Research Papers

Stronger Together: on the Articulation of Ethical Charters, Legal Tools, and Technical Documentation in ML

Giada Pistilli, Carlos Munoz Ferrandis, Yacine Jernite +1

The growing need for accountability of the people behind AI systems can be addressed by leveraging processes in three fields of study: ethics, law, and computer science. While these fields are often considered in isolation, they rely on complementary notions in their interpretation and implementation. In this work, we detail this interdependence and motivate the necessary role of collaborative governance tools in shaping a positive evolution of AI. We first contrast notions of compliance in the ethical, legal, and technical fields; we outline both their differences and where they complement each other, with a particular focus on the roles of ethical charters, licenses, and technical documentation in these interactions. We then focus on the role of values in articulating the synergies between the fields and outline specific mechanisms of interaction between them in practice. We identify how these mechanisms have played out in several open governance fora: an open collaborative workshop, a responsible licensing initiative, and a proposed regulatory framework. By leveraging complementary notions of compliance in these three domains, we can create a more comprehensive framework for governing AI systems that jointly takes into account their technical capabilities, their impact on society, and how technical specifications can inform relevant regulations. Our analysis thus underlines the necessity of joint consideration of the ethical, legal, and technical in AI ethics frameworks to be used on a larger scale to govern AI systems and how the thinking in each of these areas can inform the others.

Generative multimodal models based on diffusion models have seen tremendous growth and advances in recent years. Models such as DALL-E and Stable Diffusion have become increasingly popular and successful at creating images from texts, often combining abstract ideas. However, like other deep learning models, they also reflect social biases they inherit from their training data, which is often crawled from the internet. Manually auditing models for biases can be very time and resource consuming and is further complicated by the unbounded and unconstrained nature of inputs these models can take. Research into bias measurement and quantification has generally focused on small single-stage models working on a single modality. Thus the emergence of multistage multimodal models requires a different approach. In this paper, we propose Multimodal Composite Association Score (MCAS) as a new method of measuring gender bias in multimodal generative models. Evaluating both DALL-E 2 and Stable Diffusion using this approach uncovered the presence of gendered associations of concepts embedded within the models. We propose MCAS as an accessible and scalable method of quantifying potential bias for models with different modalities and a range of potential biases.

On Computing Probabilistic Abductive Explanations

Yacine Izza, Xuanxiang Huang, Alexey Ignatiev +3

The most widely studied explainable AI (XAI) approaches are unsound. This is the case with well-known model-agnostic explanation approaches, and it is also the case with approaches based on saliency maps. One solution is to consider intrinsic interpretability, which does not exhibit the drawback of unsoundness. Unfortunately, intrinsic interpretability can display unwieldy explanation redundancy. Formal explainability represents the alternative to these non-rigorous approaches, with one example being PI-explanations. Unfortunately, PI-explanations also exhibit important drawbacks, the most visible of which is arguably their size. Recently, it has been observed that the (absolute) rigor of PI-explanations can be traded off for a smaller explanation size, by computing the so-called relevant sets. Given some positive δ, a set S of features is δ-relevant if, when the features in S are fixed, the probability of getting the target class exceeds δ. However, even for very simple classifiers, the complexity of computing relevant sets of features is prohibitive, with the decision problem being NPPP-complete for circuit-based classifiers. In contrast with earlier negative results, this paper investigates practical approaches for computing relevant sets for a number of widely used classifiers that include Decision Trees (DTs), Naive Bayes Classifiers (NBCs), and several families of classifiers obtained from propositional languages. Moreover, the paper shows that, in practice, and for these families of classifiers, relevant sets are easy to compute. Furthermore, the experiments confirm that succinct sets of relevant features can be obtained for the families of classifiers considered.

