Research Papers

Text-based online counselling scales across geographical and stigma barriers, yet faces practitioner shortages, lacks non-verbal cues and suffers inconsistent quality assurance. Whilst artificial intelligence offers promising solutions, its use in mental health counselling raises distinct ethical challenges. This paper analyses three AI implementation approaches - autonomous counsellor bots, AI training simulators and counsellor-facing augmentation tools. Drawing on professional codes, regulatory frameworks and scholarly literature, we identify four ethical principles - privacy, fairness, autonomy and accountability - and demonstrate their distinct manifestations across implementation approaches. Textual constraints may enable AI integration whilst requiring attention to implementation-specific hazards. This conceptual paper sensitises developers, researchers and practitioners to navigate AI-enhanced counselling ethics whilst preserving human values central to mental health support.

COMPASS: A Framework for Evaluating Organization-Specific Policy Alignment in LLMs

Dasol Choi, DongGeon Lee, Brigitta Jesica Kartono +6

As large language models are deployed in high-stakes enterprise applications, from healthcare to finance, ensuring adherence to organization-specific policies has become essential. Yet existing safety evaluations focus exclusively on universal harms. We present COMPASS (Company/Organization Policy Alignment Assessment), the first systematic framework for evaluating whether LLMs comply with organizational allowlist and denylist policies. We apply COMPASS to eight diverse industry scenarios, generating and validating 5,920 queries that test both routine compliance and adversarial robustness through strategically designed edge cases. Evaluating seven state-of-the-art models, we uncover a fundamental asymmetry: models reliably handle legitimate requests (>95% accuracy) but catastrophically fail at enforcing prohibitions, refusing only 13-40% of adversarial denylist violations. These results demonstrate that current LLMs lack the robustness required for policy-critical deployments, establishing COMPASS as an essential evaluation framework for organizational AI safety.

Explainable AI (XAI) methods like SHAP and LIME produce numerical feature attributions that remain inaccessible to non expert users. Prior work has shown that Large Language Models (LLMs) can transform these outputs into natural language explanations (NLEs), but it remains unclear which factors contribute to high-quality explanations. We present a systematic factorial study investigating how Forecasting model choice, XAI method, LLM selection, and prompting strategy affect NLE quality. Our design spans four models (XGBoost (XGB), Random Forest (RF), Multilayer Perceptron (MLP), and SARIMAX - comparing black-box Machine-Learning (ML) against classical time-series approaches), three XAI conditions (SHAP, LIME, and a no-XAI baseline), three LLMs (GPT-4o, Llama-3-8B, DeepSeek-R1), and eight prompting strategies. Using G-Eval, an LLM-as-a-judge evaluation method, with dual LLM judges and four evaluation criteria, we evaluate 660 explanations for time-series forecasting. Our results suggest that: (1) XAI provides only small improvements over no-XAI baselines, and only for expert audiences; (2) LLM choice dominates all other factors, with DeepSeek-R1 outperforming GPT-4o and Llama-3; (3) we observe an interpretability paradox: in our setting, SARIMAX yielded lower NLE quality than ML models despite higher prediction accuracy; (4) zero-shot prompting is competitive with self-consistency at 7-times lower cost; and (5) chain-of-thought hurts rather than helps.

As Large Language Model (LLM) agents are increasingly tasked with high-stakes autonomous decision-making, the transparency of their reasoning processes has become a critical safety concern. While \textit{Chain-of-Thought} (CoT) prompting allows agents to generate human-readable reasoning traces, it remains unclear whether these traces are \textbf{faithful} generative drivers of the model's output or merely \textbf{post-hoc rationalizations}. We introduce \textbf{Project Ariadne}, a novel XAI framework that utilizes Structural Causal Models (SCMs) and counterfactual logic to audit the causal integrity of agentic reasoning. Unlike existing interpretability methods that rely on surface-level textual similarity, Project Ariadne performs \textbf{hard interventions} ($do$-calculus) on intermediate reasoning nodes -- systematically inverting logic, negating premises, and reversing factual claims -- to measure the \textbf{Causal Sensitivity} ($φ$) of the terminal answer. Our empirical evaluation of state-of-the-art models reveals a persistent \textit{Faithfulness Gap}. We define and detect a widespread failure mode termed \textbf{Causal Decoupling}, where agents exhibit a violation density ($ρ$) of up to $0.77$ in factual and scientific domains. In these instances, agents arrive at identical conclusions despite contradictory internal logic, proving that their reasoning traces function as "Reasoning Theater" while decision-making is governed by latent parametric priors. Our findings suggest that current agentic architectures are inherently prone to unfaithful explanation, and we propose the Ariadne Score as a new benchmark for aligning stated logic with model action.

