Research Papers

In automated planning, the need for explanations arises when there is a mismatch between a proposed plan and the user's expectation. We frame Explainable AI Planning in the context of the plan negotiation problem, in which a succession of hypothetical planning problems are generated and solved. The object of the negotiation is for the user to understand and ultimately arrive at a satisfactory plan. We present the results of a user study that demonstrates that when users ask questions about plans, those questions are contrastive, i.e. "why A rather than B?". We use the data from this study to construct a taxonomy of user questions that often arise during plan negotiation. We formally define our approach to plan negotiation through model restriction as an iterative process. This approach generates hypothetical problems and contrastive plans by restricting the model through constraints implied by user questions. We formally define model-based compilations in PDDL2.1 of each constraint derived from a user question in the taxonomy, and empirically evaluate the compilations in terms of computational complexity. The compilations were implemented as part of an explanation framework that employs iterative model restriction. We demonstrate its benefits in a second user study.

Advances in computing power, deep learning architectures, and expert labelled datasets have spurred the development of medical imaging artificial intelligence systems that rival clinical experts in a variety of scenarios. The National Institutes of Health in 2018 identified key focus areas for the future of artificial intelligence in medical imaging, creating a foundational roadmap for research in image acquisition, algorithms, data standardization, and translatable clinical decision support systems. Among the key issues raised in the report: data availability, need for novel computing architectures and explainable AI algorithms, are still relevant despite the tremendous progress made over the past few years alone. Furthermore, translational goals of data sharing, validation of performance for regulatory approval, generalizability and mitigation of unintended bias must be accounted for early in the development process. In this perspective paper we explore challenges unique to high dimensional clinical imaging data, in addition to highlighting some of the technical and ethical considerations in developing high-dimensional, multi-modality, machine learning systems for clinical decision support.

Large language models (LLMs) can sometimes report the strategies they actually use to solve tasks, yet at other times seem unable to recognize those strategies that govern their behavior. This suggests a limited degree of metacognition - the capacity to monitor one's own cognitive processes for subsequent reporting and self-control. Metacognition enhances LLMs' capabilities in solving complex tasks but also raises safety concerns, as models may obfuscate their internal processes to evade neural-activation-based oversight (e.g., safety detector). Given society's increased reliance on these models, it is critical that we understand their metacognitive abilities. To address this, we introduce a neuroscience-inspired neurofeedback paradigm that uses in-context learning to quantify metacognitive abilities of LLMs to report and control their activation patterns. We demonstrate that their abilities depend on several factors: the number of in-context examples provided, the semantic interpretability of the neural activation direction (to be reported/controlled), and the variance explained by that direction. These directions span a "metacognitive space" with dimensionality much lower than the model's neural space, suggesting LLMs can monitor only a small subset of their neural activations. Our paradigm provides empirical evidence to quantify metacognition in LLMs, with significant implications for AI safety (e.g., adversarial attack and defense).

Auditing plays a pivotal role in the development of trustworthy AI. However, current research primarily focuses on creating auditable AI documentation, which is intended for regulators and experts rather than end-users affected by AI decisions. How to communicate to members of the public that an AI has been audited and considered trustworthy remains an open challenge. This study empirically investigated certification labels as a promising solution. Through interviews (N = 12) and a census-representative survey (N = 302), we investigated end-users' attitudes toward certification labels and their effectiveness in communicating trustworthiness in low- and high-stakes AI scenarios. Based on the survey results, we demonstrate that labels can significantly increase end-users' trust and willingness to use AI in both low- and high-stakes scenarios. However, end-users' preferences for certification labels and their effect on trust and willingness to use AI were more pronounced in high-stake scenarios. Qualitative content analysis of the interviews revealed opportunities and limitations of certification labels, as well as facilitators and inhibitors for the effective use of labels in the context of AI. For example, while certification labels can mitigate data-related concerns expressed by end-users (e.g., privacy and data protection), other concerns (e.g., model performance) are more challenging to address. Our study provides valuable insights and recommendations for designing and implementing certification labels as a promising constituent within the trustworthy AI ecosystem.

The growing importance of understanding and addressing algorithmic bias in artificial intelligence (AI) has led to a surge in research on AI fairness, which often assumes that the underlying data is independent and identically distributed (IID). However, real-world data frequently exists in non-IID graph structures that capture connections among individual units. To effectively mitigate bias in AI systems, it is essential to bridge the gap between traditional fairness literature, designed for IID data, and the prevalence of non-IID graph data. This survey reviews recent advancements in fairness amidst non-IID graph data, including the newly introduced fair graph generation and the commonly studied fair graph classification. In addition, available datasets and evaluation metrics for future research are identified, the limitations of existing work are highlighted, and promising future directions are proposed.

In the last few years, AI continues demonstrating its positive impact on society while sometimes with ethically questionable consequences. Building and maintaining public trust in AI has been identified as the key to successful and sustainable innovation. This chapter discusses the challenges related to operationalizing ethical AI principles and presents an integrated view that covers high-level ethical AI principles, the general notion of trust/trustworthiness, and product/process support in the context of responsible AI, which helps improve both trust and trustworthiness of AI for a wider set of stakeholders.

