Paper deep dive
Addressing the Selection Problem in Explainable AI
Claire Vlases, Katelyn Morrison
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 8/30/2026, 2:27:15 AM
Summary
The paper formalizes the 'selection problem' in Explainable AI (XAI), where users struggle to map their natural language uncertainty to the appropriate XAI technique due to the siloed nature of existing interfaces. The authors propose a structural solution using a multi-agent Large Language Model (LLM) orchestrator that automatically translates user queries into specific XAI explanation techniques, thereby bridging the gap between user intent and system output.
Entities (12)
Relation Signals (9)
Claire Vlases → affiliatedwith → Carnegie Mellon University
confidence 95% · Claire Vlases 1 , Katelyn Morrison 1 1 Carnegie Mellon University
Katelyn Morrison → affiliatedwith → Carnegie Mellon University
confidence 95% · Katelyn Morrison 1 1 Carnegie Mellon University
Selection Problem → definedas → Failure to bridge gap between user uncertainty and technique
confidence 95% · the systematic failure of XAI interfaces to bridge the gap between a user's natural-language uncertainty and the explanation technique that resolves it.
Selection Problem → causedby → Siloed XAI Interfaces
confidence 92% · given the siloed nature of conventional XAI, users are struggling to select the appropriate XAI technique.
Multi-Agent LLM Orchestration → solves → Selection Problem
confidence 90% · We also propose a structural solution: a multi-agent LLM orchestration tool that translates the user's query to the proper XAI explanation technique.
XAI Question Bank (XAIQB) → usedtorepresent → User needs for explainability
confidence 88% · XAI Question Bank (XAIQB), an algorithm-informed XAI question bank in which user needs for explainability are represented as typical questions
Saliency Methods → identifies → Input regions driving prediction
confidence 85% · saliency methods identify which input regions drive a prediction
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Explainable AI (XAI) research has produced a plethora of explanation techniques, yet user studies repeatedly show that available explanations are not effective in practice. We argue that, given the siloed nature of conventional XAI, users are struggling to select the appropriate XAI technique. Viewing XAI through a philosophical lens, we offer a formalization of what we call the selection problem: the systematic failure of XAI interfaces to bridge the gap between a user's natural-language uncertainty and the explanation technique that resolves it. Following a logical premise-conclusion format, we show that conventional interfaces require users to translate their uncertainty into a technique selection, a challenging prerequisite to meet. We also propose a structural solution: a multi-agent LLM orchestration tool that translates the user's query to the proper XAI explanation technique. We provide an example of how this structural solution could be instantiated to address the selection problem.
Tags
Links
- Source: https://arxiv.org/abs/2608.22356v1
- Canonical: https://arxiv.org/abs/2608.22356v1
Trouble viewing inline? Open PDF directly →
Full Text
27,503 characters extracted from source content.
Expand or collapse full text
Addressing the Selection Problem in Explainable AI Claire Vlases 1 , Katelyn Morrison 1 1 Carnegie Mellon University, Pittsburgh, PA, USA cvlases, kcmorris@andrew.cmu.edu Abstract Explainable AI (XAI) research has produced a plethora of explanation techniques, yet user studies repeatedly show that available explanations are not effective in practice. We argue that, given the siloed nature of conventional XAI, users are struggling to select the appropriate XAI technique. Viewing XAI through a philosophical lens, we offer a formaliza- tion of what we call the selection problem: the sys- tematic failure of XAI interfaces to bridge the gap between a user’s natural-language uncertainty and the explanation technique that resolves it. Follow- ing a logical premise-conclusion format, we show that conventional interfaces require users to trans- late their uncertainty into a technique selection, a challenging prerequisite to meet. We also propose a structural solution: a multi-agent LLM orches- tration tool that translates the user’s query to the proper XAI explanation technique. We provide an example of how this structural solution could be in- stantiated to address the selection problem. 