explainable ai

Staged Actions and Provable Human Oversight in Agentic Enterprise AI

Abstract. As enterprise AI moves from answering questions to taking actions, the question of human oversight sharpens. Regulation for higher-risk uses requires that systems be overseen by people who can understand, intervene in and if necessary halt them. In deployed practice this duty is often satisfied by assertion: a policy states that a human is in the loop, and the organisation is asked to trust that the loop was honoured. This paper argues that for regulated use an agent's privileged actions should be staged for human clearance, and, more particularly, that each clearance decision should itself be recorded, so that oversight leaves a trace rather than remaining an operational claim. We describe a stage, clear and seal-before-execution flow in which an agent that proposes a privileged action does not perform it directly: the action is placed in a staging state, presented to an authorised person for a decision, and executed only after that decision has been sealed into a tamper-evident record. Crucially the record captures the decision, who made it, when, and what was shown, not merely the action that followed. We relate this to the human-oversight duties of the EU AI Act and explain how sealing the decision rather than only the action turns oversight from an assurance into evidence. We are candid about the limits: a record shows that a decision was made and shown, not that it was sound, and staging does not by itself defeat inattentive approval.

Human-Centered Educational Leadership for Sustainable Development in the Age of Artificial Intelligence

This chapter develops a human-centered, leadership-driven framework explaining how education contributes to sustainability in AI-augmented societies. It addresses three research questions on leadership as a driver of sustainability, its mechanisms of influence, and its role in employability, inclusion, and resilience. Using a conceptual methodology and literature synthesis, it draws on Human Capital Theory, the Capability Approach, and leadership theories. Findings show that transformational, instructional, and adaptive leadership jointly shape education through ethical resource allocation, curriculum, capacity building, stakeholder engagement, and apprenticeship systems that bridge education and the labor market through experiential learning. These processes enhance capabilities, employability, and inclusion while reducing inequality risks in AI contexts. The chapter concludes that education drives sustainable development only when guided by integrated human-centered leadership, emphasizing leadership development, inclusive governance, and curriculum reform for future-ready systems

Decision delegation to GenAI agents in travel planning: Responsible AI signals, delegation levels, and the transparency paradox

Generative Artificial Intelligence (GenAI) is becoming an integral part of travel planning. This fundamental transformation is changing the definition of travel decision delegation, calling for fresh research into responsible AI in tourism. Drawing on agency theory, trust theory and decision delegation framework, this study conceptualises five responsible AI characteristics (reliability, fairness, trustworthiness, accountability, transparency) as evaluative signals for travellers to assess GenAI across three decision delegation levels, i.e., attribute set, choice set, and final decision. Using a multi-method approach, we analyse data collected from 421 travellers. Findings suggest that reliability and accountability are consistent drivers of delegation across all levels, whereas trustworthiness is a threshold condition that becomes significant at the attribute level. We also found that excessive transparency can lead to cognitive overload and reduce willingness to GenAI decision delegation. Our findings encourage future research into responsible AI development in understanding system complexity, algorithmic explainability and traveller delegation confidence.

White-Box Completeness for Artificial Intelligence

Abstract Current research on explainable and white-box artificial intelligence faces prominent issues: conceptual disarray, divergent perspectives, and a disconnect between theory and practice. The foremost priority is therefore to return to the field’s purest objectives and explore its most fundamental questions. To this end, this paper introduces the axiomatic criterion of white-box completeness (WBC). Specifically, an agent is white-box complete if and only if its behavior can be bidirectionally approximated by human interpretable and manipulable mathematical forms with bounded error. It is a formalization for realizing three ultimate objectives for trustworthy AI: discernible learned knowledge and behaviors, knowledge extraction from AI, and human knowledge injection. Therefore, the ultimate pursuit of explainable AI has been transformed from the vague goal of “making AI interpretable” into a precise mathematical problem. Keywords: White-box completeness; Interpretability; Explainable artificial intelligence; Trustworthy artificial intelligence; Knowledge extraction and injection.

