SACAIR 2025 Springer Series

Abstract: The question of which contextual topic modelling algorithm performs best has become increasingly important as the field rapidly develops new approaches. However, existing evaluations typically focus on limited datasets and metrics, often claiming superiority for novel algorithms. This study presents a comprehensive empirical evaluation of eleven contextual topic modelling algorithms across ten diverse datasets, five numbers of topics, and four performance metrics, resulting in 22,000 metric evaluations. Rather than identifying a single superior algorithm, our results reveal clear evidence of performance complementarity: different algorithms excel on different problem instances and under different evaluation criteria. Through aggregate performance analysis, pairwise dominance comparisons, and multi-objective Pareto frontier analysis, we demonstrate that algorithmic dominance varies significantly across problem instances. Most remarkably, in 84% of cases, all algorithms are Pareto optimal when considering all metrics simultaneously, indicating that each offers unique strengths that cannot be dominated by others. These findings challenge the common practice of claiming algorithmic superiority and suggest that algorithm selection should be guided by specific problem characteristics and performance priorities rather than blanket recommendations. Our work contributes to the growing recognition that performance complementarity is fundamental to computational problems, extending this concept to contextual topic modelling and providing a foundation for future algorithm selection research. Code used to conduct this study is provided.

Abstract: Document type classification is essential for effective information retrieval and management within archival systems, particularly in low-resource languages like Afrikaans. This study examines the feasibility of utilising multilingual transformer-based language models for document classification within a South African archival context. We followed a basic linguistic approach to prepare Afrikaans text documents for classification into six categories: academic papers, media reports, books, interviews, book reviews, and theses or dissertations. We compare fine-tuned transformer models, hybrid models combining traditional classifiers with contextual embeddings, and a baseline SVM (TF-IDF) classifier, using stratified 5-fold cross-validation and a hard voting ensemble for robust evaluation. Our findings reveal that the SERENGETI transformer-based model outperformed other multilingual models, achieving a weighted F1 score of 0.964, while hybrid approaches performed competitively. However, the baseline SVM (TF-IDF) model outperformed all transformer and hybrid models, with a weighted F1 score of 0.978. This research demonstrates the potential and current limitations of neural language models and hybrid strategies for enhancing document classification in Afrikaans archival systems. If implemented, the classifier can improve indexing efforts and reduce pressure on archival personnel who handle over 5,000 new items annually.

Abstract: Automatic modulation classification (AMC) is a critical function in wireless communication systems that is used to identify the modulation type of a signal without prior knowledge. AMC has historically been tackled using likelihood or feature-based methods, yet recent research has focused on using deep neural networks (DNNs) as they outperform the classical methods in challenging signal channel conditions. However, deep learning (DL) based classifiers have a vulnerability to adversarial attacks that can significantly deteriorate their classification performance. This paper explores the robustness of different AMC classifiers to the white-box fast gradient method (FGM) and projected gradient descent (PGD) attacks under different perturbation-to-noise ratios (PNRs) and signal-to-noise ratios (SNRs) for a noisy signal channel. The investigated AMC classifiers consist of the quasi-hybrid likelihood ratio rest (QHLRT), a k-nearest neighbour (KNN) that uses higher-order cumulants, and the parameter estimation and transformation-based CNN-GRU deep neural network (PET-CGDNN). The adversarial attacks are found to have limited transferability to the QHLRT and the KNN classifiers when scaled to be imperceptible against the noise of the signal channel. Based on this finding, we propose a hybrid classifier that uses the neural rejection technique through a support vector machine (SVM) that acts as a switching mechanism to decide whether to use the KNN or PET-CGDNN to classify the modulation type. The hybrid classifier demonstrates improved robustness against attacks, while benefiting from the good performance of the DNN. Interdisciplinary motivation: Deep learning methods form an important component of future telecommunications systems. This paper combines methods from both telecommunications and deep learning to investigate the vulnerability that deep learning could expose the telecommunications systems to.

Abstract: Deep learning has revolutionised healthcare applications, achieving remarkable success in medical diagnosis and treatment prediction. However, the inherent opacity of these models presents significant challenges for clinical deployment, where interpretable explanations are crucial for patient trust and regulatory com-pliance. This paper presents a novel constraint-aware counterfactual explanation model for generating personalised dietary interventions in anaemia patients. Anaemia affects over 1.9 billion people globally, yet existing explainable AI methods fail to provide clinically feasible and culturally appropriate recommenda-tions. We develop a causal machine learning approach that integrates Pearl's caus-al hierarchy with domain-specific constraints to produce interpretable "what-if" scenarios. Our model incorporates nutritional, cultural, and economic constraints through augmented Lagrangian optimisation, ensuring recommendations remain clinically feasible whilst maintaining semantic meaningfulness. Experimental re-sults demonstrate superior performance compared to existing explainable AI methods, achieving 84.3% anaemia reversal rates (vs 71.8% best baseline), 89.1% counterfactual validity, and 4.2 interpretability scores. The model gener-ates recommendations requiring an average of 2.3 dietary changes within cogni-tive load thresholds whilst maintaining O(n log n) computational complexity suit-able for real-time clinical deployment. This work advances explainable AI in healthcare by demonstrating how domain-specific constraints can enhance both interpretability and clinical utility of counterfactual explanations for chronic dis-ease management.

Abstract: Precision agriculture requires the estimation of plant growth stages in real-time. When the plant growth stage is known, the wastage of resources in cultivation, such as nutrients and water, is reduced as only the required resources need to be supplied. Plants at different growth stages, however, have similar morphological features, which can make autonomous growth stage estimation difficult. This paper presents two feature extraction methods for growth stage estimation: one that uses a bank of Gabor filters and morphological operations, and the other that uses pre-trained convolutional neural networks (CNNs) and transfer learning. We test these methods on a publicly available plant growth stage dataset (``bccr-segset``) for two species, canola and radish, grown and captured under indoor conditions. The two proposed feature extraction methods are compared, using support vector machines and boosted trees as classifiers. We find that both methods are suitable for real-time applications, and that CNN features outperformed the hand-crafted features, both with regard to speed and accuracy. The best system (VGG-19 features, classified with a radial basis function support vector machine) obtained an accuracy of 98. 4% for both species, processing an image in 0.08 seconds.

Abstract: Deep neural networks (DNN) have a high potential for predicting and mitigating equipment failure in large industrial plants. However, plant anomalies are rare events that result in extremely unbalanced datasets, which poses a challenge for traditional DNN classifiers. Weighted loss functions such as focal loss and weighted binary cross-entropy (WBCE) have emerged as a promising approach to deal with class imbalance, where higher weightings are assigned to the anomaly class during training. This study proposes three new weighted loss function variants, i.e. weighted polynomial binary cross entropy (WPBCE) loss, weighted hinge loss and weighted squared hinge loss, and systematically evaluates these across three DNN architectures: long short-term memory (LSTM), temporal convolutional network (TCN), and multi-layer perceptron (MLP). The results show that the weighted loss function variants improve recall and yield more stable configurations across all algorithms, when compared to focal loss and WBCE, for predicting the onset of abnormal operating events in a real-world South African mineral grinding mill. Importantly, this work demonstrates that the weighted squared hinge and WPBCE, when combined with the LSTM model, offer a reliable solution for early and accurate anomaly prediction.