1 Introduction In recent years, the rapidly emerging field of explainable AI (XAI) has produced a diverse landscape of explanation techniques [ Mersha et al., 2024; Wilkinson et al., 2026; Schwalbe and Finzel, 2023; Schneider, 2024 ] . Many such explanations are unimodal and non-interactive: the user has limited control over what is shown and limited ability to re- quest follow-up information. Techniques are also often eval- uated in isolation and compared against one another to iden- tify the best configuration of a single method (e.g., [ Humer et al., 2022 ] ). Studies have explored a variety of XAI tech- niques. For example, saliency methods identify which input regions drive a prediction [ Selvaraju et al., 2019 ] . Counter- factual methods show how small input changes alter predic- tions and thereby reveal critical decision factors [ Mothilal et al., 2020 ] . Exemplar methods produce explanations by re- trieving the closest examples from a training dataset [ Hou and Wang, 2025 ] . Query-based methods let users question model perception directly [ Lakkaraju et al., 2026 ] . Each technique answers a structurally different question about AI behavior. However, the evaluation of these XAI techniques in prac- tice with real users remains limited [ Suh et al., 2025 ] . User studies show that available explanations are frequently ig- nored, misinterpreted, or used to confirm rather than scruti- nize AI outputs, leading to an inappropriate reliance on AI explanations [ Spitzer et al., 2025 ] . Some literature has at- tributed these challenges to the lack of human-centered tech- niques [ Mangold et al., 2025 ] . Recent work has begun exploring conversational and inter- active forms of explanation, including narrative-driven XAI with large language models [ Martens et al., 2026 ] and pro- posals for agentic interpretability pursued through dialogue with a generative model [ Kim et al., 2025 ] . These direc- tions share our view that explanations are better treated as an interaction than as a static output [ Rohlfing et al., 2020 ] . Our contribution formalizes the challenge of operationalizing XAI techniques within an interactive system. Before receiving any explanation when a user is collabo- rating with an AI system, that user must answer an implicit question: which technique is appropriate for what I currently want to understand? Conventional XAI systems offer expla- nation techniques as discrete controls. However, providing individual explanation methods themselves is not sufficient for users to judge, trust, or question an AI effectively [ Labarta et al., 2024 ] . Using a philosophic research lens, we formalize this as the selection problem in XAI: the failure of XAI in- terfaces to bridge the gap between a user’s natural-language expression of uncertainty and the explanation technique that resolves it. We formalize the selection problem as a property of conventional XAI interface design that places the burden of mapping uncertainty to the proper explanation technique on the user when it could be handled by the system. We then derive three requirements for a structural solu- tion: it must accept natural-language input directly; output adequate explanation techniques; and perform the translation mapping on behalf of the user. Given that a user expresses uncertainty in natural language, we propose conversational multi-agent orchestration of XAI as a possible design solu- tion. 2 The Selection Problem in XAI We formalize the selection problem as a structural property of conventional XAI interface design. The argument proceeds in arXiv:2608.22356v1 [cs.AI] 23 Aug 2026 three stages: different epistemic states require different tech- niques; users must perform a translation from their natural language uncertainty to an XAI technique selection; and the user’s mapping may be unreliable. Premise 1a. An AI model is a function from inputs to outputs. Premise 1b. An explanation technique is a function that takes an input and a model and returns an explanation. Premise 1c. Different explanation techniques may answer structurally different questions. Premise 1d. A user’s uncertainty has a specific epistemic state at any given moment. Conclusion 1. Given a user’s uncertainty, only a subset of available techniques can actually resolve it. Premise 2a. Users express uncertainty in natural language, not as technique selections. Premise 2b. Conventional XAI interfaces require a tech- nique selection as input. Conclusion 2. The user must perform a translation from nat- ural language to technique selection. Premise 3. Reliable question-to-XAI type mapping is not a skill the user population can be assumed to have. Conclusion 3. User-performed mapping is unreliable. Conclusion. Different uncertainties require different tech- niques (C1), and users cannot reliably translate their uncertainty into the right technique (C2, C3). There- fore, there exists a selection problem in which the user’s selected technique falls outside the adequate set. Premise 1a. An AI model is a function from inputs to out- puts. To view the selection problem through a philosophi- cal lens, we first define the structure of an AI model as a func- tion of inputs and outputs. An AI model M takes in an input space X , the set of all possible inputs, and maps to an output space Y , which contains all possible predictions generated by the system. The model is therefore a function M : X → Y . Under this formalization, M takes an inputx∈ X and returns a prediction y ∈ Y . Premise 1b. An explanation technique is a function that takes an input and a model and returns an explanation. With that understanding, we use the following definition of XAI: “a set of processes