on trustworthiness, reliability, competence, accuracy, and understandability; they separately rated trust in the explanation itself; and they separately reported explanation satisfaction. These measures moved differently. An explanation could improve trust calibration of model choice while not increasing satisfaction and while even increasing perceived obscurity of the explanation interface (Newen et al., 10 Sep 2025). AXTF therefore treats calibrated model trust, trust in explanation, and explanation satisfaction as distinct targets.

Clinical evidence further sharpens this point. In a breast-cancer clinical decision support experiment, high confidence scores substantially increased trust but also led to overreliance, reducing diagnostic accuracy, whereas low confidence scores decreased trust and agreement and increased diagnosis duration (Rezaeian et al., 28 Jan 2025). Some explainability conditions also increased stress. This directly supports an AXTF design rule: confidence displays and richer explanation are not intrinsically beneficial; they regulate behavior and can either improve calibration or induce automation bias.

Fairness-sensitive work extends the same logic. In visualizations of biased models, more comprehensible explanations increased perceived bias and reduced trust; the mediation path was reported as $H$8 for comprehension predicting bias perception and $H$9 for bias perception predicting trust (Kaufman et al., 31 Jul 2025). The same paper shows that visualization design can raise trust either by improving actual fairness or by merely reducing perceived bias without changing model behavior. This suggests that AXTF needs explicit safeguards against explanation policies that increase trust by hiding defect visibility rather than by improving system quality.

5. Empirical evaluation and measurement

AXTF evaluation is inherently multidimensional. The literature repeatedly rejects single-score evaluation and instead separates several constructs.

Before the table, one empirical distinction is especially important: explanation effects are heterogeneous across populations and domains. In a cross-cultural chatbot study with $C$0, Korean participants showed higher trust, more positive AI user experience, and more opportunity-oriented AI perception than German participants; political topics yielded lower trust than entertainment topics in exploratory regression; and topic-sensitive interactions appeared in explainability effects, with higher explainability improving user experience for entertainment and politics but not for health (Kang et al., 21 Mar 2025). A preliminary loan-approval study likewise reported a monotonic trust ordering from no explanation ($C$1) to feature importance ($C$2), contextual explanation ($C$3), and interactive explanation ($C$4), though it remained a small-$C$5 preliminary study (Sunny, 17 Oct 2025). Together, these findings support AXTF’s emphasis on adaptive rather than uniform explanation policies.

Construct	Operationalization in the literature	AXTF implication
Trust in model or learner	Ratings of trustworthiness, reliability, competence, accuracy, understandability	Primary calibration target
Trust in explanation	Adapted Trust in Automation items about the explanation	Must be tracked separately
Explanation satisfaction	Explanation Satisfaction Scale	Not a proxy for calibrated reliance

The same uncertainty-aware study that motivates this separation reports acceptable reliabilities for its scales, including Cronbach’s alpha values of $C$6 for explanation satisfaction, $C$7 for trust in explanation, and $C$8 for distrust in explanation (Newen et al., 10 Sep 2025). This measurement strategy is methodologically important for AXTF because it prevents explanation satisfaction from being mistaken for justified reliance. A plausible implication is that mature AXTF evaluation should combine subjective scales, behavioral differentiation between better and worse systems, uncertainty-sensitive reliance, and qualitative diagnostics about confusion points in explanation encoding.

6. Scope, limitations, and extensions

AXTF remains partly conceptual across much of the literature. The high-stakes AXTF proposal is explicitly a conceptual framework rather than a fully implemented or empirically validated system (Fernando et al., 25 Jul 2025). Related HCXAI and multilevel explainability frameworks are also architecture-oriented and normative rather than algorithmically complete; they do not supply formal trust-state estimators, explicit adaptation policies, or longitudinal trust models (Silva et al., 14 Apr 2025, Bello et al., 6 Jun 2025). This suggests that current AXTF research is strongest on requirements and architecture, and weaker on online policy learning, deployment governance, and long-term calibration under drift.

Several extension paths are already visible. For agentic LLMs, TAXAL proposes a triadic alignment of cognitive, functional, and causal dimensions, emphasizing that explanations must be understandable to humans, useful in workflows, and faithful enough for scrutiny (Herrera-Poyatos et al., 5 Sep 2025). For strategic DRL, domain-level explainability emphasizes future projections, hypothetical scenarios, risk, and uncertainty rather than only local action justification (Andrulis et al., 2020). For trust-oriented systems engineering more generally, the literature argues that systems should be engineered primarily for trustworthiness, with explainability contributing to the justificatory side of trustworthiness rather than serving as a tool for trust elicitation alone (Kästner et al., 2021).

A plausible implication is that a fully developed AXTF would need at least five additional capabilities beyond current proposals: formal trust-state estimation, explicit anti-manipulation constraints on explanation policy, online adaptation under changing user state and distribution shift, longitudinal modeling of trust formation and repair, and audit-ready governance over why a particular explanation was shown to a particular stakeholder. Under that broader interpretation, AXTF is best understood not as a single algorithm, but as an emerging research program for making explainability adaptive, uncertainty-aware, robustness-aware, fairness-aware, and accountable across heterogeneous human-AI settings.