Explainable Information Design

• Explainable information design is the intentional organization of explanatory content in AI systems that tailors information to diverse stakeholder needs.
• It involves creating artifacts, interaction patterns, and governance mechanisms that shift focus from mere model internals to context-specific, actionable insights.
• The approach emphasizes combining technical methods with user-centered design to improve accountability, trust calibration, and ethical decision making.

Searching arXiv for the cited work on explainable information design and adjacent XAI design frameworks. Explainable information design denotes the deliberate design of explanatory information in AI-supported systems: what explanatory content is provided, to whom, in what form, at what moment, and for which practical purpose. Recent work treats it not as a narrow problem of exposing model internals, but as the design of information artifacts, interaction structures, provenance traces, and governance mechanisms that mediate between computational systems and heterogeneous stakeholders. In this view, explanations are expected to support understanding, calibrated trust, scrutiny, action, and accountability within concrete sociotechnical settings rather than merely to “open the black box” in the abstract (Dhar et al., 12 Aug 2025, Zhang et al., 3 Jul 2025, Huynh et al., 2022).

1. Conceptual foundations

A central shift in the literature is the move beyond what one paper calls “technocratic XAI”: the tendency to equate explanation with the direct output of a technical method such as saliency maps, SHAP values, LIME outputs, surrogate models, or mechanistic probes. In contrast, explanation is framed as a situated design process spanning interrogation of models, selection of relevant technical insights, and communicative delivery in forms that are meaningful to specific stakeholders. A complementary line of work describes explainability as a holistic information strategy that includes onboarding explanations, moment-of-use explanations, error and failure explanations, personalization transparency, ongoing help materials, audit and incident reports, and participatory communication (Dhar et al., 12 Aug 2025, Zevenbergen et al., 2020).

A survey of human-centered explainability in interactive information systems organizes this design space around five dimensions: where explainability is needed, why it is provided, to whom it is directed, what is included in the explanation, and how it is conveyed. The same survey separates explainability as a system capability from related notions such as transparency, interpretability, understandability, controllability, trust, and accountability. This distinction matters because a technically transparent system can still fail as an explanatory system if its information is poorly timed, misaligned with user goals, or inaccessible to the intended audience (Zhang et al., 3 Jul 2025).

A recurring misconception is that explainability is exhausted by local model rationales. The case-study literature argues that explanations should be understood as any information that helps people understand and appropriately respond to an AI system over time. This includes information about capabilities, limitations, adaptation, uncertainty, and institutional response. Another misconception is that explanation is neutral. Several papers explicitly treat explanation design as an ethical and political act because choices about what is shown, hidden, simplified, or made contestable shape power, autonomy, and responsibility (Zevenbergen et al., 2020, Dhar et al., 12 Aug 2025).

2. Stakeholders, questions, and design processes

A consistent theme is that explanation design is stakeholder-specific. The “Who–What–How” framework distinguishes developers, operators, validators, and subjects, and stresses that roles should be treated as goal-driven rather than title-driven. Developers require detailed, technically faithful information for debugging and refinement; operators often need fast, local, and domain-linked rationales; validators need structured documentation, global behavior summaries, fairness information, and evidence trails; subjects need understandable reasons, recourse, and visibility into limitations and rights (Dhar et al., 12 Aug 2025).

Question-driven approaches make this stakeholder specificity operational. The “Question-Driven Design Process” centers explanation around prototypical user questions such as Output, Why, Why not, How, What if, How to be that, Performance, Data, and Model facts. The process proceeds through question elicitation, question analysis, mapping questions to modeling solutions, and iterative design and evaluation. A closely related question-bank approach uses the same logic as a boundary object between designers and AI engineers: user needs are represented as question categories, and technical methods are selected only after those question categories are prioritized (Liao et al., 2021, Liao et al., 2020).

This process orientation broadens explanation design from interface ornamentation to workflow design. One practical implication is that explanations must be distributed across the product lifecycle. The case-study literature argues for explanations at onboarding, at the moment of decision, during and after errors, in public-facing reports, and in participatory settings where communities need aggregate information and scenario exploration tools. This suggests that explainable information design is as much about information timing and organizational placement as about representational form (Zevenbergen et al., 2020).

A further extension appears in the UXR playbook literature, which groups design problems into plays such as choosing the explanation class, understanding user engagement and retention, acknowledging cognitive biases and misuse, and creating user-friendly and intuitive XAI. That work emphasizes that explanation methods such as SHAP, LIME, and counterfactuals are raw technical materials; explanation design begins when those materials are transformed into layered, contextual, bias-aware information structures that fit human goals and cognitive constraints (Naiseh et al., 19 Jun 2025).