and methods that allows human users to comprehend and trust the results and output created by machine learning algorithms” [ IBM, 2026 ] . An explana- tion is the specific output produced by the technique when run on a particular input and model. For example, gradient class activation mapping (Grad-CAM) outputs a saliency map [ Sel- varaju et al., 2019 ] ; the Diverse Counterfactual Explanations (DiCE) algorithm outputs a counterfactual [ Mothilal et al., 2020 ] ; nearest-neighbors retrieval outputs an example-based explanation [ Dong et al., 2025 ] . Premise 1c.Different explanation techniques answer structurally different questions. Each technique produces a different kind of explanation output [ Liao et al., 2020 ] . We define a finite set of techniques as E, where any individual technique e i is an element of the set, e i ∈ E. Each technique e i takes in input x and model M to produce an explanation. Premise 1d. A user’s uncertainty has a specific epistemic state at any given moment. A user seeks an XAI expla- nation when they encounter an AI prediction that leaves them uncertain about the model’s logic; they need help understand- ing why M produced prediction y for input x. We represent the specific uncertainty as their epistemic state ω. Because the epistemic state is conditioned upon the existing knowl- edge the user holds at the time of the interaction, we write ω = (x,y,φ): the input to the model x, the output of the model y, and φ, the specific gap in their understanding that prompts them to seek an explanation. The uncertainty gap φ encompasses what the user does not understand but also why they seek to understand it. The goal of an explanation shapes which technique is adequate: a user seeking recourse is served by counterfactuals that modify a small number of features, while a user seeking to learn about a domain may be better served by more global explanations [ Delaney et al., 2022 ] . Two users with the same input, output, and surface question may have different reasons for wanting an explanation, thus possessing different epistemic states. Conclusion 1. Only a subset of available techniques can resolve any given epistemic state. Because each explain- ability method is designed to generate a distinct type of out- put, its capacity to address the user’s particular inquiry is nec- essarily constrained. The utility of any given technique is siloed, relevant only to a narrow set of questions. Within the context of an epistemic state ω, only some of the techniques in E are actually useful. We denote the subset of techniques capable of resolving the user’s uncertainty as E ∗ (ω), where E ∗ (ω) ⊆ E. This is the subset of explanations that, given the user’s epistemic state ω, would actually help the user with what they want to know. The efficacy of an explanation tech- nique is defined relative to the user’s epistemic state. Premise 2a. Users express uncertainty in natural lan- guage, not as technique selections. The epistemic state of the user ω is an internal condition that must be communicated externally to the XAI system. The user must convert their sense of ambiguity into a structured natural-language form, denoted as query q. The query q serves as the only exter- nal evidence of the user’s requirements. The user’s underly- ing epistemic state is not directly accessible by the system, which can only perceive a representation of the uncertainty. The user articulates their internal uncertainty φ into a query q ∈ L, where L represents the set of all possible expressions a user might say or think. The query q functions as a verbal- ization of the underlying uncertainty φ. Notably, q is not assumed to be an accurate or complete representation of the underlying uncertainty φ. A user’s se- lection of an inadequate technique can itself signal unre- solved uncertainty, but conventional interfaces have no chan- nel through which to detect or act on that signal. Premise 2b. Conventional XAI interfaces require a tech- nique selection as input. Conventional XAI interfaces do not accept q as input. In conventional systems, either the user selects the technique or the developer fixes the choice in ad- vance on their behalf. Either way, a human agent (which we will refer to as the user, for convenience) must produce some chosen technique ˆe∈ E. Conclusion 2.The user must translate from natural language to technique selection. The user must identify which methodology within the available set is capable of re- solving their particular question, mapping their query q to a technique choice ˆe. For this translation step, the user must possess enough technical knowledge to understand the capa- bilities of the various explanation techniques. Because both the articulation and the mapping phases occur on the user side of the interaction, the system is unable to detect if a failure oc- curs in either step. The system never sees the user’s internal uncertainty, only their selected technique. The system can instead only access q, a representation of φ. A mismatch be- tween the user’s curiosity and the system’s output can