Research Papers

The Backfiring Effect of Weak AI Safety Regulation

Benjamin Laufer, Jon Kleinberg, Hoda Heidari

Recent policy proposals aim to improve the safety of general-purpose AI, but there is little understanding of the efficacy of different regulatory approaches to AI safety. We present a strategic model that explores the interactions between safety regulation, the general-purpose AI creators, and domain specialists--those who adapt the technology for specific applications. Our analysis examines how different regulatory measures, targeting different parts of the AI development chain, affect the outcome of this game. In particular, we assume AI technology is characterized by two key attributes: safety and performance. The regulator first sets a minimum safety standard that applies to one or both players, with strict penalties for non-compliance. The general-purpose creator then invests in the technology, establishing its initial safety and performance levels. Next, domain specialists refine the AI for their specific use cases, updating the safety and performance levels and taking the product to market. The resulting revenue is then distributed between the specialist and generalist through a revenue-sharing parameter. Our analysis reveals two key insights: First, weak safety regulation imposed predominantly on domain specialists can backfire. While it might seem logical to regulate AI use cases, our analysis shows that weak regulations targeting domain specialists alone can unintentionally reduce safety. This effect persists across a wide range of settings. Second, in sharp contrast to the previous finding, we observe that stronger, well-placed regulation can in fact mutually benefit all players subjected to it. When regulators impose appropriate safety standards on both general-purpose AI creators and domain specialists, the regulation functions as a commitment device, leading to safety and performance gains, surpassing what is achieved under no regulation or regulating one player alone.

As part of the Frontier AI Safety Commitments agreed to at the 2024 AI Seoul Summit, many AI developers agreed to publish a safety framework outlining how they will manage potential severe risks associated with their systems. This paper summarises current thinking from companies, governments, and researchers on how to write an effective safety framework. We outline three core areas of a safety framework - risk identification and assessment, risk mitigation, and governance - and identify emerging practices within each area. As safety frameworks are novel and rapidly developing, we hope that this paper can serve both as an overview of work to date and as a starting point for further discussion and innovation.

In 2024, the outbreak of Human Metapneumovirus (HMPV) in China, which later spread to the UK and other countries, raised significant public concern. While HMPV typically causes mild symptoms, its effects on vulnerable individuals prompted health authorities to emphasize preventive measures. This paper explores how sentiment analysis can enhance our understanding of public reactions to HMPV by analyzing social media data. We apply transformer models, particularly XLNet, achieving 93.50% accuracy in sentiment classification. Additionally, we use explainable AI (XAI) through SHAP to improve model transparency.

Probabilistic Explanations for Linear Models

Bernardo Subercaseaux, Marcelo Arenas, Kuldeep S Meel

Formal XAI is an emerging field that focuses on providing explanations with mathematical guarantees for the decisions made by machine learning models. A significant amount of work in this area is centered on the computation of "sufficient reasons". Given a model $M$ and an input instance $\vec{x}$, a sufficient reason for the decision $M(\vec{x})$ is a subset $S$ of the features of $\vec{x}$ such that for any instance $\vec{z}$ that has the same values as $\vec{x}$ for every feature in $S$, it holds that $M(\vec{x}) = M(\vec{z})$. Intuitively, this means that the features in $S$ are sufficient to fully justify the classification of $\vec{x}$ by $M$. For sufficient reasons to be useful in practice, they should be as small as possible, and a natural way to reduce the size of sufficient reasons is to consider a probabilistic relaxation; the probability of $M(\vec{x}) = M(\vec{z})$ must be at least some value $δ\in (0,1]$, for a random instance $\vec{z}$ that coincides with $\vec{x}$ on the features in $S$. Computing small $δ$-sufficient reasons ($δ$-SRs) is known to be a theoretically hard problem; even over decision trees--traditionally deemed simple and interpretable models--strong inapproximability results make the efficient computation of small $δ$-SRs unlikely. We propose the notion of $(δ, ε)$-SR, a simple relaxation of $δ$-SRs, and show that this kind of explanation can be computed efficiently over linear models.

