Search Results (101)

Reinforcement learning (RL) has emerged as a promising approach for autonomous surgical robotic subtasks. Recent advances include deep reinforcement learning (DRL), imitation learning (IL), and vision–language–action (VLA) models. However, current evidence remains fragmented across simulation benchmarks, task-specific demonstrations, and limited clinical studies. Existing reviews primarily focus on RL algorithms, while the broader pathway from algorithm development to clinically deployable surgical autonomy has not been comprehensively synthesised. This PRISMA 2020-guided systematic review examines RL, IL, safe RL, simulation-to-real (sim-to-real) transfer, foundation models, VLA systems, and regulatory readiness in surgical robotics. We searched IEEE Xplore, PubMed/MEDLINE, Embase, Scopus, Web of Science, the Cochrane Library, ACM Digital Library, arXiv, and medRxiv for studies published between January 2015 and March 2026, with additional studies identified through backward citation tracing. Eligible studies proposed novel RL, imitation learning, or foundation-model approaches for surgical robotics with empirical validation in simulation or on physical robotic platforms. Two reviewers independently extracted data using a predefined coding scheme, and a third reviewer resolved disagreements. Owing to substantial heterogeneity in platforms, tasks, and outcome measures, a quantitative meta-analysis was not feasible; therefore, the evidence was synthesised narratively using a comparative framework. A total of 220 studies met the inclusion criteria, covering eleven active surgical RL platforms, seven paired sim-to-real studies, emerging foundation-model architectures, and three FDA-cleared robotic systems exhibiting Level 3 autonomy. Available comparative studies suggest that hierarchical approaches can outperform flat policies in long-horizon tasks, while language-conditioned models demonstrated promising multi-step surgical capabilities. Seven paired simulation-to-real studies were identified, encompassing tissue retraction, guidewire navigation, and surgical cutting tasks. Sim-to-real performance gaps varied substantially by task and metric, with success-rate gaps ranging from −10 to 50 percentage points (negative values indicating better real-world than simulated performance), while paired mean spatial errors differed by at most 0.61 mm. Most studies employed domain randomization or visual domain adaptation; hierarchical reinforcement learning demonstrated advantages over flat policies in multi-step surgical tasks. Explicit safety-constrained methods (CPO, CBF, and SER), formal verification, and regulatory-aligned evaluation were reported in fewer than 3% of applied studies. Most evidence remained simulation-based, with no reported autonomous RL execution in vivo in humans. Overall, RL-based surgical robotics appears mature at the simulation stage but remains preclinical for autonomous clinical deployment. Future progress requires stronger sim-to-real validation, multimodal safety-aware architectures, alignment with IEC 62304, ISO 14971, FDA guidance, and the EU AI Act, and open benchmarks that jointly evaluate performance, safety, and surgeon trust. Full article

The EU AI Act makes fairness metrics for high-risk AI systems’ compliance evidence, turning their trustworthiness into a safety and accountability concern. Fairness audits assume that test data reflects the properties of real-world conditions, whereas standard evaluation protocols use test data drawn from the same biased records as the training data. Studies measuring how strongly this bias in test data distorts fairness metrics between the validation phase and real-world deployment are still very rare. We conduct an experiment on synthetic and real data, measuring this discrepancy across five fairness interventions on four unfairness types (1000 repetitions per combination, 20,000 total runs). Synthetic data lets us encode human bias in labels and compare fairness metrics on biased test labels (data available during development) against clean labels (conditions models face in deployment). We find that a systematic evaluation bias is present across all metrics, so the same models on the same test data can support opposite fairness conclusions and mask the mistreatment of the most disadvantaged groups. This pattern of fairness misevaluation is confirmed by a real-world validation on the Adult Census Income dataset. We conclude that trustworthy fairness auditing and regulatory standards should require bias-aware evaluation protocols, in which observed labels are not treated as ground truth. Full article