persist without any detection by the model. Premise 3. Reliable question-type classification is not a skill the user population can be assumed to have. One of the most referenced works on XAI is the XAI Question Bank (XAIQB), an algorithm-informed XAI question bank in which user needs for explainability are represented as typi- cal questions users might ask about the AI [ Liao et al., 2020 ] . Users “may hold preconceptions of what constitutes useful ex- planations for decisions,” even without a deep technical un- derstanding of AI [ Liao et al., 2020 ] . The XAIQB exists pre- cisely because users cannot be expected to produce a catego- rization or mapping of XAI techniques on their own. Empir- ical applications of the XAIQB corroborate that some ques- tions were difficult to distinguish and required interpretation during use, thus unclear to what extent the XAIQB captured the explanation needs of end-users [ Sipos et al., 2023 ] . In practice, a user-performed mapping of an uncertainty to a explanation technique produces structured and recurring failure modes. A typology of XAI failures identifies user-side failures arising from inconsistent inferences, distinguishing mismatch failures (contradictions between ML explanations and users’ expectations in terms of format), counterintuitive failures (contradictions in explanation content), and biased inferences (cognitive biases interfering with the explanation) [ Bove et al., 2025 ] . Each of these failure modes is consistent with a user who has attempted to seek an explanation without a reliable mapping from their question to a technique type. The user receives an explanation, but it does not align with what they were trying to understand. Conclusion 3. User-performed mapping is unreliable. It follows that for many users, their selection of an explanation technique ˆe is not reliably constrained to fall within the ade- quate set E ∗ (ω). The user is forced to map without the pre- requisites needed to perform that mapping reliably. This is a structural consequence of placing the mapping function f user at a point in the workflow where the user is required to act before the system can support them. Conclusion: There exists a selection problem in which the user’s selected technique falls outside the adequate set. We can write the user’s translation as a function f user : L → E that turns a natural-language query into a technique se- lection. The user is forced to perform this mapping with- out any support from the system, and doing it correctly re- quires knowing what each tool does, which question types each tool answers, and being able to accurately name what kind of question they are asking, all before seeing any expla- nation at all. This prerequisite exists on both the user and developer sides. While developers may understand the op- erational parameters of the available explanation techniques, they may lack insight into the user’s specific intent or the nature of their underlying question. End users may recog- nize their own uncertainty but lack the technical knowledge to identify the explanation tools that would resolve it. When the user fails to map their uncertainty to the appro- priate technique, and the selected explanation ˆe lands out- side of E ∗ (ω), we have what we call the selection problem of XAI: the technique the user chose is not among the tech- niques that would actually resolve their uncertainty. More formally, the selection problem occurs when ˆe /∈ E ∗ (ω).(1) The selection problem represents a failure in the explanatory process, where the output provided by the system remains functionally irrelevant to the user’s actual needs. 3 Structural Solution: Multi-Agent LLM Orchestration System The selection problem arises because conventional XAI inter- faces require the user to perform the mapping f user : L → E from query to explanation technique. The structural solution is to shift this burden to the system, a mapping function f sys . We identify three requirements for any system that replaces user-performed mapping: it must accept natural-language in- put, it must return adequate techniques as output, and the translation between them must be performed by the system rather than the user. The system accepts natural-language input. The user’s uncertainty originates in natural language L. Conventional interfaces demand a discrete technique selection, or they force a selection by only providing a single technique. The mismatch between these two forces the user to translate their query into a vocabulary the interface accepts. A structural so- lution must therefore accept q ∈ L directly, so the user is not required to map their question to technique-based terms. The system returns adequate technique routing. Each explanation technique answers a distinct question, and there is an adequate technique set E ∗ (ω) ⊆ E for each epistemic state. A solution that returned techniques outside E ∗ (ω) would deliver explanations that fail to address the user’s ac- tual uncertainty. The system must therefore aim to select techniques that can resolve the user’s expressed uncertainty. The system assists with the mapping translation. One solution could move the mapping function off the user par- tially or entirely, so that f sys : L → E ∗ (ω) is performed by the system. The user’s remaining contribution is to articulate q. Although the user’s articulation itself may not be trivial, shifting the mapping to the system opens the possibility of a collaborative, multi-turn exchange. The system could re- spond to internal uncertainty in ways a conventional interface cannot, such as eliciting the user’s goal or refining q before routing. 