AI, Volume 7, Issue 3 (March 2026) – 36 articles

Background: Deepfakes pose a growing threat to the integrity of visual media, motivating detectors that remain reliable as forgeries become increasingly realistic. Methods: We propose a deepfake detection framework built on CLIP-derived SigLIP-2 vision transformers and a multi-task design that jointly performs (i) classification and (ii) manipulated-region localization when pixel-level supervision is available. We evaluated the approach on three public benchmarks of increasing complexity—HiDF, SID_Set (SIDA), and CiFake—using each dataset’s official partitions where provided (SID_Set uses the predefined train/validation split) and a standardized preprocessing and training pipeline across experiments. Results: On HiDF, our model achieved strong performance on both video and image tracks (AUC up to 0.931 on video and 0.968 on images), yielding large gains relative to previously reported HiDF baselines under their published settings. On SID_Set, the model achieved 99.1% three-class accuracy (real/synthetic/tampered) and produced accurate localization masks for many tampered regions, while we explicitly documented the split protocol and leakage checks to support the validity of the evaluation. On CiFake, the model exceeded 95% accuracy and attained an AUC of 0.986. Conclusions: Overall, the results indicate that SigLIP-2 representations combined with multi-task training can deliver high detection accuracy and interpretable localization on challenging, realistic forgeries, while highlighting the importance of clearly stated evaluation protocols for fair comparison. Full article

We propose Trust Triangle, a Bridging Methodology that establishes evidential reliability through multi-attribution consensus, ensures external validity via statistical hypothesis testing, and enables controlled generation with RAG-anchored LLMs to transform black-box predictions into trustworthy, auditable explanations. This framework is instantiated for credit card fraud detection by integrating multi-method feature attributions with rigorous statistical validation. The resulting reliability-validity-verified insights are synthesized with high-relevance domain knowledge (relevance score > 0.7) retrieved from a real-world corpus via Retrieval-Augmented Generation (RAG). A structured Chain-of-Thought (CoT) prompt then guides an LLM to produce coherent, audit-ready case reports. Our contributions are threefold: (1) a verifiable framework for quantifying attribution reliability and validity, (2) a demonstrated end-to-end pipeline from robust prediction to semantically grounded explanation, and (3) a generalizable paradigm for Trustworthy ML in high-stakes domains. Experiments on a highly imbalanced dataset (fraud rate: 8.74%) demonstrate robust performance (PR-AUC = 0.7867), successfully identify statistically significant predictive features, and generate audit-ready reports, thereby advancing a rigorous, evidence-based pathway from model output to decision-ready support. Full article

Background: Laryngeal lesions represent a significant clinical challenge due to the complexity of the laryngeal structure, making manual diagnosis time-consuming and prone to subjective errors. Therefore, developing an accurate and lightweight automatic detection method is essential for improving the efficiency of laryngeal disease screening and diagnosis. Methods: This study proposes MSBA-YOLO, a lightweight laryngeal disease detection algorithm based on an improved YOLOv5s architecture. The method integrates FasterNet as the backbone network to reduce computational redundancy through partial convolutions and incorporates a Single-Head Self-Attention mechanism to capture long-range dependencies in complex lesion features. In addition, an MSBA-FIoU loss function is introduced to enhance the localization accuracy of multi-scale targets. Results: Experimental results show that MSBA-YOLO achieves a mean Average Precision (mAP) of 96.1% with a model size of only 6.4 MB, representing a 54.6% reduction in parameters compared with the baseline model. When deployed on the Jetson Orin Nano edge platform, the proposed method achieves real-time inference with a speed exceeding 50 FPS while maintaining low power consumption of 5.82 W. Conclusions: The results demonstrate that MSBA-YOLO effectively balances detection accuracy and computational efficiency, providing a robust and practical solution for portable and real-time clinical screening of laryngeal diseases on edge devices. Full article

This paper reports on a pilot study evaluating five large language models (ChatGPT-4, Gemini 2.0 Flash, Claude 3.5 Sonnet, Perplexity AI, and DeepSeek) in gradient acceptability judgment tasks in German. The models rated 150 contextually embedded sentences on a 5-point Likert scale across five categories: gray-zone (variable) items, canonical grammatical items, ungrammatical items, diatopically marked items, and diastratically/diaphasically marked items. All models clearly distinguish between clearly grammatical and clearly ungrammatical stimuli in unambiguous morphosyntactic contexts. Mixed-effects analyses further show that differences between models vary across stimulus categories rather than reflecting a uniform global shift in acceptability ratings. These findings indicate that current LLMs robustly capture core morphosyntactic contrasts, but that model behavior is less uniform in domains involving variation and contextual sensitivity. The study contributes to the empirical assessment of LLMs as acceptability raters and informs debates on their methodological role in linguistics. Full article