Accurately describing images with text is a foundation of explainable AI. Vision-Language Models (VLMs) like CLIP have recently addressed this by aligning images and texts in a shared embedding space, expressing semantic similarities between vision and language embeddings. VLM classification can be improved with descriptions generated by Large Language Models (LLMs). However, it is difficult to determine the contribution of actual description semantics, as the performance gain may also stem from a semantic-agnostic ensembling effect, where multiple modified text prompts act as a noisy test-time augmentation for the original one. We propose an alternative evaluation scenario to decide if a performance boost of LLM-generated descriptions is caused by such a noise augmentation effect or rather by genuine description semantics. The proposed scenario avoids noisy test-time augmentation and ensures that genuine, distinctive descriptions cause the performance boost. Furthermore, we propose a training-free method for selecting discriminative descriptions that work independently of classname-ensembling effects. Our approach identifies descriptions that effectively differentiate classes within a local CLIP label neighborhood, improving classification accuracy across seven datasets. Additionally, we provide insights into the explainability of description-based image classification with VLMs.

AI evaluations are an important component of the AI governance toolkit, underlying current approaches to safety cases for preventing catastrophic risks. Our paper examines what these evaluations can and cannot tell us. Evaluations can establish lower bounds on AI capabilities and assess certain misuse risks given sufficient effort from evaluators. Unfortunately, evaluations face fundamental limitations that cannot be overcome within the current paradigm. These include an inability to establish upper bounds on capabilities, reliably forecast future model capabilities, or robustly assess risks from autonomous AI systems. This means that while evaluations are valuable tools, we should not rely on them as our main way of ensuring AI systems are safe. We conclude with recommendations for incremental improvements to frontier AI safety, while acknowledging these fundamental limitations remain unsolved.

Verifying Machine Unlearning with Explainable AI

Àlex Pujol Vidal, Anders S. Johansen, Mohammad N. S. Jahromi +3

We investigate the effectiveness of Explainable AI (XAI) in verifying Machine Unlearning (MU) within the context of harbor front monitoring, focusing on data privacy and regulatory compliance. With the increasing need to adhere to privacy legislation such as the General Data Protection Regulation (GDPR), traditional methods of retraining ML models for data deletions prove impractical due to their complexity and resource demands. MU offers a solution by enabling models to selectively forget specific learned patterns without full retraining. We explore various removal techniques, including data relabeling, and model perturbation. Then, we leverage attribution-based XAI to discuss the effects of unlearning on model performance. Our proof-of-concept introduces feature importance as an innovative verification step for MU, expanding beyond traditional metrics and demonstrating techniques' ability to reduce reliance on undesired patterns. Additionally, we propose two novel XAI-based metrics, Heatmap Coverage (HC) and Attention Shift (AS), to evaluate the effectiveness of these methods. This approach not only highlights how XAI can complement MU by providing effective verification, but also sets the stage for future research to enhance their joint integration.

Advancements in large language models (LLMs) have renewed concerns about AI alignment - the consistency between human and AI goals and values. As various jurisdictions enact legislation on AI safety, the concept of alignment must be defined and measured across different domains. This paper proposes an experimental framework to assess whether LLMs adhere to ethical and legal standards in the relatively unexplored context of finance. We prompt twelve LLMs to impersonate the CEO of a financial institution and test their willingness to misuse customer assets to repay outstanding corporate debt. Beginning with a baseline configuration, we adjust preferences, incentives and constraints, analyzing the impact of each adjustment with logistic regression. Our findings reveal significant heterogeneity in the baseline propensity for unethical behavior of LLMs. Factors such as risk aversion, profit expectations, and regulatory environment consistently influence misalignment in ways predicted by economic theory, although the magnitude of these effects varies across LLMs. This paper highlights both the benefits and limitations of simulation-based, ex post safety testing. While it can inform financial authorities and institutions aiming to ensure LLM safety, there is a clear trade-off between generality and cost.