4 Instantiating a Multi-Agent LLM Orchestrator for XAI We provide the community with an example of how to in- stantiate this structural solution using four different XAI tech- niques in the context of medical imaging. While we believe this structural solution can be domain-agnostic, describing its potential operationalization in a specific domain helps illus- trate the functionality. Figure 1: (Top & Middle) Examples of existing public LLM in- terfaces with multi-agent functionality used for design inspiration. (Bottom) Example design of offering an LLM orchestrator “Smart” mode that routes natural-language queries from the user to one of the specialized XAI agents through function calling. The first system requirement demands natural-language understanding at the input boundary. The second requires knowledge of which techniques resolve which kinds of un- certainty, encoded in a way that supports automated routing. The third calls for a system-run process that routes a natural- language query input to an explanation technique as output. Drawing from designs in popular public LLM tools (top two images of Figure 1), such as Claude, Gemini, and Chat- GPT, we imagine how a multi-agent LLM orchestrator could help users navigate different XAI techniques as shown with the “Smart” mode in the bottom image of Figure 1. A “Smart” orchestration tool offers a blueprint: prompt an LLM to match the user’s natural-language question to the question types that known XAI techniques address, such as those in Liao’s XAIQB. Routing decisions are made by the orchestrator at the moment of query, ensuring that the techniques returned are matched to the structure of the user’s uncertainty. The orchestrator tool would thus model f sys : L→ E ∗ (ω), receiving the user’s natural language query, classifying its un- derlying type, and routing it to the best-suited explanation agent. For users who already know which explanation tech- nique they want, individual ”modes” can be invoked directly, bypassing the orchestrator. In our instantiation, the orchestra- tor itself is implemented with Gemini 2.5 Flash, configured to map query types onto the available agents. Our LLM-based orchestrator instantiation aims to address the three requirements by mapping the user’s query to a suite of XAI agents. Natural-language understanding is a native LLM capability, satisfying the first requirement. Function calling provides the routing mechanism, satisfying the sec- ond. The ability of the orchestrator to interpret a query, select techniques, and return outputs in a single pipeline absorbs the mapping function into the system, satisfying the third. 5 Future Work & Conclusion We have argued that the adoption gap in XAI is, at least in part, a problem of interface presentation. Conventional XAI interfaces force the user to map their natural-language uncer- tainty onto a discrete technique selection before any explana- tion is produced, and performing this mapping can be unreli- able. We formalized this as the selection problem and pro- posed an LLM-based multi-agent orchestrator that absorbs the mapping function into the system as one solution. Future work should consider the selection problem more closely by further investigating the premises and conclusions we present. For example, researchers could design studies to more concretely identify whether user-performed mapping is indeed unreliable and to what extent the selection problem exists for end-users versus AI system designers. Beyond that, researchers could expand these designs to other domains and XAI techniques. An immediate next step could be to compare the effectiveness of a multi-agent LLM orchestrator for XAI with that of presenting the techniques to users without the orchestrator. Technical researchers could explore how to best implement the backend of the orches- trator to ensure that the user’s natural language inquiries are routed to the correct XAI agent, whether through prompt en- gineering or by implementing multi-turn conversations. Ultimately, we hope that our formalization, description of system requirements, and example implementation of the sys- tem inspire conversations in the human-centered XAI com- munity and encourage researchers to consider the selection problem when designing XAI systems. Acknowledgments Use of Generative AI Disclosure. Claude was used during manuscript preparation to explore and refine candidate for- malizations, notation, and descriptions. The authors selected, revised, extended, and verified the