Background: Automated monitoring of crop health and the precise detection of abiotic stress, such as herbicide damage, are demanding challenges for modern agriculture. Abiotic stresses are a demanding challenge for modern agriculture, responsible for up to 82% of yield losses in major food crops. To address this, researchers are increasingly leveraging artificial intelligence (AI) to automate the detection and management of these stressors. Methods: In particular, this paper presents an instance segmentation framework to precisely detect interveinal chlorosis and leaf curling on potato leaves, two common symptoms of herbicide damage and soft wind. Within the context of precision agriculture and the need to address the inherent ambiguity in manual leaf assessment, this study employs a partial label learning approach to refine the dataset. This method utilizes an EfficientNet-b1 model to classify ambiguous samples, generating high-confidence pseudo-labels for instances that are difficult to categorize visually. The core of the proposed framework is a Mask2Former model, which is first fine-tuned on general potato leaf dataset to enhance its segmentation capabilities and then transferred on the refined, pseudo-labeled dataset. Results & Conclusions: This two-stage approach yields a highly accurate segmentation tool, achieving 89% mAP₅₀ and a pseudo-label classification accuracy of 95%, designed for integration into smart agriculture systems like ground level robotics or unmanned aerial vehicles for real-time, automated crop monitoring. Full article

How closely do the subjective perceptions simulated by Artificial Intelligence align with the subjective perceptions of human participants when evaluating an urban environment? This study serves as a pilot investigation to explore how far multimodal Large Language Models can effectively model human responses to visual stimuli based on subjective criteria. The exploratory nature of this research intends to test the feasibility of the methodology rather than provide a definitive standard. By focusing on a small set of detailed audits, a small-scale experiment performs an in-depth, qualitative examination of how machines and human assessments compare to each other in specific situations. To conduct the comparison, ratings of urban scenes were collected from human participants and two multimodal Large Language Models: ChatGPT and Gemini. After showing them an image of a sidewalk, these appraisers used a set of proposed statements to rate three sidewalks on a Likert scale. The investigation focuses on seven statements that subjectively characterize walkability factors, overall friendliness of an area, and the environment’s influence on well-being. Each participant rated each image once for all statements to establish a human baseline. The algorithms’ scores were generated using the exact same prompt, repeated multiple times to account for non-determinism. We then compared the AI’s scores to the humans’ distribution of scores and evaluated their alignment according to different experiential qualities across diverse visual environments. Full article

Smart grids require real-time ancillary services from large-scale distributed energy storage (DES), creating a conflict between second-scale physical response needs and the slow confirmation of trust mechanisms like blockchain. Traditional VPPs lack scalability and trust for massive participation, while decentralized approaches struggle with mismatched time scales. We propose a framework that decouples real-time dispatch from asynchronous settlement. An off-chain matcher uses a physics-aware model, including a novel “service holding time” ($T^{\mathrm{service}}$) constraint and power (kW) envelopes, for fast assignments. A separate on-chain proof-of-stake (PoS) layer handles incentives and penalties (slashing) asynchronously. We formulate the MILP dispatch problem and provide a fast online heuristic alongside a MINLP decomposition benchmark. Co-simulations (IEEE 33-node) show that our scheme significantly outperforms baselines in success rate and latency, is robust against non-compliant nodes due to the PoS mechanism, and thereby offers a scalable and trustworthy solution. Full article

Digital Twins (DTs) have emerged within the last decade due to the adequate maturity of several key technologies contributing to the realization of real-time virtual–physical world synchronization. Advancements in sensing, connectivity, computing processing power, and artificial intelligence have contributed to the deployment of DTs in several application sectors, such as in agriculture. This work aims to provide a scoping review of recent advancements in digital twin technologies and agricultural applications. Results indicate a special focus on plant-level models, soil moisture, and machinery, while most works are based on drone imagery combined with machine learning routines. Several works use the term DTs rather loosely, often describing systems that resemble decision support tools rather than a fully synchronized virtual–physical setup. Data integration emerges as the most important bottleneck, especially when the system mixes satellite data, local sensory data, and simulation outputs. Yet it is suggested that DTs could eventually support more adaptive and resource-efficient farm management. However, the field is still missing common frameworks and long-term evaluations. Based on this review, progress depends on better data-handling pipelines, clearer definitions of operational DTs, and more attention to economic and practical constraints faced by farmers rather than just technical proofs of concept. Full article

