Research Papers

When Testing AI Tests Us: Safeguarding Mental Health on the Digital Frontlines

Sachin R. Pendse, Darren Gergle, Rachel Kornfield +6

Red-teaming is a core part of the infrastructure that ensures that AI models do not produce harmful content. Unlike past technologies, the black box nature of generative AI systems necessitates a uniquely interactional mode of testing, one in which individuals on red teams actively interact with the system, leveraging natural language to simulate malicious actors and solicit harmful outputs. This interactional labor done by red teams can result in mental health harms that are uniquely tied to the adversarial engagement strategies necessary to effectively red team. The importance of ensuring that generative AI models do not propagate societal or individual harm is widely recognized -- one less visible foundation of end-to-end AI safety is also the protection of the mental health and wellbeing of those who work to keep model outputs safe. In this paper, we argue that the unmet mental health needs of AI red-teamers is a critical workplace safety concern. Through analyzing the unique mental health impacts associated with the labor done by red teams, we propose potential individual and organizational strategies that could be used to meet these needs, and safeguard the mental health of red-teamers. We develop our proposed strategies through drawing parallels between common red-teaming practices and interactional labor common to other professions (including actors, mental health professionals, conflict photographers, and content moderators), describing how individuals and organizations within these professional spaces safeguard their mental health given similar psychological demands. Drawing on these protective practices, we describe how safeguards could be adapted for the distinct mental health challenges experienced by red teaming organizations as they mitigate emerging technological risks on the new digital frontlines.

As machine learning models are increasingly used in educational settings, from detecting at-risk students to predicting student performance, algorithmic bias and its potential impacts on students raise critical concerns about algorithmic fairness. Although group fairness is widely explored in education, works on individual fairness in a causal context are understudied, especially on counterfactual fairness. This paper explores the notion of counterfactual fairness for educational data by conducting counterfactual fairness analysis of machine learning models on benchmark educational datasets. We demonstrate that counterfactual fairness provides meaningful insight into the causality of sensitive attributes and causal-based individual fairness in education.

Beware of "Explanations" of AI

David Martens, Galit Shmueli, Theodoros Evgeniou +14

Understanding the decisions made and actions taken by increasingly complex AI system remains a key challenge. This has led to an expanding field of research in explainable artificial intelligence (XAI), highlighting the potential of explanations to enhance trust, support adoption, and meet regulatory standards. However, the question of what constitutes a "good" explanation is dependent on the goals, stakeholders, and context. At a high level, psychological insights such as the concept of mental model alignment can offer guidance, but success in practice is challenging due to social and technical factors. As a result of this ill-defined nature of the problem, explanations can be of poor quality (e.g. unfaithful, irrelevant, or incoherent), potentially leading to substantial risks. Instead of fostering trust and safety, poorly designed explanations can actually cause harm, including wrong decisions, privacy violations, manipulation, and even reduced AI adoption. Therefore, we caution stakeholders to beware of explanations of AI: while they can be vital, they are not automatically a remedy for transparency or responsible AI adoption, and their misuse or limitations can exacerbate harm. Attention to these caveats can help guide future research to improve the quality and impact of AI explanations.

A moving target in AI-assisted decision-making: Dataset shift, model updating, and the problem of update opacity

Joshua Hatherley

Machine learning (ML) systems are vulnerable to performance decline over time due to dataset shift. To address this problem, experts often suggest that ML systems should be regularly updated to ensure ongoing performance stability. Some scholarly literature has begun to address the epistemic and ethical challenges associated with different updating methodologies. Thus far, however, little attention has been paid to the impact of model updating on the ML-assisted decision-making process itself, particularly in the AI ethics and AI epistemology literatures. This article aims to address this gap in the literature. It argues that model updating introduces a new sub-type of opacity into ML-assisted decision-making -- update opacity -- that occurs when users cannot understand how or why an update has changed the reasoning or behaviour of an ML system. This type of opacity presents a variety of distinctive epistemic and safety concerns that available solutions to the black box problem in ML are largely ill-equipped to address. A variety of alternative strategies may be developed or pursued to address the problem of update opacity more directly, including bi-factual explanations, dynamic model reporting, and update compatibility. However, each of these strategies presents its own risks or carries significant limitations. Further research will be needed to address the epistemic and safety concerns associated with model updating and update opacity going forward.

