Research Papers

With the widespread integration of AI in everyday and critical technologies, it seems inevitable to witness increasing instances of failure in AI systems. In such cases, there arises a need for technical investigations that produce legally acceptable and scientifically indisputable findings and conclusions on the causes of such failures. Inspired by the domain of cyber forensics, this paper introduces the need for the establishment of AI Forensics as a new discipline under AI safety. Furthermore, we propose a taxonomy of the subfields under this discipline, and present a discussion on the foundational challenges that lay ahead of this new research area.

Modeling AGI Safety Frameworks with Causal Influence Diagrams

Tom Everitt, Ramana Kumar, Victoria Krakovna +1

Proposals for safe AGI systems are typically made at the level of frameworks, specifying how the components of the proposed system should be trained and interact with each other. In this paper, we model and compare the most promising AGI safety frameworks using causal influence diagrams. The diagrams show the optimization objective and causal assumptions of the framework. The unified representation permits easy comparison of frameworks and their assumptions. We hope that the diagrams will serve as an accessible and visual introduction to the main AGI safety frameworks.

Taxonomy of Pathways to Dangerous AI

Roman V. Yampolskiy

In order to properly handle a dangerous Artificially Intelligent (AI) system it is important to understand how the system came to be in such a state. In popular culture (science fiction movies/books) AIs/Robots became self-aware and as a result rebel against humanity and decide to destroy it. While it is one possible scenario, it is probably the least likely path to appearance of dangerous AI. In this work, we survey, classify and analyze a number of circumstances, which might lead to arrival of malicious AI. To the best of our knowledge, this is the first attempt to systematically classify types of pathways leading to malevolent AI. Previous relevant work either surveyed specific goals/meta-rules which might lead to malevolent behavior in AIs (Özkural, 2014) or reviewed specific undesirable behaviors AGIs can exhibit at different stages of its development (Alexey Turchin, July 10 2015, July 10, 2015).

In this study, we conduct a resume-screening experiment (N=528) where people collaborate with simulated AI models exhibiting race-based preferences (bias) to evaluate candidates for 16 high and low status occupations. Simulated AI bias approximates factual and counterfactual estimates of racial bias in real-world AI systems. We investigate people's preferences for White, Black, Hispanic, and Asian candidates (represented through names and affinity groups on quality-controlled resumes) across 1,526 scenarios and measure their unconscious associations between race and status using implicit association tests (IATs), which predict discriminatory hiring decisions but have not been investigated in human-AI collaboration. When making decisions without AI or with AI that exhibits no race-based preferences, people select all candidates at equal rates. However, when interacting with AI favoring a particular group, people also favor those candidates up to 90% of the time, indicating a significant behavioral shift. The likelihood of selecting candidates whose identities do not align with common race-status stereotypes can increase by 13% if people complete an IAT before conducting resume screening. Finally, even if people think AI recommendations are low quality or not important, their decisions are still vulnerable to AI bias under certain circumstances. This work has implications for people's autonomy in AI-HITL scenarios, AI and work, design and evaluation of AI hiring systems, and strategies for mitigating bias in collaborative decision-making tasks. In particular, organizational and regulatory policy should acknowledge the complex nature of AI-HITL decision making when implementing these systems, educating people who use them, and determining which are subject to oversight.

Responsible AI (rAI) guidance increasingly promotes stakeholder involvement (SHI) during AI development. At the same time, SHI is already common in commercial software development, but with potentially different foci. This study clarifies the extent to which established SHI practices are able to contribute to rAI efforts as well as potential disconnects -- essential insights to inform and tailor future interventions that further shift industry practice towards rAI efforts. First, we analysed 56 rAI guidance documents to identify why SHI is recommended (i.e. its expected benefits for rAI) and uncovered goals such as redistributing power, improving socio-technical understandings, anticipating risks, and enhancing public oversight. To understand why and how SHI is currently practised in commercial settings, we then conducted an online survey (n=130) and semi-structured interviews (n=10) with AI practitioners. Our findings reveal that SHI in practice is primarily driven by commercial priorities (e.g. customer value, compliance) and several factors currently discourage more rAI-aligned SHI practices. This suggests that established SHI practices are largely not contributing to rAI efforts. To address this disconnect, we propose interventions and research opportunities to advance rAI development in practice.