Breast cancer (BC) remains a major global health burden, consistently standing as the foremost contributor to cancer-related illness and death among women across the world. This meta-analysis aimed to evaluate the diagnostic accuracy of AI-assisted ultrasound elastography (UE) for BC detection by considering various factors, such as AI models, segmentation, cross-validation, data augmentation, the evaluation phase, and the addition of conventional ultrasound. PubMed, Cumulative Index to Nursing and Allied Health Literature, Embase, Scopus, and Web of Science were searched from inception to 22 June 2025, for observational studies using any AI-aided UE modality in BC classification compared to histopathology. We extracted binary diagnostic accuracy data and employed the split component synthesis method for pooled outcomes. Out of 501 identified records, 39 studies (6191 samples) were included in the meta-analysis. The overall diagnostic performance showed 90.3% sensitivity (95% confidence interval [CI] 86.4–93.1%), 88.0% specificity (95% CI 83.6%–91.4), a positive likelihood ratio of 7.5 (95% CI 5.4–10.5), a negative likelihood ratio of 0.110 (95% CI 0.078–0.156), and a diagnostic odds ratio of 68.3 (95% CI 42.3–110.1). Heterogeneity was substantial (I² = 78.0%), and the funnel plot demonstrated mild positive asymmetry. Subgroup analyses indicated improved diagnostic performance in studies that employed automatic or no segmentation, cross-validation, data augmentation, retrospective designs, models evaluated during the training phase, classical machine learning approaches, and the combination of B-mode ultrasound with elastography. Despite the presence of heterogeneity and the possibility of overestimation, AI-aided UE demonstrated superior diagnostic performance compared to UE alone. Researchers should consider adopting automatic segmentation, cross-validation, augmentation, and combining UE with conventional ultrasound. Our meta-analysis also explored the potential integration of AI-aided UE into breast cancer screening practices. Full article

Generative Artificial Intelligence (GenAI) has been a viable technology for decades, yet widespread adoption in healthcare and academic settings has remained limited to research. One possible explanation for this is limited understanding about the beliefs around GenAI use amongst faculty and students training in biomedical disciplines that frequently lead to non-physician healthcare careers, including physical therapy (PT), occupational therapy (OT), allied health (AH), and biomedical engineering (BME). Furthermore, no known studies exist assessing differences that may exist across those disciplines. Given the significant number of professionals in those disciplines and the outsized impact they have on the healthcare system, investigating their beliefs around GenAI use is vital before widespread adoption. Accordingly, we investigated the perceptions of GenAI among students and faculty in the aforementioned fields that frequently lead to careers in healthcare. We found that knowledge of GenAI significantly influences comfort with its use completing college coursework including whether respondents believed it contributed to the process of completing that coursework and whether use of GenAI enhances learning. Interestingly, however, there were no statistically significant differences in perceptions of GenAI across disciplines, roles, or institution sizes. Qualitative findings revealed concerns about plagiarism, decline of critical thinking skills, and ethical challenges, while also recognizing GenAI’s potential to enhance learning efficiency and idea generation. Critically, the study results emphasize the need for proper training and guidelines to ensure GenAI is integrated responsibly into healthcare-related education. Full article

Cerebral aneurysms and arteriovenous malformations are life-threatening hemodynamic pathologies of the brain. While surgical intervention is often essential to prevent fatal outcomes, it carries significant risks both during the procedure and in the postoperative period, making the management of these conditions highly challenging. Parameters of cerebral blood flow, routinely monitored during medical interventions or with modern noninvasive high-resolution imaging methods, could potentially be utilized in machine-learning-assisted protocols for risk assessment and therapeutic prognosis. To this end, we developed a linear oscillatory model of blood velocity and pressure for clinical data acquired from neurosurgical operations. Using the method of Sparse Identification of Nonlinear Dynamics (SINDy), the parameters of our model can be reconstructed online within milliseconds from a short time series of the hemodynamic variables. The identified parameter values enable automated classification of the blood-flow pathologies by means of logistic regression, achieving a balanced accuracy of 74%. Our results demonstrate the potential of this model for both diagnostic and prognostic applications, providing a robust and interpretable framework for assessing cerebral blood vessel conditions. Full article