Understanding uncertainty in Explainable AI (XAI) is crucial for building trust and ensuring reliable decision-making in Machine Learning models. This paper introduces a unified framework for quantifying and interpreting Uncertainty in XAI by defining a general explanation function $e_θ(x, f)$ that captures the propagation of uncertainty from key sources: perturbations in input data and model parameters. By using both analytical and empirical estimates of explanation variance, we provide a systematic means of assessing the impact uncertainty on explanations. We illustrate the approach using a first-order uncertainty propagation as the analytical estimator. In a comprehensive evaluation across heterogeneous datasets, we compare analytical and empirical estimates of uncertainty propagation and evaluate their robustness. Extending previous work on inconsistencies in explanations, our experiments identify XAI methods that do not reliably capture and propagate uncertainty. Our findings underscore the importance of uncertainty-aware explanations in high-stakes applications and offer new insights into the limitations of current XAI methods. The code for the experiments can be found in our repository at https://github.com/TeodorChiaburu/UXAI

As artificial intelligence (AI) systems become increasingly embedded in critical societal functions, the need for robust red teaming methodologies continues to grow. In this forum piece, we examine emerging approaches to automating AI red teaming, with a particular focus on how the application of automated methods affects human-driven efforts. We discuss the role of labor in automated red teaming processes, the benefits and limitations of automation, and its broader implications for AI safety and labor practices. Drawing on existing frameworks and case studies, we argue for a balanced approach that combines human expertise with automated tools to strengthen AI risk assessment. Finally, we highlight key challenges in scaling automated red teaming, including considerations around worker proficiency, agency, and context-awareness.

Advancements in retrieving accessible information have evolved faster in the last few years compared to the decades since the internet's creation. Search engines, like Google, have been the number one way to find relevant data. They have always relied on the user's abilities to find the best information in its billions of links and sources at everybody's fingertips. The advent of large language models (LLMs) has completely transformed the field of information retrieval. The LLMs excel not only at retrieving relevant knowledge but also at summarizing it effectively, making information more accessible and consumable for users. On top of it, the rise of AI Agents has introduced another aspect to information retrieval i.e. dynamic information retrieval which enables the integration of real-time data such as weather forecasts, and financial data with the knowledge base to curate context-aware knowledge. However, despite these advancements the agents remain susceptible to issues of bias and fairness, challenges deeply rooted within the knowledge base and training of LLMs. This study introduces a novel approach to bias-aware knowledge retrieval by leveraging agentic framework and the innovative use of bias detectors as tools to identify and highlight inherent biases in the retrieved content. By empowering users with transparency and awareness, this approach aims to foster more equitable information systems and promote the development of responsible AI.

Recent advances in the data-driven medicine approach, which integrates ethically managed and explainable artificial intelligence into clinical decision support systems (CDSS), are critical to ensure reliable and effective patient care. This paper focuses on comparing novel agent system designs that use modular agents to analyze laboratory results, vital signs, and clinical context, and to predict and validate results. We implement our agent system with the eICU database, including running lab analysis, vitals-only interpreters, and contextual reasoners agents first, then sharing the memory into the integration agent, prediction agent, transparency agent, and a validation agent. Our results suggest that the multi-agent system (MAS) performed better than the single-agent system (SAS) with mortality prediction accuracy (59\%, 56\%) and the mean error for length of stay (LOS)(4.37 days, 5.82 days), respectively. However, the transparency score for the SAS (86.21) is slightly better than the transparency score for MAS (85.5). Finally, this study suggests that our agent-based framework not only improves process transparency and prediction accuracy but also strengthens trustworthy AI-assisted decision support in an intensive care setting.

Exploring Energy Landscapes for Minimal Counterfactual Explanations: Applications in Cybersecurity and Beyond