Explainable AI: XAI-Guided Context-Aware Data Augmentation

Melkamu Abay Mersha, Mesay Gemeda Yigezu, Atnafu Lambebo Tonja +4

Explainable AI (XAI) has emerged as a powerful tool for improving the performance of AI models, going beyond providing model transparency and interpretability. The scarcity of labeled data remains a fundamental challenge in developing robust and generalizable AI models, particularly for low-resource languages. Conventional data augmentation techniques introduce noise, cause semantic drift, disrupt contextual coherence, lack control, and lead to overfitting. To address these challenges, we propose XAI-Guided Context-Aware Data Augmentation. This novel framework leverages XAI techniques to modify less critical features while selectively preserving most task-relevant features. Our approach integrates an iterative feedback loop, which refines augmented data over multiple augmentation cycles based on explainability-driven insights and the model performance gain. Our experimental results demonstrate that XAI-SR-BT and XAI-PR-BT improve the accuracy of models on hate speech and sentiment analysis tasks by 6.6% and 8.1%, respectively, compared to the baseline, using the Amharic dataset with the XLM-R model. XAI-SR-BT and XAI-PR-BT outperform existing augmentation techniques by 4.8% and 5%, respectively, on the same dataset and model. Overall, XAI-SR-BT and XAI-PR-BT consistently outperform both baseline and conventional augmentation techniques across all tasks and models. This study provides a more controlled, interpretable, and context-aware solution to data augmentation, addressing critical limitations of existing augmentation techniques and offering a new paradigm shift for leveraging XAI techniques to enhance AI model training.

Will AI Tell Lies to Save Sick Children? Litmus-Testing AI Values Prioritization with AIRiskDilemmas

Yu Ying Chiu, Zhilin Wang, Sharan Maiya +4

Detecting AI risks becomes more challenging as stronger models emerge and find novel methods such as Alignment Faking to circumvent these detection attempts. Inspired by how risky behaviors in humans (i.e., illegal activities that may hurt others) are sometimes guided by strongly-held values, we believe that identifying values within AI models can be an early warning system for AI's risky behaviors. We create LitmusValues, an evaluation pipeline to reveal AI models' priorities on a range of AI value classes. Then, we collect AIRiskDilemmas, a diverse collection of dilemmas that pit values against one another in scenarios relevant to AI safety risks such as Power Seeking. By measuring an AI model's value prioritization using its aggregate choices, we obtain a self-consistent set of predicted value priorities that uncover potential risks. We show that values in LitmusValues (including seemingly innocuous ones like Care) can predict for both seen risky behaviors in AIRiskDilemmas and unseen risky behaviors in HarmBench.

As frontier AI systems advance toward transformative capabilities, we need a parallel transformation in how we measure and evaluate these systems to ensure safety and inform governance. While benchmarks have been the primary method for estimating model capabilities, they often fail to establish true upper bounds or predict deployment behavior. This literature review consolidates the rapidly evolving field of AI safety evaluations, proposing a systematic taxonomy around three dimensions: what properties we measure, how we measure them, and how these measurements integrate into frameworks. We show how evaluations go beyond benchmarks by measuring what models can do when pushed to the limit (capabilities), the behavioral tendencies exhibited by default (propensities), and whether our safety measures remain effective even when faced with subversive adversarial AI (control). These properties are measured through behavioral techniques like scaffolding, red teaming and supervised fine-tuning, alongside internal techniques such as representation analysis and mechanistic interpretability. We provide deeper explanations of some safety-critical capabilities like cybersecurity exploitation, deception, autonomous replication, and situational awareness, alongside concerning propensities like power-seeking and scheming. The review explores how these evaluation methods integrate into governance frameworks to translate results into concrete development decisions. We also highlight challenges to safety evaluations - proving absence of capabilities, potential model sandbagging, and incentives for "safetywashing" - while identifying promising research directions. By synthesizing scattered resources, this literature review aims to provide a central reference point for understanding AI safety evaluations.