Deep learning-based flow field prediction for microclimate pollutant dispersion represents an emerging and promising methodology, where effectively integrating meteorological, spatial, and temporal information remains a critical challenge. To address this, we propose a novel non-intrusive reduced-order model (ROM) that synergizes a Dilated Convolutional Autoencoder (DCAE) with pre-trained large language models (LLMs). The DCAE, leveraging nonlinear mapping, was employed for extracting low-dimensional spatiotemporal flow field features. These features were then combined with textual prototypes via text embedding to enable few-shot inference using the LLM-based flow field prediction method. To optimize the utilization of pre-trained LLMs, we designed a specialized textual description template tailored for pollutant dispersion data, which enhances the contextual input of meteorological conditions to guide model predictions. Experimental validation through three-dimensional urban canyon simulations conclusively demonstrated the efficacy of the convolutional autoencoder and LLM-based framework in predicting pollutant dispersion flow fields. The proposed method exhibits remarkable transfer learning capabilities across varying street canyon geometries and meteorological conditions while significantly representing a 9.85× acceleration in prediction compared to Computational Fluid Dynamics (CFD). Full article

Textual and visual representations of data play a key role in data science and artificial intelligence by supporting effective and user-friendly communication. Among existing approaches, automatic data-to-text generation aims to produce natural language descriptions from structured data sources. This paper presents an interpretable fuzzy framework for generating data to text based on linguistic contexts and computational perception networks evaluated through formal concept analysis. The proposed framework is organized into four main stages: (i) transforming numerical data sets into linguistic contexts, (ii) generating computational perceptions from linguistic contexts, (iii) building computational perceptions networks to automatically generate natural language summaries, and (iv) validating the generated texts through comparison with summaries obtained using formal concept analysis–based baselines. To the best of our knowledge, this is the first work to address the generation of linguistic summaries through an interpretable process that transforms data into linguistic contexts and subsequently into computational perceptions. Another key difference from previous work lies in the verification of the linguistic summaries generated through these computational perceptions by using a formal method. A software prototype was implemented and evaluated using real photovoltaic station data provided by a local energy operator in Puerto Real (Cádiz, Spain). Experimental results show that the proposed fuzzy framework improves the interpretability and consistency of the generated summaries when compared with others approaches, demonstrating its potential for explainable and user-centered data-to-text generation. Full article

Algorithmic management systems increasingly coordinate work, allocate resources, and support decisions in corporate, public sector, and research environments. Yet many such systems remain opaque: they optimize and score effectively but struggle to communicate rationales that are contextual, auditable, and defensible under emerging governance expectations. Large language models (LLMs) can help bridge this gap by translating quantitative signals into human-readable explanations and enabling interactive clarification. However, LLM integration also introduces new risks—hallucinated rationales, bias amplification, prompt-based security failures, and automation dependence—that must be governed rather than merely engineered. This article proposes a governance-oriented architecture for LLM-augmented algorithmic management. The model combines the following elements: an algorithmic decision core; an LLM-based cognitive interface for explanation and dialogue, and a verification and governance layer that enforces policy constraints, provenance, audit trails, and human-in-command oversight. The framework is developed through targeted conceptual synthesis and normative alignment with key governance instruments (e.g., the EU AI Act, GDPR, and ISO/IEC 42001). It is illustrated through cross-domain scenarios and complemented by a demonstrative synthetic-trace simulation that highlights transparency–latency trade-offs under verification controls. Using the demonstrative simulation (n = 120 decision events), the framework illustrates a mean baseline latency of 100.3 ms and a mean LLM-augmented latency of 115.8 ms (≈15.5% increase), a mean explanation validity proxy of 85.6%, and a simulated constraint-satisfaction rate of 94.2% (113/120 events), with failed cases routed to review. These values are presented as design-level indicators of operational plausibility and governance trade-offs, not empirical performance benchmarks or state-of-the-art comparisons. The paper contributes a conceptual and governance-oriented architectural blueprint for integrating generative AI into organisational decision systems without sacrificing accountability, compliance, or operational reliability. Full article

In this paper, we propose FDSTCN-EEG, which is a customized federated learning framework for EEG-based seizure detection that leverages deep depthwise separable temporal convolutions and asynchronous model aggregation. The network design tackles major problems in distributed healthcare AI by jointly boosting computational efficiency, training rate, and classification performance. In this paper, we propose FDSTCN-EEG, a novel federated learning framework specifically designed for EEG-based seizure detection. Our key contributions are threefold: first, high architectural efficiency with depthwise separable temporal convolutions, reducing parameters by 40.4% (9.8M to 5.8M) while maintaining accuracy of 96.96%; second, speeding up training by a factor of 38.5% compared with synchronous learning via an asynchronous aggregation protocol; finally, a privacy-preserving decentralized learning model without sharing raw EEG data and with the capability of coping with the heterogeneous clinical technology infrastructure. Extensive experiments show superior performance (accuracy: 96.96%, F1-score: 97.02%) and practical viability for real-world seizure monitoring systems. Such work introduces a practical privacy-preserving medical AI paradigm, which balances model efficiency with training scalability and clinical quality accuracy. Full article

