Research Papers

Responsible AI (rAI) guidance increasingly promotes stakeholder involvement (SHI) during AI development. At the same time, SHI is already common in commercial software development, but with potentially different foci. This study clarifies the extent to which established SHI practices are able to contribute to rAI efforts as well as potential disconnects -- essential insights to inform and tailor future interventions that further shift industry practice towards rAI efforts. First, we analysed 56 rAI guidance documents to identify why SHI is recommended (i.e. its expected benefits for rAI) and uncovered goals such as redistributing power, improving socio-technical understandings, anticipating risks, and enhancing public oversight. To understand why and how SHI is currently practised in commercial settings, we then conducted an online survey (n=130) and semi-structured interviews (n=10) with AI practitioners. Our findings reveal that SHI in practice is primarily driven by commercial priorities (e.g. customer value, compliance) and several factors currently discourage more rAI-aligned SHI practices. This suggests that established SHI practices are largely not contributing to rAI efforts. To address this disconnect, we propose interventions and research opportunities to advance rAI development in practice.

As frontier AI systems advance toward transformative capabilities, we need a parallel transformation in how we measure and evaluate these systems to ensure safety and inform governance. While benchmarks have been the primary method for estimating model capabilities, they often fail to establish true upper bounds or predict deployment behavior. This literature review consolidates the rapidly evolving field of AI safety evaluations, proposing a systematic taxonomy around three dimensions: what properties we measure, how we measure them, and how these measurements integrate into frameworks. We show how evaluations go beyond benchmarks by measuring what models can do when pushed to the limit (capabilities), the behavioral tendencies exhibited by default (propensities), and whether our safety measures remain effective even when faced with subversive adversarial AI (control). These properties are measured through behavioral techniques like scaffolding, red teaming and supervised fine-tuning, alongside internal techniques such as representation analysis and mechanistic interpretability. We provide deeper explanations of some safety-critical capabilities like cybersecurity exploitation, deception, autonomous replication, and situational awareness, alongside concerning propensities like power-seeking and scheming. The review explores how these evaluation methods integrate into governance frameworks to translate results into concrete development decisions. We also highlight challenges to safety evaluations - proving absence of capabilities, potential model sandbagging, and incentives for "safetywashing" - while identifying promising research directions. By synthesizing scattered resources, this literature review aims to provide a central reference point for understanding AI safety evaluations.

Recognizing daily activities with unobtrusive sensors in smart environments enables various healthcare applications. Monitoring how subjects perform activities at home and their changes over time can reveal early symptoms of health issues, such as cognitive decline. Most approaches in this field use deep learning models, which are often seen as black boxes mapping sensor data to activities. However, non-expert users like clinicians need to trust and understand these models' outputs. Thus, eXplainable AI (XAI) methods for Human Activity Recognition have emerged to provide intuitive natural language explanations from these models. Different XAI methods generate different explanations, and their effectiveness is typically evaluated through user surveys, that are often challenging in terms of costs and fairness. This paper proposes an automatic evaluation method using Large Language Models (LLMs) to identify, in a pool of candidates, the best XAI approach for non-expert users. Our preliminary results suggest that LLM evaluation aligns with user surveys.

The past years have presented a surge in (AI) development, fueled by breakthroughs in deep learning, increased computational power, and substantial investments in the field. Given the generative capabilities of more recent AI systems, the era of large-scale AI models has transformed various domains that intersect our daily lives. However, this progress raises concerns about the balance between technological advancement, ethical considerations, safety measures, and financial interests. Moreover, using such systems in sensitive areas amplifies our general ethical awareness, prompting a reemergence of debates on governance, regulation, and human values. However, amidst this landscape, how to bridge the principle-practice gap separating ethical discourse from the technical side of AI development remains an open problem. In response to this challenge, the present work proposes a framework to help shorten this gap: ethical problem-solving (EPS). EPS is a methodology promoting responsible, human-centric, and value-oriented AI development. The framework's core resides in translating principles into practical implementations using impact assessment surveys and a differential recommendation methodology. We utilize EPS as a blueprint to propose the implementation of Ethics as a Service Platform, which is currently available as a simple demonstration. We released all framework components openly and with a permissive license, hoping the community would adopt and extend our efforts into other contexts. Available in https://github.com/Nkluge\-correa/ethical\-problem\-solving