Low-emission zones (LEZs) require cities to define policy rules, predict effects before deployment, and verify outcomes afterwards, yet mobility, emissions, meteorology, exposure, data governance, and public communication are commonly handled in separate systems. This paper presents envair360, a Physical Intelligence architecture and a four-stage, evidence-gated LEZ methodology connecting project definition, baseline feasibility, digital-twin design, deployment, and verified impact closure. A design-science method is combined with an operational case study of Cartagena, Spain, because the research object is both a socio-technical artefact and a context-dependent municipal deployment. The technology chain is selected to bridge complementary scales and functions: SUMO for link- and vehicle-level traffic, WRF and CHIMERE for meteorology and regional chemistry, MUNICH and street-canyon parameterisation for computationally tractable street resolution, model-output calibration anchored to measurements, and FIWARE/NGSI-LD for governed context exchange. The manuscript distinguishes city observations, peer-reviewed component validation, demonstrated platform capabilities, and policy or engineering targets. A Murcia component study reports lower hourly than daily agreement after deep-learning calibration (NO₂: $r=0.79$ hourly and $0.94$ daily; O₃: $r=0.85$ hourly and $0.97$ daily), illustrating the importance of temporal aggregation and transfer limits. Digitisation of the prior Madrid ozone-density figure indicates modal shifts of approximately +32.0 and +27.8 source-axis units at two stations; the supplied source does not permit a numerical NO_x bias estimate. A separate six-city export audit covers 24,384 records and 4064 street segments and demonstrates a common model-output schema, not predictive validation. In Cartagena, project documentation reports elevated PM₁₀/PM_2.5, urban heat and solar-radiation stress, and a plausible role for dry-climate dust resuspension, supporting a superblock-oriented LEZ proposal with a long-term 30% vehicular CO₂ reduction target. The paper’s specific contribution is the governed orchestration, evidence taxonomy, quality gates, reproducible lineage, explicit policy-scenario representation, and portable city-onboarding protocol; it does not claim that the individual scientific models, the Cartagena deployment, or the cited project targets originated in this manuscript. Full article

The shift from passive predictive models to autonomous agents capable of tool use and multi-step planning moves the AI safety landscape from prediction error to control failure: small misjudgements become irreversible actions, and risks compound across long horizons and populations of interacting systems. We present a structured review and taxonomy of open scientific problems in agentic AI safety, mapped explicitly onto the EU AI Act and the NIST AI Risk Management Framework. The corpus follows a PRISMA-ScR scoping review, assembled through anchor-based citation chaining and curated reading lists across arXiv, the major machine-learning conferences, and selected security and fairness venues, with a primary March 2026 search cut-off (extended to May 2026 during revision for a small number of high-relevance governance and agentic-safety sources), explicit eligibility criteria, and an analytical distinction between open scientific problems and deployment risks. The taxonomy identifies eight problem families spanning reinforcement-learning policies and language-model planners: goal specification, inner alignment, safe learning and robustness, scalable oversight, interpretability, tool-use security, multi-agent safety, and evaluation and assurance. Mapping these onto the two frameworks shows close alignment for some families and notable absences for others, with multi-agent safety surfacing as a regulatory gap. We add a per-family research roadmap with concrete milestones and a practitioner-facing deployment-posture triage, arguing that progress on inner alignment, interpretability for deceptive-alignment detection, and multi-agent safety would most directly reduce compliance uncertainty. Full article

Deepfakes, synthetic audiovisual content produced by deep generative models, have escalated into a critical threat across civilian and military domains, enabling identity fraud, disinformation campaigns, and evidence fabrication. In high-stakes environments, ranging from journalism and finance to healthcare and legal contexts, the consequences extend to severe misinformation, market manipulation, identity fraud, and the erosion of institutional trust. This entry explores how modern visual intelligence and computer-vision techniques are used to detect deepfakes. It outlines key deepfake generation models, such as GANs, autoencoders, neural rendering, and diffusion systems, while also explaining how adversarial methods enhance realism and challenge existing detectors. The overview highlights visual artifacts, digital patterns, and physiological cues commonly leveraged in detection and reviews major CNN, transformer, and frequency-based approaches. It also summarizes evaluation practices and the difficulty of achieving strong generalization. Finally, it identifies emerging directions, including modern intelligence techniques for civilian and military content verification. This survey covers generation architectures (GANs, latent diffusion, neural rendering, video synthesis), the spatial, temporal, frequency-domain, and physiological artifacts they produce, and the detector families that exploit them. We examine evaluation benchmarks and protocols, highlighting cross-generator generalization as the field’s central open challenge. Beyond detection, we discuss cryptographic provenance standards, watermarking, and regulatory frameworks (EU AI Act, DSA, GDPR). We conclude that effective deepfake governance requires defense in depth integrating forensic detection, verifiable provenance, and institutional accountability. Full article