Introduction: Large language models (LLMs) are increasingly employed as cognitive aids in research and professional inquiry, yet their fluent outputs are frequently regarded as authoritative knowledge. We contend that this practice signifies a fundamental epistemic misalignment. Methods/Approach: Building on Peirce’s theory of inquiry, Sellars’ concept of the space of reasons, Stanovich’s tripartite model of cognition, and knowledge-building theory, we develop a conceptual framework for analyzing epistemic agency in human–LLM collaboration. Results/Argument: We demonstrate that LLM outputs fail to satisfy the conditions for knowledge because they lack reflective regulation, resistance to revision, and normative commitment. While LLMs display strong autonomous and algorithmic abilities (e.g., pattern recognition and hypothesis development), reflective control remains a distinctly human function. This asymmetry supports a principled division of epistemic labour and motivates the concept of the Knowledge-Building Partner (KBP): an AI system designed to support inquiry without claiming epistemic authority. Discussion/Implications: We identify prompt-, system-, and model-level design requirements and introduce a triangulated framework for operationalizing epistemic agency through explainable AI, discourse analysis, and rational-thinking measures. These contributions collectively reposition LLM limitations as epistemic design challenges rather than technical issues. Full article

Continuous non-invasive blood pressure monitoring holds significant promise for cardiovascular disease management, yet cuff-based methods remain limited by their intermittent nature. Machine learning approaches leveraging photoplethysmography (PPG) and electrocardiography (ECG) signals present compelling alternatives, though questions persist about which signal type contributes more predictive value. This study compares traditional machine learning models, ensemble methods, and deep learning architectures for estimating systolic blood pressure from physiological waveforms. We extracted 55 features from PPG and ECG recordings of 100 subjects in the MIMIC-III Waveform Database, yielding 3000 segments with invasive arterial blood pressure as ground truth. Data splitting was performed at the subject level (70/15/15 train/validation/test) to prevent data leakage. Evaluation included regression metrics, British Hypertension Society grading, SHAP-based explainability, and ablation studies. Among all models, LightGBM achieved the best performance with mean absolute error of 15.97 mmHg, placing it at BHS Grade D. While SHAP analysis showed ECG features contributing 54.7% of importance versus 45.3% for PPG, our ablation study revealed that PPG-only models achieved comparable performance (MAE 15.97 vs. 16.23 mmHg), with the difference not statistically significant (p = 0.226). These results suggest that PPG-only wearable devices are viable for blood pressure estimation, as adding ECG features provides no statistically significant improvement. However, all configurations achieved only BHS Grade D, indicating that personalized calibration may be necessary for clinical acceptability. Full article

The implementation of artificial intelligence (AI) in education is gaining more attention, and as a result, more research is being conducted on the views and conceptualisations of AI by educators. The understanding of teacher candidates is vital for the AI integration in education, which should be human-centred, and still, there is a lack of studies focusing mainly on teacher candidates in the field of the native language. This qualitative phenomenological research aimed to explore metaphors of 46 Turkish language teacher candidates (third- and fourth-year undergraduates in Northern Cyprus) representing their answer to the prompt “AI is like because…”. The data were collected through open-ended questions and analysed using content analysis along with expert validation. Participants produced 46 valid metaphors, which were divided into five thematic categories: (1) AI as Teacher or Learner (21.7%), (2) AI as Method/Strategy (21.7%), (3) AI as Evolving Living Organism (13%), (4) AI as Guide/Helper (21.7%), and (5) AI as Danger/Threat (21.7%). Four groups expressed positive or neutral attitudes towards AI, such as considering it a clever teacher, a useful tool, a growing entity, or a guide. One category revealed negative views, perceiving AI as a destructive force. Overall, 78.3% of participants expressed optimistic views about AI, while 21.7% of them pointed to concerns. Turkish language teacher candidates generally perceive AI as a supportive, human-like assistant in the classroom, but a few of them express concerns about its existence. These results emphasise the importance of incorporating AI literacy and ethics into teacher education. Equipping future language teachers with the skills to use AI in the classroom might be a way of implementing AI in schools that is confident, critical, and human-centred. Full article