Abstract: Class-imbalanced datasets are a common occurrence in real- world applications. The imbalance between minority and majority classes exists due to the over-representation of one class compared to another in a dataset. The class imbalance might reflect a system’s behaviour over time. However, the class imbalance causes sub-optimal performance for machine learning models that predict the system’s future behaviour. Various techniques are used to reduce the negative impact of class- imbalanced datasets on machine learning models. Data resampling techniques are one of the main techniques, and the subdivisions of data resampling techniques include oversampling and undersampling. Oversam- pling techniques have outperformed undersampling techniques in most studies, and most data resampling techniques are derived from oversampling. However, some oversampling techniques are ineffective when used on minority-class datasets that lack within-class variation and have a high-class imbalance. In this dissertation, an analysis was performed to understand the changes in within-class variation before and after over- sampling for nine datasets. Additionally, classification performance was measured for standard and hybrid oversampled datasets. A novel hybrid oversampling technique that uses k-Means and ADASYN was implemented. Hybrid oversampling techniques generated synthetic examples that marginally changed the within-class variation and had the highest recall score compared to standard oversampling techniques across nine datasets

Abstract: This study evaluates four open-source Optical Character Recognition (OCR) systems which are Tesseract, EasyOCR, PaddleOCR, and TrOCR on real world food packaging images. The aim is to assess their ability to extract ingredient lists and nutrition facts panels. Accurate OCR for packaging is important for compliance and nutrition monitoring but is challenging due to multilingual text, dense layouts, varied fonts, glare, and curved surfaces. A dataset of 231 products (1{,}628 images) was processed by all four models to assess speed and coverage, and a ground truth subset of 113 images (60 products) was created for accuracy evaluation. Metrics include Character Error Rate (CER), Word Error Rate (WER), BLEU, ROUGE-L, F1, coverage, and execution time. On the ground truth subset, Tesseract achieved the lowest CER (0.912) and the highest BLEU (0.245). EasyOCR provided a good balance between accuracy and multilingual support. PaddleOCR achieved near complete coverage but was slower because it ran on CPU only due to GPU incompatibility, and TrOCR produced the weakest results despite GPU acceleration. These results provide a packaging-specific benchmark, establish a baseline, and highlight directions for layout-aware methods and text localization.

Abstract: While fine-tuning transformer-based pre-trained speech models improves speech recognition for low resource languages, the approach increases the risk of speaker attribute bias in the resulting target language automatic speech recognition (ASR) systems. This work investigates gender bias in two state-of-the-art pre-trained speech models, MMS and Whisper, fine-tuned for ASR on three African languages: Bemba, Nyanja, and Swahili. We fine-tune models on gender-specific as well as gender-balanced datasets, and estimate and compare gender bias across different settings. Our results show varying degrees of gender bias in the fine-tuned models, even with gender-balanced fine-tuning, suggesting influence from pre-trained models. Inconsistencies in gender-specific fine-tuning further confirm the transfer of bias from pre-trained models. Additionally, an ablation study shows no relationship between training data size and gender bias.

Abstract: Panoramic dental radiographs (OPG) are the only imaging modality that captures the entire dentition in a single exposure. To support dentists with diagnosing caries it is essential to find indications for cavities on those images. While recent deep learning methods show strong results on in-distribution test sets, the generalization on out-ofdistribution datasets is mostly untested. In this study, we suggest a twostage deep learning pipeline for caries detection on single-tooth images extracted from OPGs: (1) image-level classification using a DINO-based transformer backbone and (2) instance-level segmentation using MaskR-CNN. We perform experiments on data from the University Medical Center Schleswig-Holstein (UKSH). To study generalization, we test models on an out-of-distribution set from the Federal University of Bahia (UFBA) and also evaluate a mixed-domain setting including both UKSH and UFBA data. Further we investigate the influence of strong augmentation techniques. Results show that classification performance is high on in-distribution data but significantly drops when applied to outof-distribution samples. Segmentation performance is moderate across all settings, with limited robustness under domain shift. These findings suggest that in-distribution results overestimate real-world performance and underscore the importance of evaluating domain shifts in dental AI pipelines.

Abstract: Syllables are fundamental units in speech production and carry prosodic information, but their acoustic and linguistic properties across different language families are not well understood. This study examines syllable discovery approaches across South African languages using algorithmic syllabification and S5-HuBERT, a self-supervised speech representation model that demonstrates emergent syllabic organization. We analyzed speech recordings from eleven languages representing five language families in South Africa using a systematic comparison of rule-based and data-driven syllable discovery methods. We evaluated both approaches using cross-linguistic consistency measures and acoustic quality assessments across speakers. Our analysis reveals fundamental differences between the two approaches. Algorithmic syllables demonstrate strong language-family clustering with predominantly language-specific units, while S5-HuBERT units show superior cross-linguistic sharing and weaker family effects. Speaker independence analysis across four experimental phases demonstrates that data-driven methods achieve better acoustic consistency, with the fully data-driven approach reaching near-optimal speaker generalization. These results provide empirical guidance for implementing syllable-based semantic units in multilingual text-to-speech systems for resource-scarce languages.

Abstract: The performance of deep learning models is affected by not only data quantity but also data quality. Data pruning is a process by which practitioners can reduce the size of a dataset by only keeping the most important training data points, thereby achieving similar test set performance. We empirically investigate two state-of-the-art data pruning methods under noisy and noiseless conditions and show that these methods fail in the presence of significant label noise. We highlight that the success of data pruning is distinctly affected by three factors: redundancy in the dataset, the presence of problematic samples, and interdependence between samples. We perform a detailed investigation on commonly used benchmark classification datasets and neural network architectures. We find that our observations are consistent across data distributions and training protocols.

Abstract: Transfer learning with models pretrained on ImageNet has become a standard practice in computer vision. Transfer learning refers to fine-tuning pretrained weights of a neural network on a downstream task, typically unrelated to ImageNet. However, pretrained weights that are very close to zero may yield insignificant gradients, and fail to adapt to the downstream task. This hinders the ability of the model to train effectively, and is commonly referred to as loss of neural plasticity. Loss of plasticity may prevent the model from fully adapting to the target domain, especially when the downstream dataset is atypical in nature. While this issue has been widely explored in continual learning, it remains relatively understudied in the context of transfer learning. In this work, we propose the use of a targeted weight re-initialization strategy to restore neural plasticity prior to fine-tuning. Our experiments show that both convolutional neural networks (CNNs) and vision transformers (ViTs) benefit from this approach, yielding higher test accuracy with faster convergence on several image classification benchmarks. Our method introduces negligible computational overhead and is compatible with common transfer learning pipelines.

Abstract: Language model compression through knowledge distillation has emerged as a promising approach for deploying large language models in resource-constrained environments. However, existing methods often struggle to maintain performance when distilling multilingual models, especially for low-resource languages. In this paper, we present a novel hybrid distillation approach that combines traditional knowledge distillation with a simplified attention matching mechanism, specifically designed for multilingual contexts. Our method introduces an extremely compact student model architecture, significantly smaller than conventional multilingual models. We evaluate our approach on five African languages: Kinyarwanda, Swahili, Hausa, Igbo, and Yoruba. The distilled student model AfroXLMR-Comet successfully captures both the output distribution and internal attention patterns of a larger teacher model (AfroXLMR-Large) while reducing the model size by over 85\%. Experimental results demonstrate that our hybrid approach achieves competitive performance compared to the teacher model, maintaining an accuracy within 85% of the original model's performance while requiring substantially fewer computational resources. Our work provides a practical framework for deploying efficient multilingual models in resource-constrained environments, particularly benefiting applications involving African languages.

Abstract: Cross-lingual language models enable the transfer of linguistic knowledge across languages which is beneficial for languages without large text corpora. However, these multi-lingual models typically exhibit worse performance for low-resource or typologically distant target languages compared to their high-resource counterparts. Various post-hoc methods such as projection-based alignment and adapter modules have been proposed to reduce the performance disadvantage between source and target languages. However, the use of code-switched texts remains limited. Prior research has only focused on finetuning or adapting massively multilingual models with static word substitutions, instead of incorporating code-switching directly into pretraining model design. In this work, we propose an approach that integrates code-switching during pretraining for masked language models. Instead of applying word substitutions after pretraining, we introduce a multiview probabilistic translation strategy throughout the training process. Specifically, we use IBM word alignment models to identify candidate translations for each token and sample replacements based on their translation likelihoods, while applying substitutions only to unmasked tokens. The aim is to expose the model to cross-lingual ambiguity and encourage more robust cross-lingual representations. Our results on a diverse set of eight language pairs show that our code-switching approach improves zero-shot cross-lingual natural language understanding performances for all eight languages relative to bilingual baselines. Furthermore, we also achieved performance gains on downstream named entity recognition tasks in most languages when incorporating our code-switched pretraining approach.

