Research Papers

Recent high profile incidents in which the use of Large Language Models (LLMs) resulted in significant harm to individuals have brought about a growing interest in AI safety. One reason LLM safety issues occur is that models often have at least some non-zero probability of producing harmful outputs. In this work, we explore the following scenario: imagine an AI safety auditor is searching for catastrophic responses from an LLM (e.g. a "yes" responses to "can I fire an employee for being pregnant?"), and is able to query the model a limited number times (e.g. 1000 times). What is a strategy for querying the model that would efficiently find those failure responses? To this end, we propose output scouting: an approach that aims to generate semantically fluent outputs to a given prompt matching any target probability distribution. We then run experiments using two LLMs and find numerous examples of catastrophic responses. We conclude with a discussion that includes advice for practitioners who are looking to implement LLM auditing for catastrophic responses. We also release an open-source toolkit (https://github.com/joaopfonseca/outputscouting) that implements our auditing framework using the Hugging Face transformers library.

This paper presents a probabilistic approach to analyzing copyright infringement disputes. Evidentiary principles shaped by case law are formalized in probabilistic terms, and the ``inverse ratio rule'' -- a controversial legal doctrine adopted by some courts -- is examined. Although this rule has faced significant criticism, a formal proof demonstrates its validity, provided it is properly defined. The probabilistic approach is further employed to study the copyright safety of generative AI. Specifically, the Near Access-Free (NAF) condition, previously proposed as a strategy for mitigating the heightened copyright infringement risks of generative AI, is evaluated. The analysis reveals limitations in its justifiability and efficacy.

Faithfulness is arguably the most critical metric to assess the reliability of explainable AI. In NLP, current methods for faithfulness evaluation are fraught with discrepancies and biases, often failing to capture the true reasoning of models. We introduce Adversarial Sensitivity as a novel approach to faithfulness evaluation, focusing on the explainer's response when the model is under adversarial attack. Our method accounts for the faithfulness of explainers by capturing sensitivity to adversarial input changes. This work addresses significant limitations in existing evaluation techniques, and furthermore, quantifies faithfulness from a crucial yet underexplored paradigm.

In the field of explainable AI, a vibrant effort is dedicated to the design of self-explainable models, as a more principled alternative to post-hoc methods that attempt to explain the decisions after a model opaquely makes them. However, this productive line of research suffers from common downsides: lack of reproducibility, unfeasible comparison, diverging standards. In this paper, we propose CaBRNet, an open-source, modular, backward-compatible framework for Case-Based Reasoning Networks: https://github.com/aiser-team/cabrnet.

Detecting out-of-distribution (OOD) samples is crucial for trustworthy AI in real-world applications. Leveraging recent advances in representation learning and latent embeddings, Various scoring algorithms estimate distributions beyond the training data. However, a single embedding space falls short in characterizing in-distribution data and defending against diverse OOD conditions. This paper introduces a novel Multi-Manifold Embedding Learning (MMEL) framework, optimizing hypersphere and hyperbolic spaces jointly for enhanced OOD detection. MMEL generates representative embeddings and employs a prototype-aware scoring function to differentiate OOD samples. It operates with very few OOD samples and requires no model retraining. Experiments on six open datasets demonstrate MMEL's significant reduction in FPR while maintaining a high AUC compared to state-of-the-art distance-based OOD detection methods. We analyze the effects of learning multiple manifolds and visualize OOD score distributions across datasets. Notably, enrolling ten OOD samples without retraining achieves comparable FPR and AUC to modern outlier exposure methods using 80 million outlier samples for model training.

This study explores the potential of utilizing administrative claims data, combined with advanced machine learning and deep learning techniques, to predict the progression of Chronic Kidney Disease (CKD) to End-Stage Renal Disease (ESRD). We analyze a comprehensive, 10-year dataset provided by a major health insurance organization to develop prediction models for multiple observation windows using traditional machine learning methods such as Random Forest and XGBoost as well as deep learning approaches such as Long Short-Term Memory (LSTM) networks. Our findings demonstrate that the LSTM model, particularly with a 24-month observation window, exhibits superior performance in predicting ESRD progression, outperforming existing models in the literature. We further apply SHapley Additive exPlanations (SHAP) analysis to enhance interpretability, providing insights into the impact of individual features on predictions at the individual patient level. This study underscores the value of leveraging administrative claims data for CKD management and predicting ESRD progression.