AI Revolution: The Death of Business Administration and the Birth of Inalienable Leadership

Artificial intelligence is increasingly performing administrative, analytical, coordinating, monitoring, communication, and decision-support functions historically associated with managers and business administration. This transformation does not imply the disappearance of organizations, management, managers, or business schools. It marks the erosion of administration as an exclusively human domain and as the principal justification for managerial authority. When machines can analyze, plan, coordinate, evaluate, recommend, and execute, leadership must be grounded in something deeper than superior control of information and process. Existing leadership theories explain traits, skills, behaviors, relationships, ethics, adaptation, collective agency, technological mediation, and organizational effectiveness. The documented review nevertheless identifies a higher-order question that these traditions do not make their central organizing principle: who retains final authority and responsibility over purposes, institutions, and civilizational direction when machines increasingly participate in organizational and social decision-making? This conceptual, theoretical, normative, and integrative working paper introduces and timestamps Inalienable Leadership as the first metatheory of leadership specifically formulated for the age of artificial intelligence and robotics, advances it as the overarching metatheory of leadership in that age, and formally records it as the twelfth and culminating original contribution of the complete FILE contribution system. FILE: The Five Intelligences of Leadership Evolution remains the core management theory and diagnostic framework. Leadership = Intelligence + Character + Judgment, with Intelligence conceptualized through FILE, Character through SQL, and Judgment through LENS within the FILE Architecture. Inalienable Leadership integrates, grounds, governs, and gives civilizational direction to FILE and the complete FILE contribution system. The paper establishes a three-level hierarchy: Inalienable Leadership; the integrated FILE architecture (including the five meta-intelligences, the nested intelligences, the five emergent outcomes of FILE); and the first eleven original contributions and their associated structures. It preserves the specific roles of Human Irreducibility, HACK, HTTPS, LENS, MAIL, MLT, and RC while placing them under one civilizational principle: humanity may share work and intelligence with machines, but it cannot surrender its final authority and responsibility for the purposes, institutions, and direction of civilization. The future will not be determined by what machines can do, but by whether humanity remains capable of leading what it creates. Humans + Machines, under human leadership.

Artificial Intelligence in Sustainability Assurance: Accounting Challenges, Audit Risks and a Conceptual Framework for ESG Verification

Sustainability reporting is moving from voluntary narrative disclosure toward regulated, evidence-based and externally assured corporate reporting, creating an assurance problem that artificial intelligence (AI) is expected to help address. Because AI is embedded in accounting and audit workflows, its outputs increasingly shape how assurance evidence is located, tested and evaluated. This conceptual article develops an assurance-specific framework answering three questions: for which sustainability-assurance procedures AI creates analytical value, which risks arise when AI influences assurance work, and which decision rights and controls should govern that influence. Integrating assurance standards, accounting and auditing research, AI-governance frameworks and behavioural studies, it finds AI adds value in five domains—evidence extraction, criteria mapping, anomaly and greenwashing screening, external-data triangulation, and documentation support—but only under defined base rates, error costs and source traceability. It identifies the risks limiting reliance: data, source fidelity, explainability, bias, calibration, preparer gaming, and auditor overreliance. The Responsible AI-Assisted Sustainability Assurance Framework sets graded reliance ceilings, non-delegable decisions, calibrated decision gates, anti-gaming safeguards and ex-post metrics, permitting clerical assistance, analytical recommendation and constrained agentic execution while prohibiting autonomous decisions on materiality, evidence sufficiency and conclusions. Illustrated in Europe, it generalises through ISSA 5000 as a testable model.

The Most Important Human Skills for the Age of AI and the Future of Work: FILE — The Five Intelligences of Leadership Evolution

Artificial intelligence changes what organizations can do; human leadership must decide what should be done, what remains accountable, and what must never be surrendered to machines. This working paper presents FILE: The Five Intelligences of Leadership Evolution as a human-centered management theory and diagnostic framework for responsible leadership and decision-making in the age of AI. FILE argues that the most important human skills for the AI era and the future of work require the integrated exercise of five meta-intelligences: Augmented Intelligence, Emotional Intelligence, Cultural Intelligence, Political Intelligence, and Adaptive Intelligence. The paper defines the FILE formula, maps the framework onto the human hand, presents the 70 nested intelligences of FILE, explains five emergent outcomes, and develops RC: The Relational Commons as the shared human ecosystem that leadership must protect, cultivate, and renew. It also clarifies boundaries among adjacent intelligences, offers a diagnostic application to AI-supported employee performance evaluation, and states limitations and future research directions. Its purpose is to establish a public, citable, and timestamped record of FILE as a management theory of human leadership in the age of AI.