Abstract: Social media—particularly X (formerly Twitter)—has be- come a critical platform for political discourse. It shapes public opin- ion, influences voter behaviour, and provides real-time insight into con- tentious issues. Xenophobia, defined as the “fear of strangers”, is a po- larising topic in South Africa, especially during election seasons. This paper analyses South African Twitter data from the 2016 and 2021 municipal elections and the 2019 and 2024 national elections, focusing on Xenophobia-related discourse. We develop a novel machine learning model to identify xenophobic tweets despite the removal of explicit hate speech by platform moderation. Using a labelled dataset of xenophobic tweets, we fine-tuned a transformer-based classifier that achieves over 95% F1-score in distinguishing xenophobic content. We then analyse the prevalence of xenophobic narratives over time, the peaks around elec- tion dates, and the user accounts most active in propagating xenopho- bia. Our results reveal thousands of Xenophobic tweets, peaking sharply during election periods, and show that over half of the top 20 xenophobia- spreading accounts appear affiliated with political figures or parties. We discuss implications for social media policy, election integrity, and com- munity cohesion. We also address ethical considerations such as data privacy, anonymisation of users, and bias. This work contributes a frame- work for identifying harmful election-related discourse and insights for mitigating the impact of xenophobic narratives on social media.

Abstract: Federated Learning (FL) is a distributed learning paradigm which entails the training of Machine Learning (ML) models across mul- tiple computing devices, while keeping the training data local to the devices. One of the key challenges of FL is heterogeneity of both the com- puting devices and data. This challenge might ultimately lead to the FL model instability, slow convergence, and performance degradation. This work introduces Adaptive FedProx, a new FedProx algorithm extension that dynamically modifies its proximal regularisation term in response to real-time heterogeneity detection. In order to direct adaptive regulari- sation, we present the Heterogeneity-Aware Performance Index (HAPI), a metric that measures the difference between local and global models. We uncover an important trade-off through extensive experiments on CIFAR-10 across Independent and Identically Distributed (IID), mild non-IID, and strong non-IID scenarios: Adaptive FedProx exhibits su- perior robustness to data heterogeneity, despite a 1.6% performance de- crease in homogeneous (IID) settings when compared to FedAvg (87.27% vs. 88.87%, p < 0.001). When moving from IID to strong non-IID data, Adaptive FedProx shows 3.5% better robustness with a performance drop of only 22.8% versus FedAvg’s 26.3%, and it achieves 67.36% accuracy in strong non-IID scenarios compared to FedAvg’s 65.35%. These results imply that, at the expense of a minor drop in performance in homo- geneous environments, adaptive regularisation techniques can improve federated learning’s resistance to heterogeneous data distributions.

Abstract: A third of the food produced globally and in South Africa is lost or wasted annually. Fresh fruits and vegetables (FFVs) contribute 44% of South Africa’s wastage, with temperature abuse as the main cause due to their high sensitivity and perishability. Real-time cold chain management with predictive analytics is necessary to proactively control parameters and minimise temperature breaks. While Machine Learning (ML) can predict temperature, cloud deployment causes latency, bandwidth demands, and internet dependency, hindering real-time operations. Fog computing mitigates this by localising ML predictions. This study investigates fog-based Deep Neural Networks for cold room temperature prediction in FFV cold chains, using IoT data from a South African apple cold room laboratory. SimPy, MinMax scaling, 75%/25% train/test splitting and a sequence length of 2 were used. The fog-based LSTM-GRU ensemble outperformed the cloud-based model with R2 [0.9081,0.9245] vs [0.5477, 0.5992], MAE [0.9999, 1.6353] vs [4.2983, 4.6029], MSE [10.0084,12.0685] vs [50.6492, 56.4827], latency [0.0840, 0.0847] vs [0.3375, 0.3378], and Pearson R [0.9533, 0.9618] vs [0.7670, 0.7930] (95% confidence). Paired t-test and Wilcoxon test confirm a significant difference, with fog always superior. High accuracy and low latency make fog ideal for real-time cold chains, to reduce the wastage of FFVs and associated resources, economic losses, and emissions while increasing affordability, profitability, and food security. Both had high data utilisation of 99.8%, guaranteeing analysis for transmitted data. Future work will add fog nodes, compare models, and implement asynchronous sensor fusion and temperature break and cause predictions.

Abstract: Driver distraction remains a significant contributor to road accidents; however, existing deep-learning detectors either sacrifice accuracy for speed on resource-constrained hardware, or lose generality when confronted with unseen data. This work presents an end-to-end single- stage model that, in one forward pass, jointly identifies the driver, links the driver’s body parts, recognises nearby in-cabin objects, and deter- mines whether the driver is distracted. By embedding spatial and se- mantic relationships directly into its output vector, the model avoids slow post-processing, and significantly reduces false alarms caused by objects that merely appear near the driver. Evaluated on five public datasets and an additional real-world collection that were not used for training, the proposed detector boosts the mean F1-score by 0.11 (≈20% relative) over a lightweight baseline while maintaining 39 frames per second on an NVIDIA Jetson Xavier edge device—more than three times faster than a comparable two-stage pipeline. These results demonstrate a driver-distraction detector that simultaneously achieves cross-dataset ro- bustness, real-time performance, and efficient deployment on low-power hardware.

Abstract: Current state-of-the-art Wildlife classification models are trained under the closed world setting. When exposed to unknown classes they remain overconfident in their predictions. Open-Set Recognition (OSR) aims to classify known classes while rejecting unknown samples. Several OSR methods have been proposed to model the closed-set distribution by observing the feature, logit or softmax probability space. A significant drawback of many existing approaches is the requirement to retrain the pre-trained classification model with the OSR-specific strategy. This study contributes a post-processing OSR method that measures the agreement between the models features and predicted logits. We propose a probability distribution based on an input’s distance to its Nearest-Class Mean (NCM). This NCM-based distribution is then compared with the softmax probabilities from the logit space to measure agreement between the NCM and the classification head. Our proposed strategy ranks within the top three on two evaluated datasets showing consistent performance across the two datasets. In contrast, current SOTA methods excel on a single dataset. We achieve an AUROC of 93.41 and 95.35 for African and Swedish animals. The code will be released publicly upon acceptance of this paper.

Abstract: Deep learning models, for instance, segmentation have achieved remarkable success, yet their deployment in specialised domains like ecological monitoring is constrained by the prohibitive cost of acquiring high-quality polygonal annotations. This annotation dependency creates a fundamental bottleneck, limiting both the scalability and adaptability of these models in real-world conservation scenarios. This paper introduces a data-centric workflow that addresses this challenge through an iterative, multi-stage active learning strategy enhanced with foundation models. The methodology integrates CLIP-based diversity sampling as an acquisition function with a semi-automated annotation pipeline that combines YOLOv8 detection proposals, human-in-the-loop verification, and SAM2-prompted segmentation refinement. A progressive training strategy using YOLOv11-seg with quality-controlled pseudo-labelling iteratively expands the training dataset while maintaining annotation quality standards. Validated on African Penguin monitoring using Open Images V7 data and independent SANParks field data, experimental results demonstrate that CLIP-based diversity sampling achieves mAP$_{50}$ of 0.74 with 400 training samples compared to 0.69 for random sampling, with cross-domain generalisation achieving mAP$_{50}$ of 0.81 on independent field data. The framework reduces annotation requirements while providing a practical solution for deploying instance segmentation in data-scarce domains.

Abstract: Pretraining language models for low-resource languages poses significant challenges due to scarce and poor-quality data, a lack of comprehensive evaluation benchmarks, and often limited computational resources. Research on compute-optimal language modeling typically focuses on scaling up decoder language models efficiently for high-resource languages. While some studies have investigated the down-scaling of encoder language models for low-resource languages, they often prioritize optimizing for computational constraints rather than pretraining text volume constraints. We address this research gap by analyzing the scaling behaviors of encoder language models which use the Replace Token Detection (RTD) and Masked Language Modeling (MLM) objectives under limited pretraining text volumes. By downsampling three different high-resource languages (English, French, Korean) and two low-resource languages (Xhosa and Swahili), we simulate varying degrees of data scarcity and evaluate downstream performance using established benchmarks such as the GLUE benchmark for English, FLUE for French, KLUE for Korean, and MasakhaNEWS for Xhosa and Swahili. Our findings demonstrate that optimal MLM accuracy scales logarithmically with increasing pretraining text volume across these diverse languages. Additionally, our results show that RTD models consistently outperform MLM models in low-resource scenarios, achieving superior downstream performance with pretraining text volumes smaller than 1000MB for downsampled high-resource languages. However, we find that RTD performs worse than MLM for Xhosa and Swahili. We also find that dynamic masking significantly improves MLM accuracy in these settings. Furthermore, our results show that smaller models are more effective for smaller pretraining text volumes, highlighting the importance of adjusting model size according data availability in order to maximize performance and efficiency.