3. Explanation artifacts, modalities, and interaction patterns

The literature classifies explanation artifacts along several orthogonal dimensions. One design space for ranking systems uses item count, attribute count, attribute type, visual representation, attribute aggregation, explanation level, model dependency, and model interpretability. End users are typically associated with local, item-first, low-aggregation, low-complexity explanations; analysts and model developers need broader attribute coverage, richer aggregation, and global views. This formulation makes explicit that explainable information design depends on how much of the information space is exposed and how that space is organized (Hazwani et al., 2022).

Human-centered work on end-user explainability proposes twelve end-user-friendly explanatory forms organized into feature-based, example-based, rule-based, and supplementary information. The forms include feature attribution, feature shape, feature interaction, similar examples, typical examples, counterfactual examples, decision rules, decision trees, input, output, performance, and dataset information. These forms are not treated as mutually exclusive; participants in co-design studies combine them to satisfy explanation goals such as calibrating trust, ensuring safety, detecting bias, resolving disagreement, learning from AI, improving outcomes, and communicating with others (Jin et al., 2021).

Interactive explanation systems push this further by varying level of detail and giving users control over explanation density and focus. In a literature recommender, explanations are designed at basic, intermediate, and advanced levels by combining intelligibility types such as What, Why, What if, and How. Basic explanations emphasize input and high-level rationale; intermediate explanations add more detailed rationale; advanced explanations expose algorithmic process. Users can invoke “WHAT-IF?”, “WHY?”, and “HOW?” controls to regulate complexity and personalize the explanation. Reported qualitative evidence links this interaction pattern to transparency, trust, satisfaction, and user experience (Guesmi et al., 2023).

Modality also matters. An information-theoretic comparison of voice and text treats explanation delivery as a communication channel and finds that text explanations achieve higher comprehension efficiency, while voice explanations yield improved trust calibration, with analogy-based delivery achieving the best overall trade-off. That framework models modality $M \in \{\text{text}, \text{voice}\}$ and style $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$ as parameters of the channel from model attribution vector $A$ to user representation $U$. A plausible implication is that modality choice is itself a core design variable rather than a surface presentation choice (Rajhans et al., 6 Feb 2026).

4. Formalization and optimization

Several strands of work formalize explainable information design. One strand models explanatory communication itself. In the voice-versus-text framework, explanation delivery is treated as a noisy channel from model attributions $A$ to a user’s internal representation $U$, with normalized mutual information

$$I_M = \frac{H(A) - H(A|U)}{H(A)} = \frac{I(A;U)}{H(A)}.$$

Cognitive load is modeled as

$L(M,S) = \alpha D(M,S) + \beta H(E(A,M,S)),$

and comprehension efficiency as

$CE(M,S) = \frac{I_M}{L(M,S)}.$

Trust calibration is captured by

$\mathrm{TCE} = \mathbb{E}[|T-Q|],$

and combined with comprehension efficiency in a composite score

$$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$0

This formalism recasts design trade-offs among fidelity, effort, and calibrated reliance in explicitly information-theoretic terms (Rajhans et al., 6 Feb 2026).

A second strand treats explainability-by-design as a systems architecture problem. The explainability-by-design methodology organizes work into Explanation Requirement Analysis, Explanation Technical Design, and Explanation Validation. Its technical core uses provenance patterns, provenance queries, explanation plans, explanation dictionaries, and a reusable Explanation Assistant. Logs are designed so they can be converted into provenance traces of decisions; those traces are queried; explanation plans render the results into personalized narratives. The resulting architecture treats explanatory information as a first-class product of system design rather than an after-the-fact add-on (Huynh et al., 2022).

A third strand analyzes explainability as a restriction on admissible information policies in Bayesian persuasion. In that setting, explainable signaling schemes are $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$1-partitional signaling schemes defined by deterministic and monotone partitions of a continuous state space. The paper defines the Price of Explainability as the ratio between the optimal explainable signaling scheme and the unrestricted optimum, and proves that the worst-case Price of Explainability is exactly $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$2. It also shows that exact optimization is NP-hard in general, that an $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$3-approximately optimal explainable signaling scheme can be computed in polynomial time for Lipschitz utility functions, and that for piecewise constant utility functions one can find a $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$4-approximation to the unrestricted optimum that matches the worst-case bound (Chen et al., 19 Aug 2025).

A fourth strand extends information design into adaptive multi-agent policy systems. There, explainability arises from structuring a multi-agent simulation into population, environment, behavioral, control, and diagnostics layers, and from using information-theoretic diagnostics such as entropy rate $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$5, statistical complexity $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$6, and predictive information $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$7. The paper’s regime taxonomy—CPCA, CPVA, VPCA, VPVA—makes adaptation structure explicit rather than embedding it implicitly in code. This suggests that explainable information design can operate not only at the interface level but also at the level of system architecture and experimental protocol (garrone, 24 Nov 2025).