Chronic Kidney Disease (CKD) is a major global health issue which is affecting million people around the world and with increasing rate of mortality. Mitigation of progression of CKD and better patient outcomes requires early detection. Nevertheless, limitations lie in traditional diagnostic methods, especially in resource constrained settings. This study proposes an advanced machine learning approach to enhance CKD detection by evaluating four models: Random Forest (RF), Multi-Layer Perceptron (MLP), Logistic Regression (LR), and a fine-tuned CatBoost algorithm. Specifically, among these, the fine-tuned CatBoost model demonstrated the best overall performance having an accuracy of 98.75%, an AUC of 0.9993 and a Kappa score of 97.35% of the studies. The proposed CatBoost model has used a nature inspired algorithm such as Simulated Annealing to select the most important features, Cuckoo Search to adjust outliers and grid search to fine tune its settings in such a way to achieve improved prediction accuracy. Features significance is explained by SHAP-a well-known XAI technique-for gaining transparency in the decision-making process of proposed model and bring up trust in diagnostic systems. Using SHAP, the significant clinical features were identified as specific gravity, serum creatinine, albumin, hemoglobin, and diabetes mellitus. The potential of advanced machine learning techniques in CKD detection is shown in this research, particularly for low income and middle-income healthcare settings where prompt and correct diagnoses are vital. This study seeks to provide a highly accurate, interpretable, and efficient diagnostic tool to add to efforts for early intervention and improved healthcare outcomes for all CKD patients.

Achieving balanced alignment of large language models (LLMs) in terms of Helpfulness, Honesty, and Harmlessness (3H optimization) constitutes a cornerstone of responsible AI. Existing methods like data mixture strategies face limitations, including heavy reliance on expert knowledge and conflicting optimization signals. While model merging offers parameter-level conflict-resolution strategies through integrating specialized models' parameters, its potential for 3H optimization remains underexplored. This paper systematically compares the effectiveness of model merging and data mixture methods in constructing 3H-aligned LLMs for the first time, revealing previously overlooked collaborative and conflict relationships among the 3H dimensions and discussing the advantages and drawbacks of data mixture (\textit{data-level}) and model merging (\textit{parameter-level}) methods in mitigating the conflict for balanced 3H optimization. Specially, we propose a novel \textbf{R}eweighting \textbf{E}nhanced task \textbf{S}ingular \textbf{M}erging method, \textbf{RESM}, through outlier weighting and sparsity-aware rank selection strategies to address the challenges of preference noise accumulation and layer sparsity adaptation inherent in 3H-aligned LLM merging. Extensive evaluations can verify the effectiveness and robustness of RESM compared to previous data mixture (2\%-5\% gain) and model merging (1\%-3\% gain) methods in achieving balanced LLM alignment. We release our models through \href{https://huggingface.co/Jinluan}{3H\_Merging} for further investigations.

Recent research suggests that it may be possible to build conscious AI systems now or in the near future. Conscious AI systems would arguably deserve moral consideration, and it may be the case that large numbers of conscious systems could be created and caused to suffer. Furthermore, AI systems or AI-generated characters may increasingly give the impression of being conscious, leading to debate about their moral status. Organisations involved in AI research must establish principles and policies to guide research and deployment choices and public communication concerning consciousness. Even if an organisation chooses not to study AI consciousness as such, it will still need policies in place, as those developing advanced AI systems risk inadvertently creating conscious entities. Responsible research and deployment practices are essential to address this possibility. We propose five principles for responsible research and argue that research organisations should make voluntary, public commitments to principles on these lines. Our principles concern research objectives and procedures, knowledge sharing and public communications.

Recognizing daily activities with unobtrusive sensors in smart environments enables various healthcare applications. Monitoring how subjects perform activities at home and their changes over time can reveal early symptoms of health issues, such as cognitive decline. Most approaches in this field use deep learning models, which are often seen as black boxes mapping sensor data to activities. However, non-expert users like clinicians need to trust and understand these models' outputs. Thus, eXplainable AI (XAI) methods for Human Activity Recognition have emerged to provide intuitive natural language explanations from these models. Different XAI methods generate different explanations, and their effectiveness is typically evaluated through user surveys, that are often challenging in terms of costs and fairness. This paper proposes an automatic evaluation method using Large Language Models (LLMs) to identify, in a pool of candidates, the best XAI approach for non-expert users. Our preliminary results suggest that LLM evaluation aligns with user surveys.