Abstract: Target re-identification (re-ID) is critical for surveillance and autonomous systems but faces challenges in balancing accuracy and com- putational efficiency. This paper presents a novel framework that inte- grates Mixture-of-Experts (MoE) and Knowledge Distillation (KD) to effectively leverage pre-trained foundation models. By dynamically com- bining expert models and distilling their collective knowledge into a com- pact student architecture, the proposed MoE-KD framework achieves state-of-the-art performance. Experimental evaluation demonstrates sig- nificant improvements in accuracy and computational efficiency com- pared to conventional approaches. Extensive ablation studies highlight the contributions of MoE and KD, showing improved cross-domain gen- eralisation and reduced computational overhead. The results indicate that the MoE-KD framework is well-suited for real-world deployment, significantly advancing re-ID systems.

Abstract: Optimization has been an important factor and topic of interest in training deep learning models, yet less attention has been given to how we select the optimizers we use to train these models. Hence, there is need to dive deeper into how we select the optimizers we use for training and the metrics that determine this selection. In this work, we compare the performance of 10 different optimizers in training a simple Multi-layer Perceptron model using a heart disease dataset from Kaggle. We set up a consistent training paradigm and evaluate the optimizers based on metrics such as convergence speed and stability. We also include some other Machine Learning Evaluation metrics such as AUC, Precision, and Recall, which are central metrics to classification problems. Our results show that there are trade-offs between convergence speed and stability, as optimizers like Adagrad and Adadelta, which are more stable, took longer time to converge. Across all our metrics, we choose RMSProp to be the most effective optimizer for this heart disease prediction task because it offered a balanced performance across key metrics. It achieved a precision of 0.765, recall of 0.827, and an AUC of 0.841, along with faster training time. However, it was not the most stable. We recommend that in less compute constrained environment, this method of choosing optimizers through a thorough evaluation should be adopted to increase the scientific nature and performance in training deep learning.

Abstract: Cryptocurrency-fueled financial crimes in Southern Africa have reached crisis levels, with darknet markets and peer-to-peer exchanges laundering over $ 1.2 billion annual-ly (Chainalysis, 2024: chainalysis.com/reports). Current Anti-Money Laundering (AML) tools fail to address the unique convergence of South Africa’s regulated crypto hubs and Zimbabwe’s informal USDT black markets, missing 63% of cross-border fraud (Elliptic, 2023: elliptic.co/resources). We present DarkTrace-SA, a graph ma-chine learning framework that combines Temporal General Neural Networks (TGNN) for dynamic money flow analysis, behavioural clustering optimised for low-data re-gimes, and forensic attribution modules linking on-chain transactions to real-world entities. Evaluated on 4, 200 labeled darknet transactions from publicly available datasets(Elliptic: kaggle.com/ellipticco), our model achieved 93.7% precision (22.4 improvements over benchmarks), 38.6% fewer positives, and 72.3% accuracy in iden-tifying cashout points, enabling deployable and scalable fraud detection for regulators such as the South Africa Reserve Bank(SARB) and Reserve Bank of Zimbabwe (RBZ). This study establishes a new benchmark for AI-driven forensics in emerging markets, producing policy-ready outputs that comply with Financial Action Task Force (FATF) standards. Future work will extend these privacy coins (Monero) and DeFi rug pulls.

Abstract: Due to the lack of quality data for low-resource Bantu languages, significant challenges are presented in text classification and other practical implementations. In this paper, we introduce an advanced model combining Language- Independent Data Augmentation (LiDA) with Multi-Head Attention based weighted embeddings to selectively enhance critical data points and improve text classification performance. This integration allows us to create robust data augmentation strategies that are effective across various linguistic contexts, ensuring that our model can handle the unique syntactic and semantic features of Bantu languages. This approach not only addresses the data scarcity issue but also sets a foundation for future research in low-resource language processing and classification tasks.

Abstract: The increasing availability of choices on online platforms has resulted in a rise in user expectations for personalized content across social media, entertainment, and e-commerce websites. Recommender systems, using machine learning, analyze user preferences to provide intelligent recommendations, help users manage information overload. Collaborative filtering, particularly based on latent factor models like matrix factorization, has proven successful in personalized recommenda- tionsbutencounterschallengessuchasdatasparsityandissueswithnon- linear feature representation. The use of deep neural networks for per- sonalized recommendations has garnered interest to the advancements in deep learning methodologies. To improve recommendation performance and user experience, research is focusing on applying deep learning con- cepts to address existing challenges. This paper proposes an end-to-end deep learning method to address these challenges. The fundamental idea of this method involves transforming the dense feature vector produced by embedding methods into two end-to-end deep neural network designs. Subsequently, it independently learns a low-dimensional representation and non-linear abstraction of the data. In addition, the method incorpo- rates a deep learning structure into the output layer of the networks to predict rating scores. In four real-world datasets, this proposed technique surpassed state-of-the-art models in terms of performance.

Abstract: Social media has become a primary channel for communication and public discourse, yet its high-frequency and emotionally charged nature contributes to the spread of misinformation and elevated levels of social stress. Existing misinformation detection models do not account for the psychological and social toll of this environment. This study proposes a machine learning-based Social Stress Indicator (SSI) to detect and quantify stress signals in social media conversations. Using a synthetic dataset of annotated microblogs labelled by social scientists for social stress levels, four machine learning models-Logistic Regression, Random Forest, Naive Bayes, and K-Nearest Neighbour-were trained using TF-IDF embeddings. Sentiment features from VADER and RoBERTa were also integrated. Results showed consistently high performance across models, with accuracy above 99.4% and macro F1-scores exceeding 0.994. These findings demonstrate that machine learning models can reliably detect social stress from text data. However, the overperformance may reflect dataset homogeneity, necessitating further testing on real-world, diverse social media data to confirm generalisability. Future work should expand across platforms and contexts to validate this approach for stress-aware infodemic response.

Abstract: We study reward design for prosocial behaviour in multi-agent reinforcement learning (MARL) under a social-dilemma setting. In an apple-harvesting environment where agents must both harvest apples and clean a river to sustain yields, we compare four schemes: independent rewards, fully shared team rewards, fractional team rewards, and an inequality-aware diminishing-returns shaping. Independent rewards induce over-harvesting and under-provision of cleaning, collapsing group productivity. Shared rewards achieve high total output via division of labour but produce large disparities in individual payoffs. Fractional team rewards fail to correct incentives: because self-harvesting strictly dominates spillovers from teammates, agents continue to prioritise harvesting over cleaning. In contrast, the inequality-aware shaping performs strongly: agents harvest greedily at first, then cross a utility threshold and reallocate effort to cleaning, which raises group productivity and the marginal scope for individual gains. This scheme nearly matches the shared-reward baseline in total apples while maintaining low inequality. Our contributions are: (i) an empirical demonstration of the efficiency--equity trade-off inherent in standard reward extremes, (ii) framing inequality as a controllable dimension of MARL objectives, and (iii) practical reward designs that encourage reciprocity and public-good provision. We discuss implications for training prosocial agents in larger-scale systems, including language agents, robots, and mixed human--AI teams.