final formalization and all claims/positions in the paper. References [ Bove et al., 2025 ] Clara Bove, Thibault Laugel, Marie- Jeanne Lesot, Charles Tijus, and Marcin Detyniecki. Why do explanations fail? a typology and discussion on failures in xai, 2025. [ Delaney et al., 2022 ] Eoin Delaney, Arjun Pakrashi, Derek Greene, and Mark T. Keane. Counterfactual explanations for misclassified images: How human and machine expla- nations differ, 2022. [ Dong et al., 2025 ] GenghuaDong,HenrikBostr ̈ om, Michalis Vazirgiannis, and Roman Bresson. Obtaining example-based explanations from deep neural networks, 2025. [ Hou and Wang, 2025 ] J. Hou and L. L. Wang. Explainable AI for clinical outcome prediction: A survey of clinician perceptions and preferences. In AMIA Joint Summits on Translational Science Proceedings, volume 2025, pages 215–224, June 2025. [ Humer et al., 2022 ] Christina Humer, Andreas Hinterreiter, Benedikt Leichtmann, Martina Mara, and Marc Streit. Comparing effects of attribution-based, example-based, and feature-based explanation methods on ai-assisted decision-making. OSF Preprints, 2, 2022. [ IBM, 2026 ] IBM. What is explainable AI?, 2026. [ Kim et al., 2025 ] Been Kim, John Hewitt, Neel Nanda, Noah Fiedel, and Oyvind Tafjord. Because we have llms, we can and should pursue agentic interpretability, 2025. [ Labarta et al., 2024 ] Tobias Labarta, Elizaveta Kulicheva, Ronja Froelian, Christian Geißler, Xenia Melman, and Ju- lian von Klitzing. Study on the helpfulness of explain- able artificial intelligence. In Explainable Artificial In- telligence, pages 294–312. Springer Nature Switzerland, 2024. [ Lakkaraju et al., 2026 ] Kausik Lakkaraju, Siva Likitha Val- luru, and Biplav Srivastava. Holistic explainable ai (h-xai): Extending transparency beyond developers in ai-driven de- cision making, 2026. [ Liao et al., 2020 ] Q. Vera Liao, Daniel Gruen, and Sarah Miller. Questioning the ai: Informing design practices for explainable ai user experiences. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, Honolulu, HI, USA, April 2020. ACM. [ Mangold et al., 2025 ] Aline Mangold, Juliane Zietz, Su- sanne Weinhold, and Sebastian Pannasch.On the de- sign and evaluation of human-centered explainable AI sys- tems: A systematic review and taxonomy [preprint]. arXiv preprint arXiv:2510.12201, October 2025. [ Martens et al., 2026 ] David Martens, James Hinns, Camille Dams, Mark Vergouwen, and Theodoros Evgeniou. Tell me a story! narrative-driven xai with large language mod- els, 2026. [ Mersha et al., 2024 ] Melkamu Mersha, Khang Lam, Joseph Wood, Ali AlShami, and Jugal Kalita. Explainable artifi- cial intelligence: A survey of needs, techniques, applica- tions, and future direction. Neurocomputing, 599:128111, 2024. [ Mothilal et al., 2020 ] RamaravindK.Mothilal,Amit Sharma, and Chenhao Tan. Explaining machine learning classifiers through diverse counterfactual explanations. In Proceedings of the 2020 Conference on Fairness, Ac- countability, and Transparency, FAT* ’20, page 607–617. ACM, January 2020. [ Rohlfing et al., 2020 ] Katharina J Rohlfing, Philipp Cimi- ano, Ingrid Scharlau, Tobias Matzner, Heike M Buhl, Hen- drik Buschmeier, Elena Esposito, Angela Grimminger, Barbara Hammer, Reinhold H ̈ ab-Umbach, et al. Explana- tion as a social practice: Toward a conceptual framework for the social design of ai systems. IEEE Transactions on Cognitive and Developmental Systems, 13(3):717–728, 2020. [ Schneider, 2024 ] Johannes Schneider. Explainable gener- ative AI (GenXAI): A survey, conceptualization, and re- search agenda.Artificial Intelligence Review, 57:289, 2024. [ Schwalbe and Finzel, 2023 ] Gesina Schwalbe and Bettina Finzel. A comprehensive taxonomy for explainable artifi- cial intelligence: a systematic survey of surveys on meth- ods and concepts. Data Mining and Knowledge Discovery, 38(5):3043–3101, 2023. [ Selvaraju et al., 2019 ] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In- ternational Journal of Computer Vision, 128(2):336–359, October 2019. [ Sipos et al., 2023 ] Lars Sipos, Ulrike Sch ̈ afer, Katrin Glinka, and Claudia Muller-Birn. Identifying explanation needs of end-users: Applying and extending the XAI ques- tion bank. In Mensch und Computer 2023, page 492–497. ACM, September 2023. [ Spitzer et al., 2025 ] Philipp Spitzer, Katelyn Morrison, Vi- olet Turri, Michelle Feng, Adam Perer, and Niklas K ̈ uhl. Imperfections of XAI: Phenomena influencing AI-assisted decision-making. ACM Transactions on Interactive Intel- ligent Systems, 15(3):17, September 2025. [ Suh et al., 2025 ] Ashley Suh, Isabelle Hurley, Nora Smith, and Ho Chit Siu. Fewer than 1% of explainable AI pa- pers validate explainability with humans. In Extended Ab- stracts of the CHI Conference on Human Factors in Com- puting Systems. ACM, 2025. arXiv:2503.16507. [ Wilkinson et al., 2026 ] Connor Wilkinson, John Yawney, and Stephen Andrew Gadsden. Explaining explainabil- ity: A comprehensive survey on explainable artificial in- telligence and relevant industry applications. Intelligent Systems with Applications, 30:200647, 2026.