Predictive business process monitoring increasingly leverages sophisticated prediction models. Although sophisticated models achieve consistently higher prediction accuracy than simple models, one major drawback is their lack of interpretability, which limits their adoption in practice. We thus see growing interest in explainable predictive business process monitoring, which aims to increase the interpretability of prediction models. Existing solutions focus on giving factual explanations.While factual explanations can be helpful, humans typically do not ask why a particular prediction was made, but rather why it was made instead of another prediction, i.e., humans are interested in counterfactual explanations. While research in explainable AI produced several promising techniques to generate counterfactual explanations, directly applying them to predictive process monitoring may deliver unrealistic explanations, because they ignore the underlying process constraints. We propose LORELEY, a counterfactual explanation technique for predictive process monitoring, which extends LORE, a recent explainable AI technique. We impose control flow constraints to the explanation generation process to ensure realistic counterfactual explanations. Moreover, we extend LORE to enable explaining multi-class classification models. Experimental results using a real, public dataset indicate that LORELEY can approximate the prediction models with an average fidelity of 97.69\% and generate realistic counterfactual explanations.

The last decade has witnessed many visual analytics (VA) systems that make successful applications to wide-ranging domains like urban analytics and explainable AI. However, their research rigor and contributions have been extensively challenged within the visualization community. We come in defence of VA systems by contributing two interview studies for gathering critics and responses to those criticisms. First, we interview 24 researchers to collect criticisms the review comments on their VA work. Through an iterative coding and refinement process, the interview feedback is summarized into a list of 36 common criticisms. Second, we interview 17 researchers to validate our list and collect their responses, thereby discussing implications for defending and improving the scientific values and rigor of VA systems. We highlight that the presented knowledge is deep, extensive, but also imperfect, provocative, and controversial, and thus recommend reading with an inclusive and critical eye. We hope our work can provide thoughts and foundations for conducting VA research and spark discussions to promote the research field forward more rigorously and vibrantly.

Medical students will almost inevitably encounter powerful medical AI systems early in their careers. Yet, contemporary medical education does not adequately equip students with the basic clinical proficiency in medical AI needed to use these tools safely and effectively. Education reform is urgently needed, but not easily implemented, largely due to an already jam-packed medical curricula. In this article, we propose an education reform framework as an effective and efficient solution, which we call the Embedded AI Ethics Education Framework. Unlike other calls for education reform to accommodate AI teaching that are more radical in scope, our framework is modest and incremental. It leverages existing bioethics or medical ethics curricula to develop and deliver content on the ethical issues associated with medical AI, especially the harms of technology misuse, disuse, and abuse that affect the risk-benefit analyses at the heart of healthcare. In doing so, the framework provides a simple tool for going beyond the "What?" and the "Why?" of medical AI ethics education, to answer the "How?", giving universities, course directors, and/or professors a broad road-map for equipping their students with the necessary clinical proficiency in medical AI.

Deep neural networks (DNNs) have recently been applied and used in many advanced and diverse tasks, such as medical diagnosis, automatic driving, etc. Due to the lack of transparency of the deep models, DNNs are often criticized for their prediction that cannot be explainable by human. In this paper, we propose a novel Morphological Fragmental Perturbation Pyramid (MFPP) method to solve the Explainable AI problem. In particular, we focus on the black-box scheme, which can identify the input area that is responsible for the output of the DNN without having to understand the internal architecture of the DNN. In the MFPP method, we divide the input image into multi-scale fragments and randomly mask out fragments as perturbation to generate a saliency map, which indicates the significance of each pixel for the prediction result of the black box model. Compared with the existing input sampling perturbation method, the pyramid structure fragment has proved to be more effective. It can better explore the morphological information of the input image to match its semantic information, and does not need any value inside the DNN. We qualitatively and quantitatively prove that MFPP meets and exceeds the performance of state-of-the-art (SOTA) black-box interpretation method on multiple DNN models and datasets.

Fair Allocation through Selective Information Acquisition

William Cai, Johann Gaebler, Nikhil Garg +1

Public and private institutions must often allocate scare resources under uncertainty. Banks, for example, extend credit to loan applicants based in part on their estimated likelihood of repaying a loan. But when the quality of information differs across candidates (e.g., if some applicants lack traditional credit histories), common lending strategies can lead to disparities across groups. Here we consider a setting in which decision makers -- before allocating resources -- can choose to spend some of their limited budget further screening select individuals. We present a computationally efficient algorithm for deciding whom to screen that maximizes a standard measure of social welfare. Intuitively, decision makers should screen candidates on the margin, for whom the additional information could plausibly alter the allocation. We formalize this idea by showing the problem can be reduced to solving a series of linear programs. Both on synthetic and real-world datasets, this strategy improves utility, illustrating the value of targeted information acquisition in such decisions. Further, when there is social value for distributing resources to groups for whom we have a priori poor information -- like those without credit scores -- our approach can substantially improve the allocation of limited assets.