Online, people often recount their experiences turning to conversational AI agents (e.g., ChatGPT, Claude, Copilot) for mental health support -- going so far as to replace their therapists. These anecdotes suggest that AI agents have great potential to offer accessible mental health support. However, it's unclear how to meet this potential in extreme mental health crisis use cases. In this work, we explore the first-person experience of turning to a conversational AI agent in a mental health crisis. From a testimonial survey (n = 53) of lived experiences, we find that people use AI agents to fill the in-between spaces of human support; they turn to AI due to lack of access to mental health professionals or fears of burdening others. At the same time, our interviews with mental health experts (n = 16) suggest that human-human connection is an essential positive action when managing a mental health crisis. Using the stages of change model, our results suggest that a responsible AI crisis intervention is one that increases the user's preparedness to take a positive action while de-escalating any intended negative action. We discuss the implications of designing conversational AI agents as bridges towards human-human connection rather than ends in themselves.

The proliferation of Large Language Models (LLMs) has intensified concerns about manipulative or deceptive behaviors that can undermine user autonomy, trust, and well-being. Existing safety benchmarks predominantly rely on coarse binary labels and fail to capture the nuanced psychological and social mechanisms constituting manipulation. We introduce \textbf{DarkPatterns-LLM}, a comprehensive benchmark dataset and diagnostic framework for fine-grained assessment of manipulative content in LLM outputs across seven harm categories: Legal/Power, Psychological, Emotional, Physical, Autonomy, Economic, and Societal Harm. Our framework implements a four-layer analytical pipeline comprising Multi-Granular Detection (MGD), Multi-Scale Intent Analysis (MSIAN), Threat Harmonization Protocol (THP), and Deep Contextual Risk Alignment (DCRA). The dataset contains 401 meticulously curated examples with instruction-response pairs and expert annotations. Through evaluation of state-of-the-art models including GPT-4, Claude 3.5, and LLaMA-3-70B, we observe significant performance disparities (65.2\%--89.7\%) and consistent weaknesses in detecting autonomy-undermining patterns. DarkPatterns-LLM establishes the first standardized, multi-dimensional benchmark for manipulation detection in LLMs, offering actionable diagnostics toward more trustworthy AI systems.

Machine unlearning aims to remove specific data influences from trained models, a capability essential for adhering to copyright laws and ensuring AI safety. Current unlearning metrics typically measure success by monitoring the model's performance degradation on the specific unlearning dataset ($D_u$). We argue that for Large Language Models (LLMs), this evaluation paradigm is insufficient and potentially misleading. Many real-world uses of unlearning--motivated by copyright or safety--implicitly target not only verbatim content in $D_u$, but also behaviors influenced by the broader generalizations the model derived from it. We demonstrate that LLMs can pass standard unlearning evaluation and appear to have "forgotten" the target knowledge, while simultaneously retaining strong capabilities on content that is semantically adjacent to $D_u$. This phenomenon indicates that erasing exact sentences does not necessarily equate to removing the underlying knowledge. To address this gap, we propose Proximal Surrogate Generation (PSG), an automated stress-testing framework that generates a surrogate dataset, $\tilde{D}_u$. This surrogate set is constructed to be semantically derived from $D_u$ yet sufficiently distinct in embedding space. By comparing unlearning metric scores between $D_u$ and $\tilde{D}_u$, we can stress-test the reliability of the metric itself. Our extensive evaluation across three LLM families (Llama-3-8B, Qwen2.5-7B, and Zephyr-7B-$β$), three distinct datasets, and seven standard metrics reveals widespread inconsistencies. We find that current metrics frequently overestimate unlearning success, failing to detect retained knowledge exposed by our stress-test datasets.

Significant digitalization of financial services in a short period of time has led to an urgent demand to have autonomous, transparent and real-time credit risk decision making systems. The traditional machine learning models are effective in pattern recognition, but do not have the adaptive reasoning, situational awareness, and autonomy needed in modern financial operations. As a proposal, this paper presents an Agentic AI framework, or a system where AI agents view the world of dynamic credit independent of human observers, who then make actions based on their articulable decision-making paths. The research introduces a multi-agent system with reinforcing learning, natural language reasoning, explainable AI modules, and real-time data absorption pipelines as a means of assessing the risk profiles of borrowers with few humans being involved. The processes consist of agent collaboration protocol, risk-scoring engines, interpretability layers, and continuous feedback learning cycles. Findings indicate that decision speed, transparency and responsiveness is better than traditional credit scoring models. Nevertheless, there are still some practical limitations such as risks of model drift, inconsistencies in interpreting high dimensional data and regulatory uncertainties as well as infrastructure limitations in low-resource settings. The suggested system has a high prospective to transform credit analytics and future studies ought to be directed on dynamic regulatory compliance mobilizers, new agent teamwork, adversarial robustness, and large-scale implementation in cross-country credit ecosystems.