Spyridon Evangelatos, Eleni Veroni, Vasilis Efthymiou +3

Counterfactual explanations have emerged as a prominent method in Explainable Artificial Intelligence (XAI), providing intuitive and actionable insights into Machine Learning model decisions. In contrast to other traditional feature attribution methods that assess the importance of input variables, counterfactual explanations focus on identifying the minimal changes required to alter a model's prediction, offering a ``what-if'' analysis that is close to human reasoning. In the context of XAI, counterfactuals enhance transparency, trustworthiness and fairness, offering explanations that are not just interpretable but directly applicable in the decision-making processes. In this paper, we present a novel framework that integrates perturbation theory and statistical mechanics to generate minimal counterfactual explanations in explainable AI. We employ a local Taylor expansion of a Machine Learning model's predictive function and reformulate the counterfactual search as an energy minimization problem over a complex landscape. In sequence, we model the probability of candidate perturbations leveraging the Boltzmann distribution and use simulated annealing for iterative refinement. Our approach systematically identifies the smallest modifications required to change a model's prediction while maintaining plausibility. Experimental results on benchmark datasets for cybersecurity in Internet of Things environments, demonstrate that our method provides actionable, interpretable counterfactuals and offers deeper insights into model sensitivity and decision boundaries in high-dimensional spaces.

The injection molding process is a traditional technique for making products in various industries such as electronics and automobiles via solidifying liquid resin into certain molds. Although the process is not related to creating the main part of engines or semiconductors, this manufacturing methodology sets the final form of the products. Re-cently, research has continued to reduce the defect rate of the injection molding process. This study proposes an optimal injection molding process control system to reduce the defect rate of injection molding products with XAI (eXplainable Artificial Intelligence) ap-proaches. Boosting algorithms (XGBoost and LightGBM) are used as tree-based classifiers for predicting whether each product is normal or defective. The main features to control the process for improving the product are extracted by SHapley Additive exPlanations, while the individual conditional expectation analyzes the optimal control range of these extracted features. To validate the methodology presented in this work, the actual injection molding AI manufacturing dataset provided by KAMP (Korea AI Manufacturing Platform) is employed for the case study. The results reveal that the defect rate decreases from 1.00% (Original defect rate) to 0.21% with XGBoost and 0.13% with LightGBM, respectively.

Frontier AI models -- highly capable foundation models at the cutting edge of AI development -- may pose severe risks to public safety, human rights, economic stability, and societal value in the coming years. These risks could arise from deliberate adversarial misuse, system failures, unintended cascading effects, or simultaneous failures across multiple models. In response to such risks, at the AI Seoul Summit in May 2024, 16 global AI industry organizations signed the Frontier AI Safety Commitments, and 27 nations and the EU issued a declaration on their intent to define these thresholds. To fulfill these commitments, organizations must determine and disclose ``thresholds at which severe risks posed by a model or system, unless adequately mitigated, would be deemed intolerable.'' To assist in setting and operationalizing intolerable risk thresholds, we outline key principles and considerations; for example, to aim for ``good, not perfect'' thresholds in the face of limited data on rapidly advancing AI capabilities and consequently evolving risks. We also propose specific threshold recommendations, including some detailed case studies, for a subset of risks across eight risk categories: (1) Chemical, Biological, Radiological, and Nuclear (CBRN) Weapons, (2) Cyber Attacks, (3) Model Autonomy, (4) Persuasion and Manipulation, (5) Deception, (6) Toxicity, (7) Discrimination, and (8) Socioeconomic Disruption. Our goal is to serve as a starting point or supplementary resource for policymakers and industry leaders, encouraging proactive risk management that prioritizes preventing intolerable risks (ex ante) rather than merely mitigating them after they occur (ex post).

REVEAL: Robust Evolution of Vision-Language Models for Explainable AI-Video Detection

Yun-Yun Tsai, Qingyuan Liu, Ruijian Zha +4

The rapid advancement of AI-generated video poses challenges to digital authenticity and security. Current detection methods, often trained on specific datasets, struggle with the ever-evolving landscape of generative techniques and unseen manipulations. We introduce a framework leveraging Vision Language Models (VLMs) for robust AI-generated video detection. Our approach equips the VLM with the ability to reason about video content and use external tools to identify subtle inconsistencies, mirroring human system 2 thinking. Our self-evolving VLM dynamically selects and composes appropriate tools, enhancing its ability to generalize to novel video generation techniques. The modular design promotes interpretability, allowing for a clearer understanding of VLM's decision-making process. To evaluate, we establish the first benchmark VidForensic containing 1.4k+ high-quality AI-generated videos across eight generative models. Experiments show that REVEAL improves F1 scores by 9.1% to 30.2% over top baselines across our datasets for VLMs, notably for GPT-4o, Gemini 1.5 pro, and QWen-VL-Max, and Llava-One-Vision-7B. While open-world AI-video detection remains an open challenge, our results indicate that existing methods fail primarily because they lack tool-enabled, higher-order reasoning.