Power and Play: Investigating "License to Critique" in Teams' AI Ethics Discussions

David Gray Widder, Laura Dabbish, James Herbsleb +1

Past work has sought to design AI ethics interventions--such as checklists or toolkits--to help practitioners design more ethical AI systems. However, other work demonstrates how these interventions may instead serve to limit critique to that addressed within the intervention, while rendering broader concerns illegitimate. In this paper, drawing on work examining how standards enact discursive closure and how power relations affect whether and how people raise critique, we recruit three corporate teams, and one activist team, each with prior context working with one another, to play a game designed to trigger broad discussion around AI ethics. We use this as a point of contrast to trigger reflection on their teams' past discussions, examining factors which may affect their "license to critique" in AI ethics discussions. We then report on how particular affordances of this game may influence discussion, and find that the hypothetical context created in the game is unlikely to be a viable mechanism for real world change. We discuss how power dynamics within a group and notions of "scope" affect whether people may be willing to raise critique in AI ethics discussions, and discuss our finding that games are unlikely to enable direct changes to products or practice, but may be more likely to allow members to find critically-aligned allies for future collective action.

In the ever-expanding landscape of Artificial Intelligence (AI), where innovation thrives and new products and services are continuously being delivered, ensuring that AI systems are designed and developed responsibly throughout their entire lifecycle is crucial. To this end, several AI ethics principles and guidelines have been issued to which AI systems should conform. Nevertheless, relying solely on high-level AI ethics principles is far from sufficient to ensure the responsible engineering of AI systems. In this field, AI professionals often navigate by sight. Indeed, while recommendations promoting Trustworthy AI (TAI) exist, these are often high-level statements that are difficult to translate into concrete implementation strategies. There is a significant gap between high-level AI ethics principles and low-level concrete practices for AI professionals. To address this challenge, our work presents an experience report where we develop a novel holistic framework for Trustworthy AI - designed to bridge the gap between theory and practice - and report insights from its application in an industrial case study. The framework is built on the result of a systematic review of the state of the practice, a survey, and think-aloud interviews with 34 AI practitioners. The framework, unlike most of those already in the literature, is designed to provide actionable guidelines and tools to support different types of stakeholders throughout the entire Software Development Life Cycle (SDLC). Our goal is to empower AI professionals to confidently navigate the ethical dimensions of TAI through practical insights, ensuring that the vast potential of AI is exploited responsibly for the benefit of society as a whole.

Artificial intelligence (AI) has been clearly established as a technology with the potential to revolutionize fields from healthcare to finance - if developed and deployed responsibly. This is the topic of responsible AI, which emphasizes the need to develop trustworthy AI systems that minimize bias, protect privacy, support security, and enhance transparency and accountability. Explainable AI (XAI) has been broadly considered as a building block for responsible AI (RAI), with most of the literature considering it as a solution for improved transparency. This work proposes that XAI and responsible AI are significantly more deeply entwined. In this work, we explore state-of-the-art literature on RAI and XAI technologies. Based on our findings, we demonstrate that XAI can be utilized to ensure fairness, robustness, privacy, security, and transparency in a wide range of contexts. Our findings lead us to conclude that XAI is an essential foundation for every pillar of RAI.

Explaining why a database query result is obtained is an essential task towards the goal of Explainable AI, especially nowadays where expressive database query languages such as Datalog play a critical role in the development of ontology-based applications. A standard way of explaining a query result is the so-called why-provenance, which essentially provides information about the witnesses to a query result in the form of subsets of the input database that are sufficient to derive that result. To our surprise, despite the fact that the notion of why-provenance for Datalog queries has been around for decades and intensively studied, its computational complexity remains unexplored. The goal of this work is to fill this apparent gap in the why-provenance literature. Towards this end, we pinpoint the data complexity of why-provenance for Datalog queries and key subclasses thereof. The takeaway of our work is that why-provenance for recursive queries, even if the recursion is limited to be linear, is an intractable problem, whereas for non-recursive queries is highly tractable. Having said that, we experimentally confirm, by exploiting SAT solvers, that making why-provenance for (recursive) Datalog queries work in practice is not an unrealistic goal.