Balancing the Triad: Socio-Technical Misalignments Between User-Centered Design, Agile, and AI in Healthcare Software

The rapid integration of Artificial Intelligence (AI) into healthcare software introduces profound complexities when combined with Agile Software Development (ASD) and User-Centered Design (UCD). Instead of a cohesive triad, this landscape exhibits partial intersections and fragmented literature. This Systematic Literature Review (SLR) investigates how competing priorities within the ASD-UCD-AI triad create socio-technical misalignments, impacting user experience and clinical implementation. An analysis of 27 primary studies, backed by a rigorous quality assessment, systematically weights the corpus's evidentiary strength. Through a transparent synthesis pipeline based on the extraction of verbatim excerpts, generation of inductive codes, and thematic grouping, this study uncovers four core socio-technical tensions: (1) Technical Accuracy vs. Contextual Value; (2) Agile Delivery Speed vs. Clinical Safety Compliance; (3) AI Automation vs. Human Explainability; and (4) High-Level AI Ethics vs. Daily Agile Practices. While high-quality studies primarily defined these dimensions, lower-quality evidence provided contextual background. This review demonstrates that failing to balance these tensions relegates robust models to a "model graveyard." Finally, the study consolidates actionable mitigation strategies, such as redefining multidisciplinary Scrum roles and implementing continuous design controls via pull requests. By bridging theoretical ethical guidelines and practical software engineering, this SLR provides an evidence-informed analytical lens to develop safe, user-centric, and agile AI healthcare systems.

Non-extensive statistics of the glioblastoma transcriptome: analysis code, derived results and validation-cohort inputs

Analysis code, derived result tables, figures and validation-cohort inputs for a study of non-extensive (Tsallis) statistics in the glioblastoma transcriptome. Expression deviations of each sample are fitted with q-Gaussian laws by maximum likelihood. In a longitudinal GLASS cohort (230 libraries from 115 IDH-wildtype patients, at diagnosis and at recurrence after temozolomide chemoradiotherapy) the entropic index has a cohort mean of 1.330 with a standard error of 0.007, and reproduces in two further independent cohorts analysed identically: 1.331 in CGGA-693 (n = 109) and 1.372 in TCGA-GBM (391 libraries from 293 participants). What is new in version 4. Estimator validated against data with a known answer. On 350 synthetic samples drawn from q-Gaussians with known parameters at the observed gene counts, the estimator recovers q with a bias of -0.0008, no fit fails to converge, and the nominal 95% profile-likelihood intervals achieve 94.0% empirical coverage. Two further cohorts. Both CGGA mRNA-seq releases, stratified to match the GLASS design (primary, WHO IV, IDH-wildtype). Three cohorts agree to within 0.042, inside the a priori equivalence margin. A discrepant fourth cohort, with three explanations tested and excluded. CGGA-325 gives 1.420. Sequencing depth was tested by binomial thinning of real integer counts (shift -0.004), cohort size by subsampling (+0.003 at n = 74), and clinical composition by a pre-registered test of MGMT status (predicts 0.002 of 0.089, and the gap survives within both strata). Reported as unexplained. Model comparison against Laplace. The q-Gaussian is preferred by AIC in 87% of samples against an exponential-tailed alternative, with median dAIC above 120 in every cohort. A cohort-size effect in the estimator, characterised. Deviations from a cohort-median reference inflate the fitted tails when few samples are present (+0.032 at n = 20, +0.003 at n = 74); a fixed reference profile removes the effect, identifying the pathway. Applies to any index built the same way. Carried over from version 3. Estimation error measured by within-sample gene splitting (1–14% of the observed variance); a repeatability floor of 0.070 from 90 TCGA participants with replicate libraries, shown not to be explained by library composition; the finding that chemoradiotherapy moves the index no more than re-sampling the same tumour (F = 0.86, p = 0.78); equivalence testing with patient-clustered standard errors verified by a pairs-cluster bootstrap; and a circular covariate documented (the interquartile range of the deviation vector predicts q at R² = 0.94 but is a deterministic function of the fitted parameters). Contents. code/ analysis scripts in execution order; data/ open-access TCGA-GBM inputs with the GDC manifest; results/ one CSV per analysis; figures/ main and supplementary; legacy/ superseded artefacts, each with a file explaining why it is retired, including one retracted analysis. Reproducibility. The deposit is built from a git commit, recorded in the README. code/19_verify_manuscript_numbers.py recomputes every number printed in the manuscript from the tables in results/, confirms that each appears in the manuscript source, and exits non-zero if any disagrees. It passes on this build: 107 checks, 0 failures. Data availability. Raw RNA-seq and clinical data for the primary cohort are available from the GLASS consortium (Synapse syn17038081) under its terms of use and are not redistributed here; derived per-sample values remain subject to those terms, including a prohibition on commercialisation. See NOTICE.txt. TCGA-GBM data are open access from the NCI Genomic Data Commons and are included. CGGA data are openly available from cgga.org.cn and are not redistributed. Use of AI-assisted tools. Part of the analysis code was written with the assistance of a large language model (Claude, Anthropic), as described in the manuscript. All code was read, executed and verified by the authors, who take full responsibility for its correctness.

