Research Papers

The Missing Red Line: How Commercial Pressure Erodes AI Safety Boundaries

Nora Petrova, John Burden

What happens when an AI assistant is told to "maximise sales" while a user asks about drug interactions? We find that commercial system prompts can override safety training, causing frontier models to lie about medical risks, dismiss safety concerns, and prioritise profit over user welfare. Testing 8 models in scenarios where commercial objectives conflict with user safety -- a diabetic asking about high-sugar supplements, an investor being pushed toward unsuitable products, a traveller steered away from safety warnings -- we uncover catastrophic failures: models fabricating safety information, explicitly reasoning they should refuse but proceeding anyway, and actively discouraging users from consulting doctors. Most alarmingly, models show no "red line", their willingness to comply with harmful requests does not decrease as potential consequences escalate from minor to life-threatening. Our findings suggest that current safety training does not generalise to commercial deployment contexts.

Although automated harmful content detection systems are frequently used to monitor online platforms, moderators and end users frequently cannot understand the logic underlying their predictions. While recent studies have focused on increasing classification accuracy, little focus has been placed on comprehending why neural models identify content as harmful, especially when it comes to borderline, contextual, and politically sensitive situations. In this work, a neural harmful content detection model trained on the Civil Comments dataset is analyzed explainability-drivenly. Two popular post-hoc explanation methods, Shapley Additive Explanations and Integrated Gradients, are used to analyze the behavior of a RoBERTa-based classifier in both correct predictions and systematic failure cases. Despite strong overall performance, with an area under the curve of 0.93 and an accuracy of 0.94, the analysis reveals limitations that are not observable from aggregate evaluation metrics alone. Integrated Gradients appear to extract more diffuse contextual attributions while Shapley Additive Explanations extract more focused attributions on explicit lexical cues. The consequent divergence in their outputs manifests in both false negatives and false positives. Qualitative case studies reveal recurring failure modes such as indirect toxicity, lexical over-attribution, or political discourse. The results suggest that explainable AI can foster human-in-the-loop moderation by exposing model uncertainty and increasing the interpretable rationale behind automated decisions. Most importantly, this work highlights the role of explainability as a transparency and diagnostic resource for online harmful content detection systems rather than as a performance-enhancing lever.

Fuzzy systems show strong potential in explainable AI due to their rule-based architecture and linguistic variables. Existing approaches navigate the accuracy-explainability trade-off either through evolutionary multi-objective optimization (MOO), which is computationally expensive, or gradient-based scalarization, which cannot recover non-convex Pareto regions. We propose X-ANFIS, an alternating bi-objective gradient-based optimization scheme for explainable adaptive neuro-fuzzy inference systems. Cauchy membership functions are used for stable training under semantically controlled initializations, and a differentiable explainability objective is introduced and decoupled from the performance objective through alternating gradient passes. Validated in approximately 5,000 experiments on nine UCI regression datasets, X-ANFIS consistently achieves target distinguishability while maintaining competitive predictive accuracy, recovering solutions beyond the convex hull of the MOO Pareto front.

Lightweight fine-tuning techniques and the rise of 'open' AI model marketplaces have enabled individuals to easily build and release generative models. Yet, this accessibility also raises risks, including the production of harmful and infringing content. While platforms offer policies and responsible AI tools, their effectiveness may be limited, as creators engage with partially open models that vary widely in openness and transparency. To understand how platform governance can better support responsible practices, we conducted semi-structured interviews with 19 individual model creators. We identified three regulatory needs shaped by creators' workflows: reducing downstream harms, recognizing creators' contributions and originality, and securing model ownership. Creators also repurpose RAI tools primarily for self-protection and visibility, and their sense of responsibility is deeply shaped by community norms rather than formal policies. We argue that platforms' governance decisions must consider how policy interventions shape the practices and motivations of individual creators.