Explainable machine learning (ML) enables human learning from ML, human appeal of automated model decisions, regulatory compliance, and security audits of ML models. Explainable ML (i.e. explainable artificial intelligence or XAI) has been implemented in numerous open source and commercial packages and explainable ML is also an important, mandatory, or embedded aspect of commercial predictive modeling in industries like financial services. However, like many technologies, explainable ML can be misused, particularly as a faulty safeguard for harmful black-boxes, e.g. fairwashing or scaffolding, and for other malevolent purposes like stealing models and sensitive training data. To promote best-practice discussions for this already in-flight technology, this short text presents internal definitions and a few examples before covering the proposed guidelines. This text concludes with a seemingly natural argument for the use of interpretable models and explanatory, debugging, and disparate impact testing methods in life- or mission-critical ML systems.

Phishing has been a prevalent cyber threat that manipulates users into revealing sensitive private information through deceptive tactics, designed to masquerade as trustworthy entities. Over the years, proactively detection of phishing URLs (or websites) has been established as an widely-accepted defense approach. In literature, we often find supervised Machine Learning (ML) models with highly competitive performance for detecting phishing websites based on the extracted features from both phishing and benign (i.e., legitimate) websites. However, it is still unclear if these features or indicators are dependent on a particular dataset or they are generalized for overall phishing detection. In this paper, we delve deeper into this issue by analyzing two publicly available phishing URL datasets, where each dataset has its own set of unique and overlapping features related to URL string and website contents. We want to investigate if overlapping features are similar in nature across datasets and how does the model perform when trained on one dataset and tested on the other. We conduct practical experiments and leverage explainable AI (XAI) methods such as SHAP plots to provide insights into different features' contributions in case of phishing detection to answer our primary question, "Can features for phishing URL detection be trusted across diverse dataset?". Our case study experiment results show that features for phishing URL detection can often be dataset-dependent and thus may not be trusted across different datasets even though they share same set of feature behaviors.

Large Language Models (LLMs) are being adopted across a wide range of tasks, including decision-making processes in industries where bias in AI systems is a significant concern. Recent research indicates that LLMs can harbor implicit biases even when they pass explicit bias evaluations. Building upon the frameworks of the LLM Implicit Association Test (IAT) Bias and LLM Decision Bias, this study highlights that newer or larger language models do not automatically exhibit reduced bias; in some cases, they displayed higher bias scores than their predecessors, such as in Meta's Llama series and OpenAI's GPT models. This suggests that increasing model complexity without deliberate bias mitigation strategies can unintentionally amplify existing biases. The variability in bias scores within and across providers underscores the need for standardized evaluation metrics and benchmarks for bias assessment. The lack of consistency indicates that bias mitigation is not yet a universally prioritized goal in model development, which can lead to unfair or discriminatory outcomes. By broadening the detection of implicit bias, this research provides a more comprehensive understanding of the biases present in advanced models and underscores the critical importance of addressing these issues to ensure the development of fair and responsible AI systems.

Researchers, government bodies, and organizations have been repeatedly calling for a shift in the responsible AI community from general principles to tangible and operationalizable practices in mitigating the potential sociotechnical harms of AI. Frameworks like the NIST AI RMF embody an emerging consensus on recommended practices in operationalizing sociotechnical harm mitigation. However, private sector organizations currently lag far behind this emerging consensus. Implementation is sporadic and selective at best. At worst, it is ineffective and can risk serving as a misleading veneer of trustworthy processes, providing an appearance of legitimacy to substantively harmful practices. In this paper, we provide a foundation for a framework for evaluating where organizations sit relative to the emerging consensus on sociotechnical harm mitigation best practices: a flexible maturity model based on the NIST AI RMF.