Abstract: Cross-lingual transfer learning is crucial for developing NLP technologies for low-resource African languages, yet the optimal modeling strategy remains an open question. This paper presents a comprehensive benchmark for cross-lingual transfer from Kinyarwanda to Kirundi, two closely related Bantu languages. We evaluate the performance of traditional monolingual architectures (BiGRU, CNN, Char-CNN) against three distinct multilingual transformer models: the global mBERT, the Africa-centric AfriBERT, and the language-family specific BantuBERTa. Our evaluation covers both zero-shot and fine-tuned sen-timent classification, and critically, we measure the degree of catastrophic for getting on the source language after fine-tuning. Our results demonstrate that the regionally-focused AfriBERT achieves the highest cross-lingual accuracy (88.3%) after fine-tuning. Furthermore, we find that large-scale pre-training is essential for robustness; mBERT and AfriBERT exhibit minimal forgetting, while BantuBERTa and the traditional models suffer a severe performance degradation. This study highlights the superiority of regionally-focused multilingual models for transfer between related African languages and establishes catastrophic for getting as a critical evaluation metric for such tasks.

Abstract: Code-switching has become the modus operandi of internet communi- cation in many communities such as South Africans, who are domestically multi- lingual. This phenomenon has made processing textual data increasingly complex due to non-standard ways of writing, spontaneous word replacements, and other challenges. Pre-trained multilingual models have shown elevated text processing capabilities in various similar downstream tasks such as language identification, dialect detection, and language family discrimination. In this study, we exten- sively investigate the use of pre-trained multilingual models - AfroXLMR, and Serengeti for code-switching detection on five South African languages: Sesotho, Setswana, IsiZulu, IsiXhosa, and English, with English used interchangeably with the other four languages, including various transfer learning settings. Addition- ally, we explore the modelling of known switching pairs within a dataset through explicit cross-lingual embeddings extracted using projection models: VecMap, Muse, and Canonical Correlation Analyses (CCA). The resulting cross-lingual embeddings are used to replace the embedding layer of a pre-trained multilingual model without additional training. Concretely, our results show that performance gains can be realized by closing the representational gap between the languages of the code-switched dataset with known codes, using cross-lingual representations. Moreover, expanding code-switched datasets with datasets of closely related lan- guages improves code-switching classification, especially in cases with minimal training examples

Abstract: The integration of AI into cybersecurity is essential for addressing complex and evolving threats. However, much of the existing AI implementation research emphasises either technical or social dimensions, neglecting their socio-technical interdependence. This study addresses this gap by identifying the critical success factors (CSFs) for AI-driven cybersecurity implementation through a socio-technical lens. Using an interpretive case study of a South African state-owned entity, the research draws on thematic analysis of in-depth interviews with technical staff and end-users. Findings reveal that successful implementation depends on the interplay between technical and social elements. Key technical CSFs include data quality, scalability, automation, and efficiency, while social CSFs encompass change acceptance, top management support, user awareness, ethical considerations, human oversight, and usability. Crucially, the study confirms that neither technical nor social factors alone are sufficient and that effective implementation depends on their interdependence. By applying a socio-technical perspective, the research offers a more balanced understanding of AI-driven cybersecurity and presents a framework to support practitioners in implementing socially integrated, technically robust solutions. Future research should further examine how human-AI collaboration can be socio-technically integrated to enhance trust, ensure ethical compliance, and improve the operational reliability of AI-enabled cybersecurity systems within organisational settings.

Abstract: As Large Language Models (LLMs) increasingly support and automate business-critical workflows, the need for robust evaluation frameworks becomes paramount. This paper proposes a system-level testing approach designed to assess the performance and reliability of LLM-based applications integrated into enterprise processes. Moving beyond model-centric benchmarks, the framework adopts principles from software engineering, including black-box and dynamic testing, to evaluate real-world outcomes in retrieval-augmented generation (RAG) systems. It features modular metrics such as context precision, hallucination detection, business tonality alignment, and functional correctness, many harnessing the LLM-as-a-Judge methodology. Empirical evaluations in four use cases demonstrate how this approach enables organizations to validate not only the accuracy of the system but also its business relevance.

Abstract: The emergence of generative artificial intelligence (GenAI) and its applicability within higher education institutions (HEIs) has gained momentum worldwide. GenAI tools have caused a paradigm shift in education, including students’ research activities. Despite various studies being conducted on GenAI tools in education, most research remains concentrated on developed countries, with limited attention to how these technologies are perceived in developing nations. Therefore, this study explores the usage and perceptions of GenAI tools among post-graduate students enrolled for Postgraduate diplomas, Honours and Master’s degrees at a private HEI in South Africa. Using a mixed-methods approach, the study surveyed 75 students to understand their usage and perceptions of GenAI tools for supporting research activities. The findings reveal that almost three-quarters of the students use GenAI tools, particularly ChatGPT, and have a positive attitude towards the use of GenAI tools to support their research activities. The high usage of GenAI tools is attributed to their capability to generate research ideas, summarise articles, and simplify difficult concepts. Over a quarter of the surveyed students do not use GenAI tools due to concerns about plagiarism, bias, privacy and the potential to impair cognitive development. 78% of the students are familiar with the institution's policy on GenAI. On this basis, it is then recommended that HEIs should assert balance in integrating GenAI tools to support students’ research activities. HEIs should further invest in universal frameworks that will serve as guidelines for using GenAI tools in scholarly activities without violating academic and ethical integrity.

Abstract: The environmental impacts of large language models (LLMs) often remain invisible in business adoption. This paper presents an awareness framework to support the sustainable selection of LLMs, developed using a design science research approach within the marketing department of a major European engineering and technology company. Addressing the lack of transparency and emissions data from LLM providers, the prototype calculates electricity use, carbon emissions, and material impacts of inference tasks and visualises them in an interactive dashboard. Evaluation workshops with stakeholders from marketing, sustainability, and AI strategy confirmed the framework’s potential to foster awareness, support sustainable decision-making, and align AI use with corporate environmental goals and the UN Sustainable Development Goals (SDG). The framework is transferable to other business contexts.

Abstract: This study synthesizes insights from 14 in-depth expert interviews to examine the state of artificial intelligence (AI) in Africa across six thematic areas: infrastructure development, governance frameworks, cultural preservation, linguistic equity, the startup ecosystem, and youth empowerment. It aims to identify key challenges, opportunities, user sentiments, and context-specific strategies for responsible AI growth on the continent. Verbatim transcripts were analyzed through thematic coding and sentiment analysis. Key themes and subthemes were mapped, and sentiment trends were quantified using AI-assisted coding tools and manual validation. Visualizations (heatmaps, sentiment distributions, co-occurrence matrices, etc.) were generated to contextualize the findings. The analysis revealed uneven AI readiness in Africa. Less than 1% of the global AI computing capacity is based in Africa, with significant infrastructure gaps despite pockets of progress. Governance is characterized by a strong developmental vision (for example, the African Union’s Continental AI Strategy) but suffers from fragmented policies and lagging implementation. Cultural and linguistic diversity efforts have yielded positive sentiment—for example, community-driven projects for language preservation—whereas discussions on infrastructure and policy skewed more negative. Private AI investment remains concentrated in a few tech hubs, leaving other regions behind. Nonetheless, interviewees highlighted emerging opportunities: public-private partnerships for infrastructure, grassroots AI innovation in local languages, and ambitious youth-focused AI capacity-building programs.

Abstract: This study examines the factors that influence the use of autonomous vehicles (AVs) on traditional road infrastructure in the Western Cape, with a focus on AV compatibility with existing infrastructure. The Diffusion of Innovations Theory serves as a conceptual framework for assessing the potential and challenges of AV implementation. A qualitative approach was employed through semi-structured interviews with experts from the Departments of Mobility, Infrastructure, and Environmental Affairs in the Western Cape. A thematic analysis of the interviews indicates that the Western Cape’s paved roads are largely suitable for AV implementation. However, some adaptation would be necessary in rural areas with gravel roads and limited connectivity. Findings indicate that environmental and economic factors, such as funding limitations and public purchasing preferences, negatively influence AV adoption. Political advantage, however, may positively influence the diffusion process. Surprisingly, the study suggests that AVs may need to adapt to existing road infrastructure rather than vice versa, which contrasts with the established literature on AV implementation in developed regions.