With the introduction of Artificial Intelligence (AI) and related technologies in our daily lives, fear and anxiety about their misuse as well as the hidden biases in their creation have led to a demand for regulation to address such issues. Yet blindly regulating an innovation process that is not well understood, may stifle this process and reduce benefits that society may gain from the generated technology, even under the best intentions. In this paper, starting from a baseline model that captures the fundamental dynamics of a race for domain supremacy using AI technology, we demonstrate how socially unwanted outcomes may be produced when sanctioning is applied unconditionally to risk-taking, i.e. potentially unsafe, behaviours. As an alternative to resolve the detrimental effect of over-regulation, we propose a voluntary commitment approach wherein technologists have the freedom of choice between independently pursuing their course of actions or establishing binding agreements to act safely, with sanctioning of those that do not abide to what they pledged. Overall, this work reveals for the first time how voluntary commitments, with sanctions either by peers or an institution, leads to socially beneficial outcomes in all scenarios envisageable in a short-term race towards domain supremacy through AI technology. These results are directly relevant for the design of governance and regulatory policies that aim to ensure an ethical and responsible AI technology development process.

Traditional machine learning models often prioritize predictive accuracy, often at the expense of model transparency and interpretability. The lack of transparency makes it difficult for organizations to comply with regulatory requirements and gain stakeholders trust. In this research, we propose a fraud detection framework that combines a stacking ensemble of well-known gradient boosting models: XGBoost, LightGBM, and CatBoost. In addition, explainable artificial intelligence (XAI) techniques are used to enhance the transparency and interpretability of the model's decisions. We used SHAP (SHapley Additive Explanations) for feature selection to identify the most important features. Further efforts were made to explain the model's predictions using Local Interpretable Model-Agnostic Explanation (LIME), Partial Dependence Plots (PDP), and Permutation Feature Importance (PFI). The IEEE-CIS Fraud Detection dataset, which includes more than 590,000 real transaction records, was used to evaluate the proposed model. The model achieved a high performance with an accuracy of 99% and an AUC-ROC score of 0.99, outperforming several recent related approaches. These results indicate that combining high prediction accuracy with transparent interpretability is possible and could lead to a more ethical and trustworthy solution in financial fraud detection.

The Clever Hans Effect in Unsupervised Learning

Jacob Kauffmann, Jonas Dippel, Lukas Ruff +3

Unsupervised learning has become an essential building block of AI systems. The representations it produces, e.g. in foundation models, are critical to a wide variety of downstream applications. It is therefore important to carefully examine unsupervised models to ensure not only that they produce accurate predictions, but also that these predictions are not "right for the wrong reasons", the so-called Clever Hans (CH) effect. Using specially developed Explainable AI techniques, we show for the first time that CH effects are widespread in unsupervised learning. Our empirical findings are enriched by theoretical insights, which interestingly point to inductive biases in the unsupervised learning machine as a primary source of CH effects. Overall, our work sheds light on unexplored risks associated with practical applications of unsupervised learning and suggests ways to make unsupervised learning more robust.

In many AI applications today, the predominance of black-box machine learning models, due to their typically higher accuracy, amplifies the need for Explainable AI (XAI). Existing XAI approaches, such as the widely used SHAP values or counterfactual (CF) explanations, are arguably often too technical for users to understand and act upon. To enhance comprehension of explanations of AI decisions and the overall user experience, we introduce XAIstories, which leverage Large Language Models to provide narratives about how AI predictions are made: SHAPstories do so based on SHAP explanations, while CFstories do so for CF explanations. We study the impact of our approach on users' experience and understanding of AI predictions. Our results are striking: over 90% of the surveyed general audience finds the narratives generated by SHAPstories convincing. Data scientists primarily see the value of SHAPstories in communicating explanations to a general audience, with 83% of data scientists indicating they are likely to use SHAPstories for this purpose. In an image classification setting, CFstories are considered more or equally convincing as the users' own crafted stories by more than 75% of the participants. CFstories additionally bring a tenfold speed gain in creating a narrative. We also find that SHAPstories help users to more accurately summarize and understand AI decisions, in a credit scoring setting we test, correctly answering comprehension questions significantly more often than they do when only SHAP values are provided. The results thereby suggest that XAIstories may significantly help explaining and understanding AI predictions, ultimately supporting better decision-making in various applications.

Deep neural networks have become increasingly popular for analyzing ECG data because of their ability to accurately identify cardiac conditions and hidden clinical factors. However, the lack of transparency due to the black box nature of these models is a common concern. To address this issue, explainable AI (XAI) methods can be employed. In this study, we present a comprehensive analysis of post-hoc XAI methods, investigating the local (attributions per sample) and global (based on domain expert concepts) perspectives. We have established a set of sanity checks to identify sensible attribution methods, and we provide quantitative evidence in accordance with expert rules. This dataset-wide analysis goes beyond anecdotal evidence by aggregating data across patient subgroups. Furthermore, we demonstrate how these XAI techniques can be utilized for knowledge discovery, such as identifying subtypes of myocardial infarction. We believe that these proposed methods can serve as building blocks for a complementary assessment of the internal validity during a certification process, as well as for knowledge discovery in the field of ECG analysis.

While the emerging research field of explainable artificial intelligence (XAI) claims to address the lack of explainability in high-performance machine learning models, in practice, XAI targets developers rather than actual end-users. Unsurprisingly, end-users are often unwilling to use XAI-based decision support systems. Similarly, there is limited interdisciplinary research on end-users' behavior during XAI explanations usage, rendering it unknown how explanations may impact cognitive load and further affect end-user performance. Therefore, we conducted an empirical study with 271 prospective physicians, measuring their cognitive load, task performance, and task time for distinct implementation-independent XAI explanation types using a COVID-19 use case. We found that these explanation types strongly influence end-users' cognitive load, task performance, and task time. Further, we contextualized a mental efficiency metric, ranking local XAI explanation types best, to provide recommendations for future applications and implications for sociotechnical XAI research.