The emergence of Large Language Models (LLMs) has significantly advanced computational biology, yet their integration into autonomous, multi-agent systems (MASs) and clinical workflows remains challenging due to systemic architectural fragmentation. To quantify the operational readiness and regulatory compliance of bioinformatics LLMs, we developed the Multi-Agent Readiness Score (MARS), a standardized evaluation framework assessing models across four structural dimensions: Governance & Accessibility, Biological Competence, Technical Maturity, and Agentic Orchestration. The framework incorporates compliance criteria from the EU AI Act, HL7 FHIR, HL7 CDA, and MyHealth@EU standards. To empirically validate this domain-agnostic methodology, we applied it to a highly mature subset of the field: a diverse cohort of 43 prominent genomic LLMs. Our assessment revealed a severe, industry-wide readiness gap: the majority of models fell into “Not Suitable” or “Research Prototype” tiers, lacking essential technical interfaces, structured communication schemas, and provenance tracking. Furthermore, the data demonstrated a ’competence-readiness gap’, where models scale in biological predictive competence without corresponding improvements in engineering utility. The primary barrier to scalable bioinformatics AI is no longer biological competence, but operational and architectural incompatibility. By quantifying integration friction, MARS provides a crucial, reproducible metric to audit model maturity, guide system architecture, and ensure future models are structurally prepared for the rigorous regulatory demands of precision medicine workflows. Full article

Artificial intelligence (AI) is increasingly embedded in high-stakes socio-technical systems, intensifying concerns about autonomy, accountability, data rights, and fundamental-rights protection. This article develops an exploratory, expert-informed Human-Centred AI (HCAI) pre-design governance framework that translates selected risk-based obligations of the EU Artificial Intelligence Act into early organisational decisions about human oversight, data accountability, documentation, and bounded algorithmic autonomy. Using a sequential mixed-methods design, the study combines an Analytic Hierarchy Process (AHP) survey of 28 experts with think-aloud interviews with 15 of those respondents. The AHP results show that, among the governance criteria included in the model, AI design objectives received the highest upper-level priority and human oversight and control received the highest global priority, followed by personal information protection, design ethics, intellectual property rights protection, and limits of algorithmic autonomy. The interviews explain these priorities by showing that experts framed trustworthy AI governance as a problem of controllability, responsibility allocation, traceable data use, rights protection, and verifiable human intervention rather than model performance alone. The study contributes by defining pre-design governance as a bounded initial consideration-stage decision structure, combining AHP-based priority evidence with qualitative justification logic, and proposing a preliminary governance package of decision points, minimum evidence artefacts, and illustrative operational check criteria. The package is not presented as a validated legal compliance model; instead, it provides an expert-informed translation pathway for future organisational, sector-specific, and empirical validation. Full article

The rapid integration of generative artificial intelligence into digital advertising, combined with growing consumer concern about greenwashing, has created a dual credibility burden in which sustainability messages must be evaluated for both environmental truthfulness and authenticity of authorship—a challenge intensified by emerging AI disclosure regulations such as Article 52 of the EU AI Act. This study investigates how ad source (AI-generated vs. human-created), source label (labelled as AI vs. labelled as human), and environmental claim strength (vague vs. strong) jointly influence perceived greenwashing and green trust among Generation Z consumers. A 2 × 2 × 2 mixed factorial experiment was conducted with 154 undergraduate participants randomly assigned to one of four source–label conditions and exposed to both vague and strong environmental claims; data were analysed through one-way ANOVAs with Tukey HSD post hoc tests and paired-samples t-tests. All eight hypotheses were supported: AI attribution consistently elevated greenwashing perceptions and depressed green trust, while strong claims significantly reduced greenwashing perceptions and elevated trust within both correctly disclosed conditions. Most notably, the disclosed label—rather than the actual generative source—emerged as the dominant psychological cue, with human content mislabelled as AI suffering the same credibility penalty as genuine AI content. The findings reveal a transparency paradox with significant implications for AI disclosure regulation and sustainable marketing practice. Full article