Purpose: The aim is to evaluate the effectiveness of artificial intelligence (AI)-based automatic segmentation as a predictive tool for clinical peer review in prostate cancer patients treated with hypofractionated radiotherapy. Methodology: A retrospective analysis was conducted on 62 patients treated across three Italian centers between 2020 and 2025. CT images were segmented using software based on 3D U-net models. Three workflows were compared: manual segmentation (C man), automatic segmentation (C AI), and AI-based segmentation adjusted by clinicians (C adj). Quantitative metrics used for comparison included the Dice Similarity Coefficient (DSC) and Hausdorff Distance (HDmax). Statistical analysis involved Welch’s t-test and Cohen’s d for effect size. Results: The results showed a significant improvement in agreement between C AI and C adj compared to C man. Median DSC for CTV increased from 0.80 (C man) to 0.92 (C adj), while HDmax decreased from 12.33 mm to 9.22 mm. Similar improvements were observed for the bladder and anorectum. All differences were statistically significant (p < 0.0001), with large effect sizes (Cohen’s d > 0.8). Discussion: AI use demonstrated a reduction in interobserver variability and segmentation time, enhancing workflow standardization. The C adj workflow, where the physician acts as a reviewer of AI-generated contours, proved effective and potentially integrable into clinical peer review. The predictive peer review refers to a preliminary support step in the clinical review process rather than a substitute for medical decision-making. Full article

The effectiveness of deep learning methods in image segmentation has led to interest in their deployment for 3D point cloud segmentation, particularly in the context of pre-grasp identification of a unique object amongst distractors. However, existing 3D object datasets are not ideal for training and evaluation of these methods. Datasets developed for grasp planning are often CAD models that are too clean for sim-to-real transfer. Real-world datasets can lack texture information or have been collected using sets of objects and/or specialized sensor setups that are hard to reproduce. In this work, we introduce the MiniMarket80 dataset to address this gap.The dataset consists of 1200 colored point cloud partial views, each of 80 standard grocery objects, collected with widely used Realsense RGB-D cameras (D415 and D435) under variable lighting conditions. We also provide a complete pipeline to generate a per-object segmentation dataset from these partial views suitable for use in training. We use this dataset to evaluate 11 state-of-the-art point cloud segmentation methods. Only four of these are able to (partially) segment the target object in a real-world test, still producing significant false positives and false negatives. Full article

Recent advancements in XAI have radically changed the way that AI systems are evaluated, as transparency and trustworthiness are now valued as highly as performance. This is especially true in medical applications, as, in order for such tools to be used in practical applications, interpretability is a key requirement for clinical adoption. Electroencephalography (EEG) analysis, in particular, has seen a significant rise in research, as the difficult and complex nature of EEG signals benefits from these methods, enabling researchers and practitioners to gain new insights from the vast amount of data that is now available. This survey presents a comprehensive analysis of the latest trends and advancements in XAI for EEG analysis. First, we provide a brief overview of fundamental EEG tasks, available datasets, and AI model approaches used for analysis. Then, we classify XAI methods using well-established taxonomies in XAI research, such as locality and generalization of explanations. By exploring all relevant XAI techniques in EEG analysis, our study offers researchers a clear perspective on the current state of the field and identifies potential research gaps. Our review indicates that current XAI approaches for EEG often face limitations in robustness, consistency, and neuroscientific grounding. These findings highlight the need for more reliable and domain-informed explainability methods to support trustworthy EEG analysis in research and clinical practice. Full article

Convolutional neural networks (CNNs) can achieve high accuracy in photovoltaic (PV) fault recognition from RGB imagery, yet their decisions may rely on shortcut cues induced by heterogeneous backgrounds, viewpoints, and class imbalance. This work presents a multi-method explainability audit on the Kaggle PV Panel Defect Dataset (six classes), comparing five architectures (Baseline CNN, VGG16, ResNet50, InceptionV3, EfficientNetB0). Explanations are obtained with LIME superpixel surrogates (reported together with kernel-weighted surrogate fidelity), occlusion sensitivity (quantified via IoU@Top10% against consistent proxy masks, Shannon entropy, and Hoyer sparsity), and Integrated Gradients evaluated by deletion–insertion faithfulness and a Faithfulness Gap. While ResNet50 yields the best predictive performance, EfficientNetB0 shows the most consistent faithfulness evidence and stable panel-centered attributions. The analysis highlights class-dependent vulnerability to context cues, especially for the Clean and damaged classes, and supports using quantitative explainability diagnostics during model selection and dataset curation to mitigate shortcuts in vision-based PV monitoring. Full article