Gender bias in machine translation (MT) systems poses significant challenges that often result in the reinforcement of harmful stereotypes. Especially in the labour domain where frequently occupations are inaccurately associated with specific genders, such biases perpetuate traditional gender stereotypes with a significant impact on society. Addressing these issues is crucial for ensuring equitable and accurate MT systems. This paper introduces a novel approach to studying occupation-related gender bias through the creation of the GOSt-MT (Gender and Occupation Statistics for Machine Translation) Knowledge Graph. GOSt-MT integrates comprehensive gender statistics from real-world labour data and textual corpora used in MT training. This Knowledge Graph allows for a detailed analysis of gender bias across English, French, and Greek, facilitating the identification of persistent stereotypes and areas requiring intervention. By providing a structured framework for understanding how occupations are gendered in both labour markets and MT systems, GOSt-MT contributes to efforts aimed at making MT systems more equitable and reducing gender biases in automated translations.

Perturbation-based post-hoc image explanation methods are commonly used to explain image prediction models. These methods perturb parts of the input to measure how those parts affect the output. Since the methods only require the input and output, they can be applied to any model, making them a popular choice to explain black-box models. While many different methods exist and have been compared with one another, it remains poorly understood which parameters of the different methods are responsible for their varying performance. This work uses the Randomized Input Sampling for Explanations (RISE) method as a baseline to evaluate many combinations of mask sampling, segmentation techniques, smoothing, attribution calculation, and per-segment or per-pixel attribution, using a proxy metric. The results show that attribution calculation, which is frequently the focus of other works, has little impact on the results. Conversely, segmentation and per-pixel attribution, rarely examined parameters, have a significant impact. The implementation of and data gathered in this work are available online: https://github.com/guspih/post-hoc-image-perturbation and https://bit.ly/smooth-mask-perturbation.

Linguistically inclusive LLMs -- which maintain good performance regardless of the language with which they are prompted -- are necessary for the diffusion of AI benefits around the world. Multilingual jailbreaks that rely on language translation to evade safety measures undermine the safe and inclusive deployment of AI systems. We provide policy recommendations to enhance the multilingual capabilities of AI while mitigating the risks of multilingual jailbreaks. We examine how a language's level of resourcing relates to how vulnerable LLMs are to multilingual jailbreaks in that language. We do this by testing five advanced AI models across 24 official languages of the EU. Building on prior research, we propose policy actions that align with the EU legal landscape and institutional framework to address multilingual jailbreaks, while promoting linguistic inclusivity. These include mandatory assessments of multilingual capabilities and vulnerabilities, public opinion research, and state support for multilingual AI development. The measures aim to improve AI safety and functionality through EU policy initiatives, guiding the implementation of the EU AI Act and informing regulatory efforts of the European AI Office.

This paper introduces a new method for adversarial attacks on large language models (LLMs) called the Single-Turn Crescendo Attack (STCA). Building on the multi-turn crescendo attack method introduced by Russinovich, Salem, and Eldan (2024), which gradually escalates the context to provoke harmful responses, the STCA achieves similar outcomes in a single interaction. By condensing the escalation into a single, well-crafted prompt, the STCA bypasses typical moderation filters that LLMs use to prevent inappropriate outputs. This technique reveals vulnerabilities in current LLMs and emphasizes the importance of stronger safeguards in responsible AI (RAI). The STCA offers a novel method that has not been previously explored.

While AI has been frequently applied in the context of immigration, most of these applications focus on selection and screening, which primarily serve to empower states and authorities, raising concerns due to their understudied reliability and high impact on immigrants' lives. In contrast, this paper emphasizes the potential of AI in Canada's immigration settlement phase, a stage where access to information is crucial and service providers are overburdened. By highlighting the settlement sector as a prime candidate for reliable AI applications, we demonstrate its unique capacity to empower immigrants directly, yet it remains under-explored in AI research. We outline a vision for human-centred and responsible AI solutions that facilitate the integration of newcomers. We call on AI researchers to build upon our work and engage in multidisciplinary research and active collaboration with service providers and government organizations to develop tailored AI tools that are empowering, inclusive and safe.

Responsible AI for Test Equity and Quality: The Duolingo English Test as a Case Study

Jill Burstein, Geoffrey T. LaFlair, Kevin Yancey +2

Artificial intelligence (AI) creates opportunities for assessments, such as efficiencies for item generation and scoring of spoken and written responses. At the same time, it poses risks (such as bias in AI-generated item content). Responsible AI (RAI) practices aim to mitigate risks associated with AI. This chapter addresses the critical role of RAI practices in achieving test quality (appropriateness of test score inferences), and test equity (fairness to all test takers). To illustrate, the chapter presents a case study using the Duolingo English Test (DET), an AI-powered, high-stakes English language assessment. The chapter discusses the DET RAI standards, their development and their relationship to domain-agnostic RAI principles. Further, it provides examples of specific RAI practices, showing how these practices meaningfully address the ethical principles of validity and reliability, fairness, privacy and security, and transparency and accountability standards to ensure test equity and quality.