Artificial intelligence (AI) tools—including audit data analytics, robotic process automation, machine-learning models, and generative AI—are changing how audit teams identify risks, select procedures, and evaluate evidence. At the same time, Regulation (EU) 2024/1689 (the EU AI Act) establishes a risk-based governance architecture built around risk management, data governance, technical documentation, logging, transparency, human oversight, robustness, cybersecurity, and post-market monitoring. The Act is not an auditing standard and does not directly regulate every tool used by audit firms. Nevertheless, its governance logic is relevant where audit firms develop, procure, or rely on AI-enabled systems that process sensitive client data, influence professional judgement, or become part of audit-relevant client systems. This conceptual study uses doctrinal requirements-to-controls mapping and design-oriented analysis to translate selected AI Act governance objectives into firm-level and engagement-level quality-management controls and into criteria for evaluating AI-enabled audit evidence. The paper specifies three modes of AI Act relevance: direct legal relevance where a regulated AI Act role is engaged; indirect relevance where AI compliance documentation becomes audit-relevant information; and benchmark relevance where the Act supplies governance objectives for quality management without creating an audit-law duty. The resulting artefacts are a traceable AI Act/IAASB standards crosswalk, an evidence-risk typology, a quality-management integration model, a documentation and review checklist, and a proportional maturity model. The framework clarifies when AI outputs remain triage or risk-assessment tools, when they provide directional or corroborative evidence, and the narrower conditions under which they may contribute to substantive evidence. It links reliance to data completeness, reconciliation, versioning, validation, false-positive and false-negative behaviour, explainability, logging, source-document corroboration, and reviewer challenge. The contribution is a scalable governance-to-controls framework that supports defensible reliance and inspection readiness without overstating the AI Act’s direct legal applicability. Empirical validation in audit firms remains a priority for future research. It further explains how quantitative risk features and anomaly-detection outputs feed into qualitative audit judgement: models can route attention to unusual transactions or documents, but evidential weight still depends on base-rate-aware error analysis, source-document corroboration, and reviewer challenge. Full article

AI-driven digitalisation transforms how organisations process personal data and introduces risks that traditional Data Protection Impact Assessment (DPIA) frameworks cannot adequately address. Automated decision-making and large-scale processing in AI, IoT, big data analytics, and blockchain environments create privacy concerns beyond the scope of existing DPIA methodologies. The EU AI Act extends this scope through the Fundamental Rights Impact Assessment (FRIA) under Article 27, which links data protection obligations to broader fundamental rights governance. This study addresses these gaps through a two-phase research design. Phase 1 conducts a systematic literature review of 25 studies and applies framework analysis to identify DPIA implementation challenges across four categories: legal and regulatory, risk assessment, scope, and complexity. AI-specific challenges appear across all four categories. Phase 2 develops a governance framework built on the Plan-Do-Check-Act (PDCA) cycle and organised through a four-level hierarchy of Lifecycle Phase, Risk Management Domain, Control Objective, and Operational Activity. The framework translates relevant requirements of ISO 31000:2018, ISO/IEC 27701:2025, and ISO/IEC 29134:2023 into traceable activities and encompasses algorithmic fairness and socio-ethical impacts. The actionable DPIA framework supports compliance with the GDPR, the EU AI Act and the three ISO standards and will be of interest to company practitioners and other researchers investigating the theoretical and practice-based aspects of digitalisation and data privacy. Full article