Artificial intelligence (AI) is rapidly transforming virology by strengthening pandemic preparedness, enhancing our molecular understanding of virus–host interactions, and accelerating the discovery and development of novel antiviral therapies. Yet, the same technologies also pose urgent biosecurity risks, particularly by enabling the development of bioweapons or identifying strategies that maximize harm. This paper presents a critical content analysis of current and emerging AI applications in virology, including tools used to detect synthetic alterations in viral genomes, assess the severity of new variants, and design clinical vectors for gene therapy. It also highlights the potential for misuse, whether intentional or due to poor data quality and flawed model training. Drawing on case studies, public databases, and documented applications from research institutions and biotechnology firms, the analysis shows that AI can integrate large datasets to reduce reliance on animal testing in drug development, improve therapeutic precision, and allocate resources more effectively during outbreaks. However, the increasing accessibility of AI tools and genomic data also creates vulnerabilities, especially as models become capable of autonomously interpreting the scientific literature and mining bioinformatics databases. To address this dual-use dilemma, the paper proposes targeted and adaptable policy recommendations for governments, research institutions, and commercial biotech firms, emphasizing pre-emptive oversight, responsible innovation, and ethical AI deployment. These recommendations are designed for immediate relevance yet flexible enough to evolve alongside the expanding role of AI in global health. Full article

This article evaluates the use of probing classifiers to modify the internal hidden state of a chess-playing transformer, which has been trained on sequences of chess moves and can generate new moves with prompted. Probing classifiers are a technique for understanding and modifying the operation of neural networks in which a smaller classifier is trained to use the model’s internal representation to learn a probing task. The aim of this research is to discover whether the learned model possesses an editable internal representation of the chess game, despite being trained without explicit information about the rules of chess. We contrast the performance of standard linear probes against Sparse Autoencoders (SAEs), a latent space interpretability technique designed to decompose polysemantic concepts into atomic features via an overcomplete basis. Our experiments demonstrate that linear probes trained directly on the residual stream significantly outperform probes based on SAE latents. When quantifying the success of interventions via the probability of legal moves, linear probe edits achieved an 88% success rate, whereas SAE-based edits yielded only 41%. These findings suggest that while SAEs are valuable for specific interpretability tasks, they do not enhance the controllability of hidden states compared to raw vectors. Finally, we show that the residual stream respects the Markovian property of chess, validating the feasibility of applying consistent edits across different time steps for the same board state. Full article

In Industry 4.0 environments, operators and production managers frequently query industrial databases for production monitoring, quality control, and equipment maintenance using natural language. Existing Chinese NL2SQL systems often process semantic, program, and schema information in a single encoder, which leads to semantic-program interference and frequent structural or schema errors in the generated SQL. We present CIR-SQL, a dual-model framework that separates intent recognition from SQL generation via structured intermediate representations, decoupling semantic understanding from program synthesis. CIR-SQL employs a seven-category intent classification system (simple_select, count_query, filter_query, max_min_query, sort_query, join_query, group_by_query) and leverages large language models for intent recognition and structured information extraction. A three-level hierarchical backtracking strategy (SQL, context, intent) further improves robustness by correcting different error types. The architecture is particularly suited to Industry 4.0 scenarios where Chinese-speaking operators interact with complex industrial databases containing production data, quality metrics, and equipment status information. Full article

Background: Machine learning (ML) has been widely adopted in decision-making, making fairness a central ethical and scientific priority. We developed the Themis chatbot, a Large Language Model (LLM) system designed to explain concepts of ML fairness in an accessible, conversational format. Methods: The development followed four stages: (1) curating a document corpus of 286 peer-reviewed publications on ML fairness; (2) development of Themis by combining a modern LLM (OpenAI’s GPT-4o) with Retrieval Augmented Generation (RAG); (3) creation of a 340-item benchmark dataset, the FairnessQA; and (4) evaluating performance against state-of-the-art non-augmented LLMs (DeepSeek R1, GPT-4o, GPT-5, and Grok 3). Results: For the multiple-choice questions, Themis achieved an accuracy of 96.7%, outperforming DeepSeek R1 (90.0%), GPT-4o (89.3%), GPT-5 (92.0%), and Grok 3 (86.7%), and the overall difference was statistically significant (χ²(4) = 10.1, p = 0.038). In the closed-ended questions, Themis achieved the highest accuracy (96.7%), while competing models ranged from 78.0% to 84.0%, and the overall difference was significant (χ²(4) = 23.9, p < 0.001). In the open-ended questions, Themis achieved the highest mean scores for correctness (M = 4.62), completeness (M = 4.59), and usefulness (M = 4.56), and differences were statistically significant (correctness: F(4, 195) = 20.91, p < 0.001; completeness: F(4, 195) = 7.76, p < 0.001; usefulness: F(4, 195) = 2.90, p < 0.001). By consolidating scattered research into an interactive assistant, Themis makes fairness concepts more accessible to educators, researchers, and policymakers. This work demonstrates that retrieval-augmented systems can enhance the public understanding of machine learning fairness at scale. Full article