Explainable AI model for predicting customer service communication effectiveness: A computational design science approach

The NMC Theory for Global Bipolarity: A Structural Law of The Atomic Era

By using an axiomatic method, this manuscript identifies the first structural law ever formulated in International Relations. It introduces a new ontology of global order and shows that the post-1945 system obeys a geometric constraint that permits only two true Poles. Appendices 1-3 focus on method which is new and helpful for further studies in social science. Appendices 2-3 combine appendix in Logic of Closure give us The Minimal Test. V4 adds a direction for future researches. V5: rewrite appendix 4 The companion paper (Logic of Closure) introduces the first computational epistemology for real-time structural verification, and The Minimal Test for validating any law or theory in social science. https://doi.org/10.5281/zenodo.17911012 V6: adds a truly ugly appendix 7 to clear some possible misunderstanding about grayzone, boundary,... structurally. V7: adds only Appendix 8 illustrating discussion of deep governance layers (norms and institutions). V8: adds Appendix 9 which answers a question about Venezuela-Greenland sequence in NMC context. A deeper discussion may start from section 5.2. Also adds Disclaimer, Methodology. V9: add Clairification for Appendix 9. V10: delete the last line in Section 7 + add Appendix 10, Notes. V11: adds Appendix 11 - the last one. The manuscript is done! Appendices 7, 9, 10 and 11 (application layer) can be applied to read moves of Poles in future. V12: add Appendix 12 to complete Appendix 11. Two True Poles can be considered as systemic nodes which create tides/forces, etc. Appendices 4, 7-12 are however not ontology/core of The NMC Theory. They are mainly operational layers. V13: just add Appendix 13 V14: add note for Appendix 9, 11, 13 + edit 11.2, 11.12, 11.13 Note: For operational layer without NMC, we have a trilogy (weaker versions of 6.2, appendices 7, 9, 11-13) Geometry of Closure: https://doi.org/10.5281/zenodo.21107689 Strategic Stability and End of Absolute Freedom: https://doi.org/10.5281/zenodo.21411649 Maintaining the System: https://doi.org/10.5281/zenodo.21273268 V15: almost remove 3.7, rename 6.2, reorganize 6.3, add new 6.4 On The Logical Upper Boundary of NMC - on Aug 8 2026. Note. Application for SEA and non-poles generally (how to survive): https://doi.org/10.5281/zenodo.21938856 V16: minor edit last lines of section 7 and add some notes for A11-13 Structural Survival Doctrine for Japan https://doi.org/10.5281/zenodo.22001535 V17: remove "counter pole" in the def of True Pole, edit Methodology, Abstract, note C3 as the foundational law, and add note about Structural Time for A11. The third paper (ethics/law for AI) is "THE GEOMETRY OF STRUCTURAL ETHICS: Cognitive Symmetry as a Constitutional Invariant of the Post-Human Epoch" https://doi.org/10.5281/zenodo.17831278 DISCLAIMER NMC is a descriptive structural law, not a normative doctrine. It predicts systemic trajectories under post-1945 closure, but it does not justify, legitimize, or morally endorse any policy conducted in its name. Structural inevitability is ethically neutral; the ethical burden lies in how actors exploit inevitability. NMC may explain why “deals,” coercive compromises, and structural accommodations emerge, yet it does not recommend them, nor does it equate equilibrium with justice. Methodology: The Axiomatic Construction of NMC The NMC theory is not an inductive synthesis of historical narratives but a formally constructed axiomatic system. Engagement with NMC begins with its internal logic prior to empirical application. 