As Artificial Intelligence (AI) continues to advance rapidly, it becomes increasingly important to consider AI's ethical and societal implications. In this paper, we present a bottom-up mapping of the current state of research at the intersection of Human-Centered AI, Ethical, and Responsible AI (HCER-AI) by thematically reviewing and analyzing 164 research papers from leading conferences in ethical, social, and human factors of AI: AIES, CHI, CSCW, and FAccT. The ongoing research in HCER-AI places emphasis on governance, fairness, and explainability. These conferences, however, concentrate on specific themes rather than encompassing all aspects. While AIES has fewer papers on HCER-AI, it emphasizes governance and rarely publishes papers about privacy, security, and human flourishing. FAccT publishes more on governance and lacks papers on privacy, security, and human flourishing. CHI and CSCW, as more established conferences, have a broader research portfolio. We find that the current emphasis on governance and fairness in AI research may not adequately address the potential unforeseen and unknown implications of AI. Therefore, we recommend that future research should expand its scope and diversify resources to prepare for these potential consequences. This could involve exploring additional areas such as privacy, security, human flourishing, and explainability.

The need for explainable AI (XAI) is well established but relatively little has been published outside of the supervised learning paradigm. This paper focuses on a principled approach to applying explainability and interpretability to the task of unsupervised anomaly detection. We argue that explainability is principally an algorithmic task and interpretability is principally a cognitive task, and draw on insights from the cognitive sciences to propose a general-purpose method for practical diagnosis using explained anomalies. We define Attribution Error, and demonstrate, using real-world labeled datasets, that our method based on Integrated Gradients (IG) yields significantly lower attribution errors than alternative methods.

How to Assess Trustworthy AI in Practice

Roberto V. Zicari, Julia Amann, Frédérick Bruneault +17

This report is a methodological reflection on Z-Inspection$^{\small{\circledR}}$. Z-Inspection$^{\small{\circledR}}$ is a holistic process used to evaluate the trustworthiness of AI-based technologies at different stages of the AI lifecycle. It focuses, in particular, on the identification and discussion of ethical issues and tensions through the elaboration of socio-technical scenarios. It uses the general European Union's High-Level Expert Group's (EU HLEG) guidelines for trustworthy AI. This report illustrates for both AI researchers and AI practitioners how the EU HLEG guidelines for trustworthy AI can be applied in practice. We share the lessons learned from conducting a series of independent assessments to evaluate the trustworthiness of AI systems in healthcare. We also share key recommendations and practical suggestions on how to ensure a rigorous trustworthy AI assessment throughout the life-cycle of an AI system.

Benchmarks are seen as the cornerstone for measuring technical progress in Artificial Intelligence (AI) research and have been developed for a variety of tasks ranging from question answering to facial recognition. An increasingly prominent research area in AI is ethics, which currently has no set of benchmarks nor commonly accepted way for measuring the 'ethicality' of an AI system. In this paper, drawing upon research in moral philosophy and metaethics, we argue that it is impossible to develop such a benchmark. As such, alternative mechanisms are necessary for evaluating whether an AI system is 'ethical'. This is especially pressing in light of the prevalence of applied, industrial AI research. We argue that it makes more sense to talk about 'values' (and 'value alignment') rather than 'ethics' when considering the possible actions of present and future AI systems. We further highlight that, because values are unambiguously relative, focusing on values forces us to consider explicitly what the values are and whose values they are. Shifting the emphasis from ethics to values therefore gives rise to several new ways of understanding how researchers might advance research programmes for robustly safe or beneficial AI. We conclude by highlighting a number of possible ways forward for the field as a whole, and we advocate for different approaches towards more value-aligned AI research.