Artificial intelligence (AI)-driven augmented reality (AR) systems are becoming increasingly integrated into daily life, and with this growth comes a greater need for explainability in real-time user interactions. Traditional explainable AI (XAI) methods, which often rely on feature-based or example-based explanations, struggle to deliver dynamic, context-specific, personalized, and human-centric insights for everyday AR users. These methods typically address separate explainability dimensions (e.g., when, what, how) with different explanation techniques, resulting in unrealistic and fragmented experiences for seamless AR interactions. To address this challenge, we propose PILAR, a novel framework that leverages a pre-trained large language model (LLM) to generate context-aware, personalized explanations, offering a more intuitive and trustworthy experience in real-time AI-powered AR systems. Unlike traditional methods, which rely on multiple techniques for different aspects of explanation, PILAR employs a unified LLM-based approach that dynamically adapts explanations to the user's needs, fostering greater trust and engagement. We implement the PILAR concept in a real-world AR application (e.g., personalized recipe recommendations), an open-source prototype that integrates real-time object detection, recipe recommendation, and LLM-based personalized explanations of the recommended recipes based on users' dietary preferences. We evaluate the effectiveness of PILAR through a user study with 16 participants performing AR-based recipe recommendation tasks, comparing an LLM-based explanation interface to a traditional template-based one. Results show that the LLM-based interface significantly enhances user performance and experience, with participants completing tasks 40% faster and reporting greater satisfaction, ease of use, and perceived transparency.

The integration of generative Artificial Intelligence into the digital ecosystem necessitates a critical re-evaluation of Indian criminal jurisprudence regarding computational forensics integrity. While algorithmic efficiency enhances evidence extraction, a research gap exists regarding the Digital Personal Data Protection Act, 2023's compatibility with adversarial AI threats, specifically anti-forensics and deepfakes. This study scrutinizes the AI "dual-use" dilemma, functioning as both a cyber-threat vector and forensic automation mechanism, to delineate privacy boundaries in high-stakes investigations. Employing a doctrinal legal methodology, the research synthesizes statutory analysis of the DPDP Act with global ethical frameworks (IEEE, EU) to evaluate regulatory efficacy. Preliminary results indicate that while Machine Learning offers high accuracy in pattern recognition, it introduces vulnerabilities regarding data poisoning and algorithmic bias. Findings highlight a critical tension between the Act's data minimization principles and forensic data retention requirements. Furthermore, the paper identifies that existing legal definitions inadequately encompass AI-driven "tool crimes" and "target crimes." Consequently, the research proposes a "human-centric" forensic model prioritizing explainable AI (XAI) to ensure evidence admissibility. These implications suggest that synchronizing Indian privacy statutes with international forensic standards is imperative to mitigate synthetic media risks, establishing a roadmap for future legislative amendments and technical standardization.

Nowadays, Artificial Intelligence (AI), particularly Machine Learning (ML) and Large Language Models (LLMs), is widely applied across various contexts. However, the corresponding models often operate as black boxes, leading them to unintentionally act unfairly towards different demographic groups. This has led to a growing focus on fairness in AI software recently, alongside the traditional focus on the effectiveness of AI models. Through 26 semi-structured interviews with practitioners from different application domains and with varied backgrounds across 23 countries, we conducted research on fairness requirements in AI from software engineering perspective. Our study assesses the participants' awareness of fairness in AI / ML software and its application within the Software Development Life Cycle (SDLC), from translating fairness concerns into requirements to assessing their arising early in the SDLC. It also examines fairness through the key assessment dimensions of implementation, validation, evaluation, and how it is balanced with trade-offs involving other priorities, such as addressing all the software functionalities and meeting critical delivery deadlines. Findings of our thematic qualitative analysis show that while our participants recognize the aforementioned AI fairness dimensions, practices are inconsistent, and fairness is often deprioritized with noticeable knowledge gaps. This highlights the need for agreement with relevant stakeholders on well-defined, contextually appropriate fairness definitions, the corresponding evaluation metrics, and formalized processes to better integrate fairness into AI/ML projects.

Link-analysis algorithms, such as PageRank, are instrumental in understanding the structural dynamics of networks by evaluating the importance of individual vertices based on their connectivity. Recently, with the rising importance of responsible AI, the question of fairness in link-analysis algorithms has gained traction. In this paper, we present a new approach for incorporating group fairness into the PageRank algorithm by reweighting the transition probabilities in the underlying transition matrix. We formulate the problem of achieving fair PageRank by seeking to minimize the fairness loss, which is the difference between the original group-wise PageRank distribution and a target PageRank distribution. We further define a group-adapted fairness notion, which accounts for group homophily by considering random walks with group-biased restart for each group. Since the fairness loss is non-convex, we propose an efficient projected gradient-descent method for computing locally-optimal edge weights. Unlike earlier approaches, we do not recommend adding new edges to the network, nor do we adjust the restart vector. Instead, we keep the topology of the underlying network unchanged and only modify the relative importance of existing edges. We empirically compare our approach with state-of-the-art baselines and demonstrate the efficacy of our method, where very small changes in the transition matrix lead to significant improvement in the fairness of the PageRank algorithm.