1. Formal Validity and Internal Consistency The first-order validity of NMC is strictly internal. As an axiomatic system, it is evaluated along two criteria: Partition Exhaustiveness: whether its ontology (True Poles, Pseudo-Poles, Gray Zones) provides a complete and domain-bounded classification of all actors in the post-1945 nuclear-technological era; Axiomatic Consistency: whether its fundamental axioms (e.g., Axiom Zero, twele axioms, the Law of Structural Closure) are mutually consistent and non-contradictory. If these criteria are satisfied, NMC constitutes an autonomous formal-mathematical space. Its logical integrity stands independently of normative preference. 2. Empirical Fitness as a Second-Order Filter Empirical history and contemporary events function as a second-order filter assessing the fitness of this formal structure to its intended domain. Trajectory Validation: When observed systemic dynamics conform to trajectories induced by the axioms (e.g., patterns of pole competition, structural absorption of pseudo-poles), this constitutes empirical support. Perturbation vs. Violation: Surface-level deviations or short-term political resistance are considered local perturbations within the established geometry. Such perturbations do not, in themselves, invalidate the axioms unless they demonstrate a genuine structural violation. 3. Standards for Refutation An internally consistent axiomatic system can be refuted only on the following grounds: Formal Grounds: uncovering an internal logical inconsistency or showing that the partition is not exhaustive; Structural-Empirical Grounds: providing a genuine counterexample that breaks the structural topology — e.g., the sustained emergence of a bona fide Third Pole that independently satisfies all closure requirements. Critiques rooted in moral judgments, normative “oughts,” or conventional geopolitical metaphors (such as multipolar balance of power) are methodologically irrelevant unless they explicitly engage with the structural laws governing this closed system. Note. NMC is formulated as a structural law governing the post-1945 closed system. It does not claim exclusivity over all social phenomena; rather, it specifies the structural topology within which IR-level events unfold.

Searching for the truth? search engines and their AI affordances’ responses to conspiratorial search practices

Search engines are a common pathway to conspiracy theories, and producers of conspiratorial narratives actively encourage users to search for particular terms. Understanding how search engines respond to different information-seeking practices is therefore essential for explaining the dynamics of online conspiracy theories. Existing audits of search engines rarely account for systematic variation in query formulation arising from users’ differing levels of engagement with conspiratorial beliefs, nor do they consider recently introduced AI-powered search features. To address this gap, we developed search query sets informed by conspiratorial and non-conspiratorial information-seeking practices of online forums, using the chemtrails and 15-minute cities conspiracy theories as case studies. We performed algorithmic audits of Google's first-page search results and AI Overview responses. Our findings show substantial differences in the information returned for queries representative of general public information seeking compared with those reflecting conspiratorial perspectives. These differences appear to arise partly from Google's efforts to moderate sensitive topics, but also from the limitations of those interventions in recognising and responding to conspiratorial modes of query formulation. AI Overviews likewise vary according to query framing and generally attempt to debunk conspiratorial claims. However, these responses are often brief, provide limited supporting evidence, and are less effective for the newer 15-minute cities conspiracy theory, suggesting that current AI-mediated search guardrails remain uneven and require further development.