5. Sociotechnical, legal, and governance dimensions

The sociotechnical literature argues that explanation design cannot be reduced to maximizing intelligibility or user trust. One paper identifies epistemic inequality, social inequality, and obscuring accountability and governance as central risks. Explanations that are available in rich form to developers or institutional actors but only weakly or opaquely to affected subjects can reinforce asymmetries of knowledge and power. In this perspective, explainable information design must decide not only how to expose model reasoning but also who is enabled to understand, contest, and act on that reasoning (Dhar et al., 12 Aug 2025).

Seamful XAI pushes this argument further by focusing on “mismatches and cracks between assumptions made in designing and developing AI systems and the reality of their deployment contexts.” Instead of hiding seams, seamful design treats them as resources for explanation. Its process begins with envisioning breakdowns, then identifying seams across the AI lifecycle, and finally deciding which seams to reveal to support actionability, contestability, and appropriation. The result is a model of explanation in which policy lag, data mismatch, infrastructure fragility, and workflow misalignment become explanation content rather than background conditions (Ehsan et al., 2022).

A legal extension formalizes this contestability requirement. Legally-informed XAI distinguishes legally informative information, which directly describes laws, regulations, rights, obligations, and liability, from legally actionable information, which can be used in legal proceedings even if it is not itself a legal rule. It organizes requirements around three stakeholder groups—decision makers, decision subjects, and legal representatives—and argues that explanations in high-stakes domains must support concrete action and pushback against AI determinations. This reframes explainable information design as part of legal communication infrastructure, not merely user assistance (Mansi et al., 14 Apr 2025).

Explainability-by-design contributes a complementary governance mechanism through provenance. Its definition of explanation explicitly includes enabling recipients to understand the decision-making process and, where necessary, to take action such as contesting a decision or correcting the process. The provenance-oriented architecture is therefore also a governance architecture: it renders data basis, influences, and responsibility auditable, queryable, and narratable (Huynh et al., 2022).

6. Evaluation, applications, and open problems

A systematic survey of 100 papers and 121 user studies shows that evaluation of explainable information design is already multi-dimensional. It categorizes measurements into intrinsic properties, format and presentation, usability, experiential outcomes, ethics, and interaction with explanations. The same survey reports that explanation designs vary across dashboards, desktop applications, web tools, chatbots, mobile apps, pop-up messages, games, virtual reality, and wearable devices, and across text, graphics, video, voice, and multimodal combinations. This indicates that explanation quality is inseparable from platform, interaction pattern, and usage context (Zhang et al., 3 Jul 2025).

Application domains in the literature are correspondingly broad. Recommendation research uses explainable and interactive structures to support cocoon-breaking, user control, and multiple levels of detail. Clinical decision support uses layered local and global explanations, uncertainty displays, and task-sensitive interaction. Policy simulation uses causal graphs, information-theoretic diagnostics, and clustering to make regimes legible and contestable. In each case, the explanation artifact is embedded in a larger information system rather than standing alone (Liu et al., 2023, Guesmi et al., 2023, garrone, 24 Nov 2025).

Several open problems recur. Synthetic modeling and simplified cognitive assumptions remain a limitation in modality studies; empirical estimation of parameters such as $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$8, $$S \in \{\text{brief}, \text{detailed}, \text{analogy}\}$$9, trust functions, and realistic information retention remains future work. In multi-agent settings, estimation of entropy rate and statistical complexity from finite data, principled choice of observables, and the translation of formal diagnostics into human-readable narratives remain unresolved. In legally informed settings, more domain-specific mappings between explanation artifacts and legal action are needed. Across all settings, there is tension between completeness and simplicity, transparency and cognitive load, explanation fidelity and practical usability, and seamlessness and the deliberate exposure of seams (Rajhans et al., 6 Feb 2026, garrone, 24 Nov 2025, Mansi et al., 14 Apr 2025, Ehsan et al., 2022).

Taken together, the literature supports a broad definition: explainable information design is the practice of structuring, selecting, and presenting explanatory information so that model behavior, system behavior, and sociotechnical constraints become understandable, usable, and contestable for specific audiences in specific contexts. This includes interface design, interaction design, information architecture, provenance engineering, modality choice, optimization under explainability constraints, and governance design. A plausible implication is that future work will increasingly treat explanation not as a single artifact but as an adaptive, multi-layer, multimodal information environment whose quality must be judged simultaneously by fidelity, comprehensibility, calibrated trust, actionability, and accountability (Dhar et al., 12 Aug 2025, Zhang et al., 3 Jul 2025, Huynh et al., 2022).