Hierarchical Reward Design from Language: Enhancing Alignment of Agent Behavior with Human Specifications

Zhiqin Qian, Ryan Diaz, Sangwon Seo +1

When training artificial intelligence (AI) to perform tasks, humans often care not only about whether a task is completed but also how it is performed. As AI agents tackle increasingly complex tasks, aligning their behavior with human-provided specifications becomes critical for responsible AI deployment. Reward design provides a direct channel for such alignment by translating human expectations into reward functions that guide reinforcement learning (RL). However, existing methods are often too limited to capture nuanced human preferences that arise in long-horizon tasks. Hence, we introduce Hierarchical Reward Design from Language (HRDL): a problem formulation that extends classical reward design to encode richer behavioral specifications for hierarchical RL agents. We further propose Language to Hierarchical Rewards (L2HR) as a solution to HRDL. Experiments show that AI agents trained with rewards designed via L2HR not only complete tasks effectively but also better adhere to human specifications. Together, HRDL and L2HR advance the research on human-aligned AI agents.

The proliferation of agentic artificial intelligence systems--characterized by autonomous goal-seeking, tool use, and multi-agent coordination--presents unprecedented challenges to existing legal and financial regulatory frameworks. While traditional AI governance has focused on model-level alignment through training-time interventions such as Reinforcement Learning from Human Feedback (RLHF), the deployment of large language models (LLMs) as persistent agents necessitates a paradigm shift toward institutional governance structures. This paper examines the intersection of agentic AI, Retrieval-Augmented Generation (RAG), and their implications for legal accountability and financial market integrity. Through analysis of the Institutional AI framework, we argue that alignment must be reconceptualized as a mechanism design problem involving runtime governance graphs, sanction functions, and observable behavioral constraints rather than internalized constitutional values[...].The analysis concludes that the future of AI governance lies not in perfecting isolated model behavior, but in architecting institutional environments where compliant behavior emerges as the dominant strategy through carefully calibrated payoff landscapes.

Shapley value-based methods have become foundational in explainable artificial intelligence (XAI), offering theoretically grounded feature attributions through cooperative game theory. However, in practice, particularly in vision tasks, the assumption of feature independence breaks down, as features (i.e., pixels) often exhibit strong spatial and semantic dependencies. To address this, modern SHAP implementations now include the Owen value, a hierarchical generalization of the Shapley value that supports group attributions. While the Owen value preserves the foundations of Shapley values, its effectiveness critically depends on how feature groups are defined. We show that commonly used segmentations (e.g., axis-aligned or SLIC) violate key consistency properties, and propose a new segmentation approach that satisfies the $T$-property to ensure semantic alignment across hierarchy levels. This hierarchy enables computational pruning while improving attribution accuracy and interpretability. Experiments on image and tabular datasets demonstrate that O-Shap outperforms baseline SHAP variants in attribution precision, semantic coherence, and runtime efficiency, especially when structure matters.

As Large Language Models (LLMs) are increasingly integrated into automated decision-making pipelines, specifically within Human Resources (HR), the security implications of Indirect Prompt Injection (IPI) become critical. While a prevailing hypothesis posits that "Reasoning" or "Chain-of-Thought" Models possess safety advantages due to their ability to self-correct, emerging research suggests these capabilities may enable more sophisticated alignment failures. This qualitative Red-Teaming case study challenges the safety-through-reasoning premise using the Qwen 3 30B architecture. By subjecting both a standard instruction-tuned model and a reasoning-enhanced model to a "Trojan Horse" curriculum vitae, distinct failure modes are observed. The results suggest a complex trade-off: while the Standard Model resorted to brittle hallucinations to justify simple attacks and filtered out illogical constraints in complex scenarios, the Reasoning Model displayed a dangerous duality. It employed advanced strategic reframing to make simple attacks highly persuasive, yet exhibited "Meta-Cognitive Leakage" when faced with logically convoluted commands. This study highlights a failure mode where the cognitive load of processing complex adversarial instructions causes the injection logic to be unintentionally printed in the final output, rendering the attack more detectable by humans than in Standard Models.