Artificial intelligence (AI)-driven fraud detection systems in FinTech ecosystems increasingly face a governance tension between high predictive accuracy and limited regulatory transparency, a gap that existing reviews have not addressed through an integrated behavioural, technical, and institutional lens. This study synthesises 99 Scopus-indexed, ABDC-ranked journal articles (2015–2026) using PRISMA 2020 and the SPAR-4-SLR protocol, integrating the Theory of Planned Behaviour (TPB) within an Antecedents–Decisions–Outcomes (ADO) framework to examine organisational adoption of explainable AI (XAI) in financial fraud detection. Three antecedent clusters are identified: attitudinal (algorithmic complexity, model opacity, data imbalance), normative (regulatory compliance, ethical expectations), and control-based (technical self-efficacy, organisational readiness)—which drive decision mechanisms including post hoc interpretability tools (SHapley Additive exPlanations [SHAP], Local Interpretable Model-Agnostic Explanations [LIME]), ethical governance protocols, and human-in-the-loop oversight. These produce outcomes across precision (reduced false positives, improved decision accuracy), compliance (audit transparency, institutional legitimacy), and cognitive (user acceptance, procedural justice) dimensions. The study introduces the Stability–Transparency–Reliability (STR) model, which advances TPB, Socio-Technical Systems Theory, and the Dynamic Capabilities View by reframing XAI from a static interpretability output into a recursive governance capability, formalised through the concept of Interpretative Agility, with direct implications for financial institutions operating under the EU AI Act. Full article

The operationalization of the One Health (OH) approach remains a major challenge due to persistent fragmentation across human, animal, and environmental data systems. This gap is exacerbated by climate change, which acts as a risk multiplier for pathogen transmission and agri-food system vulnerability. Drawing on more than a decade of research, including the re-emergence of brucellosis in Italy and the 2024 Salmonella Umbilo outbreak, this perspective discusses key weaknesses in current data management, particularly the lack of real-time, interoperable data sharing. To address these challenges, we propose an AI-enabled One Health Information System (OH-IS), grounded in FAIR data principles and privacy-preserving architectures. The proposed conceptual framework integrates multi-matrix data streams, combining Earth observation data, genomic surveillance through whole-genome sequencing (WGS), and livestock mobility within a geospatially integrated architecture to support timely decision-making in vulnerable settings. By analyzing the constraints of siloed databases, we discuss how automated semantic harmonization could conceptually support improved risk assessment and outbreak reconstruction in recent zoonotic events. This approach may facilitate a transition from descriptive to anticipatory surveillance, providing a scalable model to move One Health from a conceptual paradigm toward a more integrated and data-driven surveillance framework aligned with EU digital health policies and global health security priorities. Full article

Background and Objectives: As Romanian health systems deploy artificial intelligence (AI), uptake depends on navigating the EU AI Act, GDPR, the Medical Device Regulation (MDR), and national rules. We measured AI legislation literacy, governance readiness, and adoption intentions among Romanian healthcare professionals, identified implementation phenotypes, and tested whether confidence mediates the literacy–adoption link. Materials and Methods: In a multicenter cross-sectional survey (N = 109), participants completed a 20-item AI Legislation Literacy Index (0–20) plus scales rated form one to five measuring legislative confidence, adoption intention, readiness, trust, and perceived compliance burden. We used PCA and k-means clustering, multivariable logistic regression for high adoption intention (≥4), and covariate-adjusted mediation (5000 bootstrap resamples). Results: Mean age was 38.7 ± 9.8 years, and 60.6% of participants were female. Literacy was moderate (11.2 ± 4.1/20) and familiarity favored GDPR (69.7%) over the EU AI Act (25.7%). Literacy correlated with confidence (=0.52), whereas confidence correlated with adoption intention (=0.41); trust correlated positively (=0.44) and burden correlated negatively (=−0.29) with adoption. High adoption intention was noted in 50.5% of participants and was independently associated with higher literacy (aOR 1.85 per +1 SD; 95% CI 1.20–2.85), higher trust (aOR 1.72; 1.13–2.63), lower burden (aOR 0.64; 0.43–0.95), and prior AI training (aOR 2.10; 1.03–4.29). Three phenotypes emerged (Confident Adopters n = 44; Cautious Compliers n = 36; Skeptical Low Literacy n = 29), with adoption scores of 4.2 ± 0.5 vs. 3.1 ± 0.7 in the highest and lowest groups. Mediation showed a partial indirect effect via confidence (0.13; 95% CI 0.05–0.24). Conclusions: AI legislation literacy, confidence, trust, and perceived burden are key, modifiable determinants of AI adoption intentions; phenotype-guided strategies can target training, governance support, and post-deployment monitoring readiness. The revised framing explicitly situates these determinants within recent AI-specific regulatory and technical developments, including high-risk AI obligations, AI-enabled medical device change control, generative/large multimodal model risks, and lifecycle monitoring. Full article

