SACAIR 2025 Proceedings

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 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: 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: 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.