Abstract: Interpreting the complex and multifactorial risk factors driving Cholera outbreaks remains a critical challenge for public health, particularly across diverse environmental and socio-economic contexts. This paper presents an integrated agentic framework that combines explainable machine learning (ML), statistical analysis, and a language model-powered question-answering system to support Cholera risk interpretation and public health decision-making. Using a multi-country dataset spanning 2000–2025, the framework applies three interpretable ML models—Explainable Boosting Machines (EBM), Natural Gradient Boosting (NGBoost), and TabNet—to predict Cholera incidence based on environmental, socio-economic, and infrastructural variables. In parallel, statistical methods including Pearson and Spearman correlation, and multivariate linear regression are used to validate and quantify associations between predictors and disease outcomes. A LangChain-powered agent, implemented with LangGraph, is integrated into the system to interpret model outputs, analyse tabular results, and generate expert-like responses to natural language queries. The agent draws evidence from multiple CSV-based analyses—including feature importance scores, correlation matrices, regression coefficients, and model performance comparisons to provide grounded, interpretable answers and policy recommendations. A Streamlit interface enables interactive exploration of Cholera risk factors by researchers, health professionals, and policy stakeholders. Results show strong agreement among models on key predictors, such as rainfall frequency, stagnant water presence, and open defecation, with statistically significant relationships confirmed through regression analysis. The EBM model achieved the lowest RMSE (0.421), indicating superior predictive performance. This work demonstrates how explainable AI and LLM agents can be combined into a transparent, interpretable, and actionable framework for public health analytics, offering valuable insights into data-driven disease prevention strategies.

Abstract: Generative artificial intelligence (AI) applications have enhanced democratization of infor-mation to the extent that industry professionals can automate routine tasks, gain insight from complex data, and execute tasks more efficiently through generation of text, image, and audio content. Although these applications augment human capabilities, there are concerns about veracity of AI prompting, which results in hallucinations that have dire consequences on clinical workflow of healthcare professionals. The impact of prompting patterns on op-timization of clinical workflows at points-of-care remains nascent with limited evidence especially in overburdened healthcare sectors of sub-Saharan Africa (SSA). The review explores existing literature on how generative AI prompt engineering optimizes clinical workflow of healthcare professionals by adopting Arksey and O'Malley five-stage scoping review framework to analyze peer-reviewed publications. A comprehensive search strategy was conducted in scholarly databases, including PubMed, IEEE Xplore, and Google Scholar between 2019 and 2025. The study highlights AI prompt engineering strategies, how prompting affects clinical and administrative activities, and how limitations of generative AI prompting can be addressed. Evidence of generative AI prompt engineering are limited in SSA while the Global North and China are the most dominant regions in the discourse. Con-sultations, clinical decision support, record summary and documentation, research and pre-scription recommendations are activities in which AI prompting is perceived as most signifi-cant. To conclude, this study provides insights for health managers, healthcare profession-als, data scientists, ethicists, health IT experts, human-computer interaction practitioners, and researchers on standardizing the integration of generative AI utilization at points-of-care.

Abstract: Visual inspection remains a common approach for assessing composite insu-lators, with unmanned aerial vehicles (UAVs) increasingly preferred due to their efficiency and reduced error rates. Recent developments have integrat-ed artificial intelligence (AI) algorithms directly into UAV hardware to ena-ble faster processing; however, such systems require optimized models ow-ing to limited onboard computing resources. The recently introduced YOLOv10-N model, which offers greater efficiency compared to its prede-cessors, demonstrates potential for detecting insulator defects on resource-constrained UAV platforms. This study evaluates the effectiveness of YOLOv10-N for this application.

Abstract: The deployment of Large Language Models (LLMs) in enterprise en- vironments faces significant challenges including high computational costs, data privacy concerns, and network dependencies. This paper presents a comprehen- sive framework for deploying Small Language Models (SLMs) with fewer than 7 billion parameters on edge devices, demonstrating how agentic architectures can overcome the inherent limitations of reduced model capacity. We introduce three key contributions: (1) a novel multi-agent benchmarking framework that employs role-based evaluation (Proctor, Student, Grader) to reduce bias and provide robust performance metrics; (2) a three-phase task planning pipeline specifically designed for SLMs that decomposes planning into subtask identification, dependency rea- soning, and schema-constrained generation, significantly improving reliability com- pared to single-pass approaches; and (3) real-world implementations including a multilingual translation system achieving 3-4x latency improvements over cloud services while maintaining data sovereignty. Our extensive evaluation across mul- tiple SLMs (3B-7B parameters) demonstrates that models like Phi-4 can achieve CEFR C1-level translation quality for high-resource languages and strong summa- rization performance (0.883 G-Eval score) while running on commodity hardware. Through practical deployments using WebLLM for browser-based inference and lo- cal hosting solutions, we show that SLMs can effectively serve enterprise needs in privacy-sensitive, bandwidth-constrained, or air-gapped environments. Our findings indicate that carefully orchestrated SLM-based systems represent a viable alternative to cloud-based LLMs for organizations prioritizing data sovereignty, cost efficiency, and edge deployment capabilities.

Abstract: Teenagers face serious difficulties as a result of mobile bullying, which calls for creative solutions. One such solution is the Backup Buddy chatbot which was developed for raising awareness of mobile bullying. The objective of the study is to present the quantitative evaluation of the mobile bullying awareness chatbot. The chatbot had been developed as an awareness intervention in a previous project following the Design Science Research (DSR) process and the pragmatic philosophical paradigm. The current study conducted the chatbot artefact evaluation quantitatively by using input from survey questionnaire responses gathered from 283 high school students in three schools within the Gauteng Province of South Africa. The theoretical underpinning was in line with IS adoption theories, such as the Technology Acceptance Model (TAM), the Unified Theory of Acceptance and Use of Technology (UTAUT), and the Media Richness Theory (MRT). Statistical analysis was performed on the quantitative survey data. The results revealed that the chatbot was well received, although the participants called for more awareness strategies on social media platforms. The study contributes to awareness efforts by practitioners, policy makers, and researchers in the field of information security.

Abstract: This paper explores how artificial intelligence can be harnessed to boost and supercharge scoping reviews. To demonstrate its power (comprehensive-ness), its efficiency (cost effectiveness) and affordability (cost effective-ness), we explore the use of AI in obstetrics and gynaecology. We carried out: (1) a traditional scoping review, and (2) a traditional scoping review boosted by artificial intelligence tools. We also provide an overview of the extensive and varied literature related to our topic of interest: the use of AI in obstetrics and gynaecology. We compare and contrast the outcomes of the different options. We conclude that the use of AI tools can enrich and ex-tend the scope of scoping reviews (a so-called Hybrid Review), but only if the prompts are carefully and thoughtfully crafted.

Abstract: This study investigates how cultural proximity affects the ability to detect AI-generated fake news by comparing South African participants with those from other nationalities. As Large Language Models (LLMs) like ChatGPT increasingly enable the creation of sophisticated fake news, understanding human detection capabilities becomes crucial, particularly across different cultural contexts. We conducted a survey where 45 participants (24 South Africans, 21 from other nationalities) evaluated 10 authentic South African news articles and 10 AI-generated fake versions. Results revealed an asymmetric pattern: South Africans demonstrated superior performance in detecting authentic news about their country (61% accuracy) compared to other participants (48%), but performed worse at identifying fake news (36% vs. 44%). Our analysis further reveals that South Africans relied more on content knowledge and contextual understanding when judging credibility, while participants from other countries emphasized formal linguistic features such as grammar and structure. Overall detection accuracy was similar between groups (49% vs. 46%), suggesting that cultural proximity confers advantages for verifying authentic information while potentially creating vulnerabilities when evaluating fabricated content. These insights contribute to understanding cross-cultural dimensions of misinformation detection and inform strategies for combating AI-generated fake news in increasingly globalized information ecosystems where content crosses cultural and geographical boundaries.