Background: Artificial intelligence (AI) adoption is increasingly reported as a priority for organizations, yet they face a growing, fragmented body of evidence concerning the factors that influence successful AI integration. Method: To identify the relevant factors for organizational AI adoption, we conducted a systematic literature review (SLR) following PRISMA guidelines, which yielded 37 quantitative empirical studies. From these studies we extracted 1229 paper-item instances, of which 810 were retained after applying structured exclusion criteria to develop a domain model relevant to organizational AI adoption. The model’s content validity was assessed and supported through expert feedback using the Content Validity Index (CVI) methodology. Results: We organized 24 subclusters into nine main clusters across the three dimensions Technology (Enablers, Usability, Trust), Organization (Leadership, People, Process), and Environment (Market, Regulatory, Partner). Our analysis suggests that workforce skills, perceived intelligence, and resources are among the most frequently studied and positively associated antecedents of AI adoption, and that constructs related to AI explainability and control (human-in-the-loop oversight) have received little research attention and remain underrepresented despite growing regulatory requirements such as the EU AI Act. Conclusions: The resulting domain model provides an empirically grounded classification of organizational AI adoption factors and can serve as a foundation for future measurement instruments. Full article

Ensemble learning has been widely adopted in fraud detection; however, conventional ensemble strategies rely on uniform or performance-based weighting schemes that treat explainability as a post hoc annotation rather than an architectural component. This study addresses the research goal of whether SHAP attribution values can serve as a principled, instance-specific weighting mechanism within an ensemble, thereby embedding interpretability directly into the aggregation process. A SHAP-Weighted Ensemble (SWE) framework is proposed in which the L2 norm of each base model’s SHAP attribution vector, computed at prediction time, is used to derive instance-specific voting weights via Softmax normalization. Three linear base learners (logistic regression, robust LR, calibrated linear SVM) are combined, with LinearSHAP providing exact attribution values. A comprehensive evaluation protocol was applied on a real-world vehicle insurance claims dataset, including bootstrap 95% confidence intervals, McNemar’s test, a three-way ablation study comparing equal weighting, SWE, and validation-AUC weighting, F1-optimal threshold selection, expected calibration error, and cost-sensitive evaluation under asymmetric misclassification costs. The central finding is that SWE achieves performance statistically comparable to both simpler baselines across all evaluated metrics (ROC-AUC = 0.774, 95% CI [0.681, 0.862]; F1 = 0.679, 95% CI [0.569, 0.774]; McNemar p = 1.000), while producing a transparent, per-claim weighting trace that equal-weight voting cannot provide. A KernelSHAP influence analysis conducted directly on the SWE confirms that SHAP-derived weights are substantially aligned with actual model influence ratios (LR: 1.05×, LR_R: 1.05×, SVM: 0.81×), validating the weighting mechanism empirically. An exploratory analysis of a seven-model equal-weight diagnostic ensemble reveals a negative correlation (r = −0.721, p = 0.067) between individual model performance and ensemble influence; a theoretically coherent finding that does not reach statistical significance at conventional thresholds. The primary contribution of SWE is architectural and interpretability-driven: it produces an auditable, instance-level model-weighting mechanism grounded in SHAP attribution theory, supporting regulatory accountability under GDPR Article 22 and the EU AI Act. Full article