Research on metal–organic frameworks (MOFs) bridges the fields of chemistry and materials science. MOFs consist of metal ions linked together by long organic molecules. These materials are known for their high porosity and large surface area, with numerous applications ranging from storage of various gases to medical uses. Recent developments show that artificial intelligence (AI) is revolutionizing the discovery and design of MOFs. Despite these advancements in AI-driven approaches in MOFs, many challenges remain in processes such as data quality assurance and experimental validation. In this perspective, we highlight recent progress in MOFs and discuss the role of AI in this truly interdisciplinary field. Full article

The deep water environment is complex and variable, and underwater images are easily affected by light scattering, water absorption and other factors, resulting in blurred details and color distortion. The existing enhancement methods generally suffer from poor model generalization, parallel task conflicts, and imbalance between real-time performance and optimization effect. To address this challenge, this study proposes a dual-branch lightweight network named DualOadamNet for underwater image processing with optical-aware detail augmentation. The model is based on the branches of global and local feature extraction, and the Optical-Aware Detail Augmentation Module with the characteristics of human visual simulation repair is introduced to repair the image naturally. Combined with pixel rearrangement operation, the model achieves efficient feature scale extraction. The experimental results on UIEB, EUVP and LSUI datasets show that the proposed method achieves average full-color evaluation metrics of 24.63 for PSNR, 0.919 for SSIM, and 3.267 for UIQM. Additionally, the real-time enhancement speed of a 1080p resolution image is 85.299 FPS. Full article

Network intrusion detection is a core component of modern cybersecurity, but it remains challenging due to highly imbalanced traffic, heterogeneous feature types, and a presence of short-term temporal dependencies in network flows. Traditional machine learning models often rely on handcrafted features and struggle with complex attack patterns, while deep learning approaches may become overly complex or difficult to interpret. In this paper, we propose a neural intrusion detection method that combines structured feature preprocessing with a compact hybrid architecture. Numerical and categorical traffic features are processed separately using robust normalisation and trainable embeddings, and then merged into an unified representation. The proposed model builds on a multi-scale feature interaction block, followed by channel-wise attention and a single bidirectional gated recurrent unit layer with attention pooling to capture short-term temporal behavior. The method is evaluated on two widely used benchmark datasets, i.e., the CIC-IDS2017 and CSE-CIC-IDS2018 dataset. Experimental results show that the proposed approach consistently outperforms the classical machine learning baselines and achieves competitive or superior performance compared to the recent deep learning methods proposed in the literature. The results confirm that the proposed architectural choices effectively capture both feature interactions and temporal patterns in network traffic. Full article

Emergency triage is a critical healthcare action that could be improved through the use of artificial intelligence (AI) systems, as these have been shown to achieve accuracy rates of approximately 70–90% for LLMs and AUC values ranging from 0.75 to 0.95 for common AI models. However, these systems face challenges related to the rights and interests of the individuals involved. The European Union’s normative framework, including not only data protection regulations but also the AI Act and medical device regulations, imposes conditions on the use of AI, and these are analyzed here. Our conclusions reveal that Article 22 of the General Data Protection Regulation (GDPR) makes it difficult to justify the establishment of fully automated decision-making models for triage. That accountability obligations for implementers (Fundamental Rights Impact Assessments: FRIAs) and data controllers (data protection impact assessments: DPIAs) can contribute to better design of AI-based decision-making in triage. Furthermore, with regard to the information rights set out in the GDPR, these have been complemented by the right to an explanation under Art. 86 AI Act in the use of high-risk AI systems. Unfortunately, regulation relating to general-purpose AI models may create some gaps in this framework. The implementation of AI systems for automated decision-making in triage has the potential to improve medical care, but their use requires clarification of applicable regulations and safeguards for patients’ rights. Full article