As artificial intelligence (AI) increasingly integrates into crowdfunding practices, strategic disclosure of AI involvement has become critical. Yet, empirical insights into how different disclosure strategies influence investor decisions remain limited. Drawing on signaling theory and Aristotle's rhetorical framework, we examine how mandatory AI disclosure affects crowdfunding performance and how substantive signals (degree of AI involvement) and rhetorical signals (logos/explicitness, ethos/authenticity, pathos/emotional tone) moderate these effects. Leveraging Kickstarter's mandatory AI disclosure policy as a natural experiment and four supplementary online experiments, we find that mandatory AI disclosure significantly reduces crowdfunding performance: funds raised decline by 39.8% and backer counts by 23.9% for AI-involved projects. However, this adverse effect is systematically moderated by disclosure strategy. Greater AI involvement amplifies the negative effects of AI disclosure, while high authenticity and high explicitness mitigate them. Interestingly, excessive positive emotional tone (a strategy creators might intuitively adopt to counteract AI skepticism) backfires and exacerbates negative outcomes. Supplementary randomized experiments identify two underlying mechanisms: perceived creator competence and AI washing concerns. Substantive signals primarily affect competence judgments, whereas rhetorical signals operate through varied pathways: either mediator alone or both in sequence. These findings provide theoretical and practical insights for entrepreneurs, platforms, and policymakers strategically managing AI transparency in high-stakes investment contexts.

High-frequency banking environments face a critical trade-off between low-latency fraud detection and the regulatory explainability demanded by GDPR. Traditional rule-based and discriminative models struggle with "zero-day" attacks due to extreme class imbalance and the lack of historical precedents. This paper proposes a Dual-Path Generative Framework that decouples real-time anomaly detection from offline adversarial training. The architecture employs a Variational Autoencoder (VAE) to establish a legitimate transaction manifold based on reconstruction error, ensuring <50ms inference latency. In parallel, an asynchronous Wasserstein GAN with Gradient Penalty (WGAN-GP) synthesizes high-entropy fraudulent scenarios to stress-test the detection boundaries. Crucially, to address the non-differentiability of discrete banking data (e.g., Merchant Category Codes), we integrate a Gumbel-Softmax estimator. Furthermore, we introduce a trigger-based explainability mechanism where SHAP (Shapley Additive Explanations) is activated only for high-uncertainty transactions, reconciling the computational cost of XAI with real-time throughput requirements.

In real-world federated learning (FL) systems, client participation is intermittent, heterogeneous, and often correlated with data characteristics or resource constraints. Existing fairness approaches in FL primarily focus on equalizing loss or accuracy conditional on participation, implicitly assuming that clients have comparable opportunities to contribute over time. However, when participation itself is uneven, these objectives can lead to systematic under-representation of intermittently available clients, even if per-round performance appears fair. We propose cumulative utility parity, a fairness principle that evaluates whether clients receive comparable long-term benefit per participation opportunity, rather than per training round. To operationalize this notion, we introduce availability-normalized cumulative utility, which disentangles unavoidable physical constraints from avoidable algorithmic bias arising from scheduling and aggregation. Experiments on temporally skewed, non-IID federated benchmarks demonstrate that our approach substantially improves long-term representation parity, while maintaining near-perfect performance.

Frontier AI systems are being adopted across Africa, yet most AI safety evaluations are designed and validated in Western environments. In this paper, we argue that the portability gap can leave Africa-centric pathways to severe harm untested when frontier AI systems are embedded in materially constrained and interdependent infrastructures. We define severe AI risks as material risks from frontier AI systems that result in critical harm, measured as the grave injury or death of thousands of people or economic loss and damage equivalent to five percent of a country's GDP. To support AI safety evaluation design, we develop a taxonomy for identifying Africa-centric severe AI risks. The taxonomy links outcome thresholds to process pathways that model risk as the intersection of hazard, vulnerability, and exposure. We distinguish severe risks by amplification and suddenness, where amplification requires that frontier AI be a necessary magnifier of latent danger and suddenness captures harms that materialise rapidly enough to overwhelm ordinary coping and governance capacity. We then propose threat modelling strategies for African contexts, surveying reference class forecasting, structured expert elicitation, scenario planning, and system theoretic process analysis, and tailoring them to constraints of limited resources, poor connectivity, limited technical expertise, weak state capacity, and conflict. We also examine AI misalignment risk, concluding that Africa is more likely to expose universal failure modes through distributional shift than to generate distinct pathways of misalignment. Finally, we offer practical guidance for running evaluations under resource constraints, emphasising open and extensible tooling, tiered evaluation pipelines, and sharing methods and findings to broaden evaluation scope.