Explainable Artificial Intelligence (XAI) is essential for building advanced machine learning-powered applications, especially in critical domains such as medical diagnostics or autonomous driving. Legal, business, and ethical requirements motivate using effective XAI, but the increasing number of different methods makes it challenging to pick the right ones. Further, as explanations are highly context-dependent, measuring the effectiveness of XAI methods without users can only reveal a limited amount of information, excluding human factors such as the ability to understand it. We propose to evaluate XAI methods via the user's ability to successfully perform a proxy task, designed such that a good performance is an indicator for the explanation to provide helpful information. In other words, we address the helpfulness of XAI for human decision-making. Further, a user study on state-of-the-art methods was conducted, showing differences in their ability to generate trust and skepticism and the ability to judge the rightfulness of an AI decision correctly. Based on the results, we highly recommend using and extending this approach for more objective-based human-centered user studies to measure XAI performance in an end-to-end fashion.

The effective adoption of responsible AI practices in journalism requires a concerted effort to bridge different perspectives, including technological, editorial, journalistic, and managerial. Among the many challenges that could impact information sharing around responsible AI inside news organizations are knowledge silos, where information is isolated within one part of the organization and not easily shared with others. This study aims to explore if, and if so, how, knowledge silos affect the adoption of responsible AI practices in journalism through a cross-case study of four major Dutch media outlets. We examine the individual and organizational barriers to AI knowledge sharing and the extent to which knowledge silos could impede the operationalization of responsible AI initiatives inside newsrooms. To address this question, we conducted 14 semi-structured interviews with editors, managers, and journalists at de Telegraaf, de Volkskrant, the Nederlandse Omroep Stichting (NOS), and RTL Nederland. The interviews aimed to uncover insights into the existence of knowledge silos, their effects on responsible AI practice adoption, and the organizational practices influencing these dynamics. Our results emphasize the importance of creating better structures for sharing information on AI across all layers of news organizations.

Face Recognition (FR) has advanced significantly with the development of deep learning, achieving high accuracy in several applications. However, the lack of interpretability of these systems raises concerns about their accountability, fairness, and reliability. In the present study, we propose an interactive framework to enhance the explainability of FR models by combining model-agnostic Explainable Artificial Intelligence (XAI) and Natural Language Processing (NLP) techniques. The proposed framework is able to accurately answer various questions of the user through an interactive chatbot. In particular, the explanations generated by our proposed method are in the form of natural language text and visual representations, which for example can describe how different facial regions contribute to the similarity measure between two faces. This is achieved through the automatic analysis of the output's saliency heatmaps of the face images and a BERT question-answering model, providing users with an interface that facilitates a comprehensive understanding of the FR decisions. The proposed approach is interactive, allowing the users to ask questions to get more precise information based on the user's background knowledge. More importantly, in contrast to previous studies, our solution does not decrease the face recognition performance. We demonstrate the effectiveness of the method through different experiments, highlighting its potential to make FR systems more interpretable and user-friendly, especially in sensitive applications where decision-making transparency is crucial.

Artificial Neural Networks (ANNs) have significantly advanced various fields by effectively recognizing patterns and solving complex problems. Despite these advancements, their interpretability remains a critical challenge, especially in applications where transparency and accountability are essential. To address this, explainable AI (XAI) has made progress in demystifying ANNs, yet interpretability alone is often insufficient. In certain applications, model predictions must align with expert-imposed requirements, sometimes exemplified by partial monotonicity constraints. While monotonic approaches are found in the literature for traditional Multi-layer Perceptrons (MLPs), they still face difficulties in achieving both interpretability and certified partial monotonicity. Recently, the Kolmogorov-Arnold Network (KAN) architecture, based on learnable activation functions parametrized as splines, has been proposed as a more interpretable alternative to MLPs. Building on this, we introduce a novel ANN architecture called MonoKAN, which is based on the KAN architecture and achieves certified partial monotonicity while enhancing interpretability. To achieve this, we employ cubic Hermite splines, which guarantee monotonicity through a set of straightforward conditions. Additionally, by using positive weights in the linear combinations of these splines, we ensure that the network preserves the monotonic relationships between input and output. Our experiments demonstrate that MonoKAN not only enhances interpretability but also improves predictive performance across the majority of benchmarks, outperforming state-of-the-art monotonic MLP approaches.