Abstract: Large Language Models (LLMs) are increasingly deployed in interactive public information systems, yet often fail to adapt their outputs to diverse user expectations. This paper presents a persona-aware alignment pipeline that fine-tunes LLMs using synthetic dialogue data to improve the stylistic alignment and communicative relevance of generated responses. The approach is applied in the context of RepoChat, a dialogue system developed with Nagra, the Swiss agency for the disposal of radioactive waste. RepoChat provides public access to complex technical and regulatory information and must respond appropriately to diverse audiences, including citizens, journalists, politicians, and subject matter experts. We introduce a modular training process combining retrieval-augmented answer generation with persona-specific rewriting, supervised fine-tuning, and preference-based optimization. To evaluate the system, we employ both automated assessments using LLM-as-a-Judge methods and qualitative user feedback through interviews. Findings show that the fine-tuned model demonstrates moderate and inconsistent improvements in tone, clarity, and user-perceived alignment—particularly for non-expert audiences. However, limitations remain in handling emotional nuance and maintaining consistency across multi-turn dialogue. This work contributes a reproducible alignment pipeline for persona-sensitive LLM deployment and highlights the value of synthetic training data in human-centred, high-stakes communication domains.

Abstract: The transition from traditional deep learning models to foundational models represents a paradigm shift in artificial intelligence, particularly in medical applications with development of several unimodal and multimodal base models including MedGemma, Biomedclip, DINO group of models and MedImageInsight. However, this paradigm shift from narrow to generalist models is not without challenges. Foundation models need large datasets and computational requirements for their initial training hence can only be trained by large corporations. Due to their generative nature, they are prone to hallucinations where the models make up an output - which increases risk in medicine with potential for harm. The fundamental question facing healthcare deployers is not whether a model can pass standardized tests like USMLE exams, but whether it can be safely and effectively deployed in specific clinical contexts. Current evaluation frameworks that heavily rely on multiple-choice questions (MCQs) fail to capture real-world deployment scenarios and human interaction patterns. We convened a two day datathon on July 16th - 17th 2025. Our chosen theme was a single question “Can we deploy MedGemma for chest Xray reporting in South Africa?” rather than "How well does this model perform on standardized tests?" We sought a practical approach for radiologists to validate and incrementally implement a foundation model into their workflow at a pace that they are comfortable with. Rather than risking the clinical inaccuracies of attempting a full diagnosis, we focused on more accessible tasks that the model could automate. We introduce a three-pronged evaluation framework designed to bridge the gap between model capabilities and deployment readiness that provides grounding to overcome hallucinations by using a 1) RAG based approach, 2) Support for personalised deployment and 3) Hierarchical decision making with several Go/No-Go thresholds. These features are then aggregated into a single score similar to the threshold levels of deterministic models that simplify deployment, and continuous real world surveillance evaluation dashboards.

Abstract: Yuval Noah Harari, a historian and author of popular sci ence books, explores Information in his latest work Nexus, focusing on how it connects humans through networks. He presents two perspectives — the naïve and the more complete view — on Information as a rep resentation of reality balanced with social Order. This paper examines the analytics workflow as an example of Information within these views, then introduces an Agentic AI perspective where Agents generate mul tiple, stochastic truths. Finally, we consider resulting issues of Order, including regulation, alignment, and perception of accuracy.

Abstract: Generative Artificial Intelligence (Gen AI) is an emerging technology that has the potential to influence cybersecurity within organisations. The transformative capabilities of Gen AI allow it to adapt to the rapidly evolving cyber space that organisations operate in. This study represents primary efforts to understand Gen AI’s capabilities for cybersecurity and its uses within organisations. The study was conducted through a literature analysis to investigate the discourse on Gen AI usage for cybersecurity based on published academic papers. The findings point to a discourse focused on the capabilities of Gen AI for cybersecurity within organisations, such as threat detection, and automated testing to name a few. The study also highlights the benefits associated with the usage of Gen AI for cybersecurity as well as the concerns that arise within organisations. Further research will be important to ascertain employee perceptions of the usage of Gen AI for cybersecurity within organisations.

Abstract: Algorithms often reinforce societal biases and stereotypes. This is especially concerning for minorities, who are disproportionately impacted by it, thereby threatening their further marginalization. Data fundamentalists frame this issue of algorithmic bias as stemming from data bias, indicated by the underrepresentation of some groups (minorities) in the datasets. Consequently, measures adopted to resolve algorithmic bias have been data-focused. A relatively recent data-focused measure adopted to address this issue is the deployment of what I term artificially generated minorities (AGMs)—synthetic data used to increase the representation of underrepresented groups (minorities) in datasets used for training algorithms. Data fundamentalists make two central claims about AGMs, which I term: the representation claim, which holds that AGMs are representative of minorities, and the normative intervention claim, which holds that AGMs are a solution to algorithmic bias and, by extension, societal bias. In this paper, I argue that AGMs do not meet these claims. First, I demonstrate that AGMs do not capture the experience of historic and systemic oppression, which defines minority status. Hence, I contend that they do not meaningfully represent minorities. Second, I demonstrate that while AGMs facilitate the realization of the futuristic component of an adequate normative intervention, they undermine the reparative component. Thus, I contend that AGMs do not adequately address algorithmic bias and, by extension, societal bias. Finally, I briefly highlight that the failure of AGMs to meet these claims indicates that a data-focused framing of algorithmic bias is overly simplistic and does not account for all the complexities involved in the issue of algorithmic bias and its correction.

Abstract: This paper discusses ELSA (Ethical, Legal, and Social Aspects) as an emerging methodology for transdisciplinary AI research, based on a 1,5 year-long case study in Lombardijen, the Netherlands. This neighbourhood is paradoxically under-resourced—meaning historically neglected and stigmatised as a 'problem district'—yet over-researched, meaning scrutinised by countless academics who engage in what has been called ‘drive-by’ research—driving by, extracting data, and disappearing, often without community benefits. Though understandable, the community's ensuing alienation from governmental and academic institutions is particularly problematic for socially responsible AI development because citizen’s contextual knowledge becomes overlooked in public deliberation on AI's impacts. This raises our research question: How can citizens in low-trust neighbourhoods be meaningfully and reciprocally engaged in transdisciplinary AI research, and what does an ELSA approach offer in this regard? After tracing ELSA's development as a methodology and presenting our case study (engaging, amongst others, residents, students, municipality and housing corporations), this paper details our experiences in Lombardijen respectively from ethical, legal, social, and technical perspectives—showing not only how we addressed community distrust and academic alienation, but also how we developed transformative AI research aligning with community interests. Ultimately, these findings demonstrate how ELSA can support more accountable, reciprocal, and meaningful public engagement with AI in transdisciplinary fashion. Yet, importantly, by also candidly discussing the limitations of our work—ultimately arguing for more situated responsibility as a precondition for these types of research to succeed—we contribute insights and practical guidance for future work in transdisciplinary AI research, broadly speaking.

Abstract: Artificial intelligence (AI) offers numerous benefits, including increased automation levels, but it can also cause harm to individuals and communities, which raises concerns. The field of human-centered AI (HCAI) was created to address human involvement and the consideration of human rights and values involved in the development of AI systems. Document analysis of governance frameworks was used as a research approach to identify components that contribute to human-centred AI solutions. Four authoritative bodies were selected as the principal sources of AI-related principles, standards, and guidelines, based on their recognized authority, global relevance, and comprehensive regulatory scope. Thirty-seven human-centred components were extracted and classified into five classification categories: human-centred values and ethics, user experience and human interaction, data and model governance, technical robustness and system performance, and AI system capabilities and design considerations. The identified components can be used to develop AI solutions that are human-centred and uphold the legal integrity, fundamental freedoms, and principles of democratic governance. Including these components in a formal development methodology can assist in developing AI solutions that are human-centered, free from bias, beneficial to humans, supportive of human capabilities, and aligned with ethical, transparent, reliable, trustworthy, and explainable principles.

Abstract: The increasing presence of artificial agents (AAs) in everyday life has foregrounded a profound paradox in human-machine interaction: while people instinctively engage with AAs as if they possess consciousness and emotions, they often intellectually deny these very attributions. This contradiction has led to a dominant research paradigm, which this paper terms the "deception strategy," that dismisses such empathic behavior as a fallacy rooted in anthropomorphic illusions. This paper argues that this view is flawed, as it relies on (1) a rigid ontological divide between humans and machines and a (2) simplistic distinction between appearance and reality. Drawing from phenomenologically inspired enactivism, this paper proposes an alternative framework that reinterprets these empathic responses not as deceptive projections, but as the constitutive elements of a new form of social relation. By introducing the concept of the "otheroid," this paper offers a novel category for artificial entities that are experienced as neither fully human nor purely mechanical, thereby embracing the dynamic, reciprocal, and embodied nature of our interactions in an increasingly technologically mediated world.

Abstract: Granting legal status to Artificial intelligence (AI) could address much of the uncertainty that currently surrounds accountability and liability of AI systems. As AI lacks sentience and expression of its own will, it cannot be deemed a natural person however juristic or legal personality may possibly be imposed on it. Consideration of the South African common law requirements for juristic personality suggest that AI as a non-human entity meets the prerequisites and consequently should be bestowed with juristic personality in the same manner in which it is bestowed on companies and similar corporations. The research argues that AI meets the South African common law requirements for bestowing juristic personality on non-human entities and postulates that on this basis the South African Companies Act may serve as the primary guiding legislation on which basis a specific AI statute and associated regulations may be drafted and entrenched. By means of a doctrinal desktop study this research considers arguments for and against bestowing juristic personality on AI systems and argues in favour of juristic personality being bestowed on AI discussing the potential benefits that may arise from same. The research by no means attempts to provide a definitive answer to the questions posed but seeks to stimulate academic discussions and contribute to the existing body of knowledge exploring the legal status of AI from a uniquely South African perspective.

Abstract: While the systemic biases and discriminatory outcomes of AI are increasing-ly well-documented, this article argues that these dynamics also facilitate a subtler, less examined form of neo-colonialism that disproportionately af-fects Africa due to enduring historical and structural power imbalances in the global digital ecosystem. A pertinent question, however, is whether unethical biases are deliberate or unintended. This paper examines the intrinsic algo-rithmic biases that disproportionately marginalize communities in Africa, critically analyzing whether these computational systems replicate and rein-force entrenched historical power asymmetries, and whether that is done in-tentionally, through the lens of postcolonial theory. Through an analysis of ownership structures and capital flows in AI development and deployment, this paper demonstrates that digital discrimination is not an unintended con-sequence but rather a deliberate outcome of technological ecosystems that reflect and reproduce colonial-era hierarchies of power and knowledge.

Abstract: The growing popularity and adoption of generative artificial intelligence (AI) has given rise to the new and worrying phenomenon of deepfake por-nography. A type of synthetic media generated by AI in which a person’s face is superimposed onto existing pornographic material, creating a new, hyper-realistic version in which a new person is depicted performing in pornographic content they did not actually participate in. Classified as an act of technology-facilitated gender-based violence (TFGBV), the broad function of deepfake pornography is to sexually humiliate women with the purpose of reinforcing women’s domination and subjugation at the hands of men.

Abstract: This empirical study investigates how well can Artificial Intelligence (AI) draft a South African High Court judgment using OpenAI’s ChatGPT-4o large language model and case data from De Jager v Netcare Limited and Others 2025 JDR 0793 (GP)—an informational privacy and data protection dispute. Drawing from local and international experiments, the study develops a value-driven prompting framework and distills a rubric for evaluating the quality of AI-driven judgments. Twenty sequential prompts were developed to generate an artificial judgment which was then compared with the human decision with respect to structure, factual accuracy, use of legal authority, legal reasoning and clarity. ChatGPT-4o led across all five metrics, though formatting inadequacies were a weakness. This study did not assess the substantive legal correctness of either judgment. However, it is interesting that both judgments reached nearly identical outcomes on the main substantive legal issues, viz privacy, statutory compliance (Protection of Personal Information Act 4 of 2013) and admissibility of surveillance data. This may be attributable to their common reliance on similar submissions, suggesting that persuasive pleadings can channel both human and machine reasoning to similar ends. These findings support the potential of AI as a judicial aid but reinforce the need for robust human oversight.

Abstract: Agency law, a bedrock of commercial legal systems globally, facilitates commerce by enabling individuals (agents) to act on behalf of others (principals). This doctrine encompasses a wide range of applications, from employer-employee relationships to complex supply chains and is a tool used to achieve commercial convenience. Yet, with the increasing deployment of artificial intelligence (AI) agents, particularly machine learning systems acting autonomously in data-driven environments, the traditional application of the doctrine of agency faces unprecedented pressure. In South Africa, commercial law frameworks have yet to adequately account for the functional implications of AI systems acting in place of human agents. As these technologies become entrenched in private-sector decision-making from smart contracts and procurement systems to customer service bots and autonomous trading, they reveal structural ambiguities in the legal rules regarding agency.

Abstract: As AI becomes integrated into everyday life, it also becomes a powerful force in shaping whose voices are heard, whose knowledge is validated, and whose experienc-es are recognized as legitimate. While AI technologies are often portrayed as objective tools for decision-making, their development and deployment are deeply entangled with the social and political structures in which they operate. This makes them not only susceptible to contributing to epistemic injustice but capable of amplifying it at scale. In this paper we draw on the recent theoretical innovation of Rebera et al who mobilize the idea of hermeneutic harm and use it to identify potential sites of epistem-ic injustice. Hermeneutic harm is “emotional and psychological pain caused by a pro-longed inability to make sense of an event (or events) in one’s life” [1]. While not specific to AI, hermeneutic harms can be exacerbated by AI-systems. Hermeneutic harm can arise in cases where agent-like AI-systems cause a harm, which under nor-mal circumstances would be explainable by appeal to the intentions, motivations, or actions of human agents. Drawing on the rich analytic resources of the epistemic injustice literature, we identify three forms of AI-induced hermeneutic harm: obscured self-understanding, asymmetrical testimonial obstruction, and bureaucratic displacement. We then high-light how each of these hermeneutic harms are trigged by specific kinds of epistemic injustice and can illustrate instances of wrongful exclusion from meaning and sense-making practices.

Abstract: With the increasing encroachment of Big Data and AI governing critical domains, the legal perspective has resorted to suggesting novel human rights to protect individuals and communities and ensure meaningful human control, accountability, and responsible AI governance. This paper explores one of such rights: the right to a human decision-maker. After a brief outline of the growing advocacy for this right, primarily following the work of Onora O’Neill, I sketch the various challenges of positing a new positive human, i.e., universal, right, including arduous allocation and questions on feasibility. Not all hope is lost, however, as I argue that framing a human decision-maker as a novel right is in fact unnecessary. Extending an argument on moral relationships by Christine Korsgaard, I argue that the structure of the claim right and correlative duty relationships necessitates human decision-makers. Here we see that an essential reciprocity of rights that requires both role-reversibility and constitutive symmetry between the right-holders and decision-makers, or duty-bearers. Importantly, while this extends Korsgaard’s account of the moral law, I show why this also applies to democratic positive law. This argument about the essential moral and legal value of human intervention therefore sidesteps well-known difficulties of positive right allocation and shows that a human decision-maker is not merely a novel, desirable pursuit of responsible AI development and implementation, but is a constitutive feature of rights tout court.

Abstract: Many real-world decisions follow rules that hold in general but allow exceptions, such as “birds usually fly, unless they are penguins.” Most interpretable classifiers struggle to capture this pattern, leading to explanations that feel less aligned with human reasoning. This paper introduces the Defeasible Horn Classifier with Exceptions (DHCE), a symbolic model that makes this structure explicit: each default rule is paired with its linked exception, so predictions can be explained step by step without post-hoc tools. DHCE is learned using Answer Set Programming, which searches for globally optimal rule sets while balancing accuracy and simplicity. The resulting models consist of ranked Horn rules that provide full traceability: users can see both why a decision applies and why it may be overridden. On three standard benchmarks, DHCE achieved comparable test error on two datasets and slightly lower error on one. For context, we align our experimental setup with the interpretable short Boolean formulas classifier of Jaakkola et al. and observe similar error magnitudes while providing an explicit “normally…unless…” narrative. By making overrides explicit, DHCE delivers accuracy alongside step-by-step explanations, which is promising for domains where knowing when a rule no longer applies is as important as the prediction itself.