Research Papers

Explainable artificial intelligence (XAI) methods lack ground truth. In its place, method developers have relied on axioms to determine desirable properties for their explanations' behavior. For high stakes uses of machine learning that require explainability, it is not sufficient to rely on axioms as the implementation, or its usage, can fail to live up to the ideal. As a result, there exists active research on validating the performance of XAI methods. The need for validation is especially magnified in domains with a reliance on XAI. A procedure frequently used to assess their utility, and to some extent their fidelity, is an ablation study. By perturbing the input variables in rank order of importance, the goal is to assess the sensitivity of the model's performance. Perturbing important variables should correlate with larger decreases in measures of model capability than perturbing less important features. While the intent is clear, the actual implementation details have not been studied rigorously for tabular data. Using five datasets, three XAI methods, four baselines, and three perturbations, we aim to show 1) how varying perturbations and adding simple guardrails can help to avoid potentially flawed conclusions, 2) how treatment of categorical variables is an important consideration in both post-hoc explainability and ablation studies, and 3) how to identify useful baselines for XAI methods and viable perturbations for ablation studies.

Ethics in AI through the Practitioner's View: A Grounded Theory Literature Review

Aastha Pant, Rashina Hoda, Chakkrit Tantithamthavorn +1

The term ethics is widely used, explored, and debated in the context of developing Artificial Intelligence (AI) based software systems. In recent years, numerous incidents have raised the profile of ethical issues in AI development and led to public concerns about the proliferation of AI technology in our everyday lives. But what do we know about the views and experiences of those who develop these systems -- the AI practitioners? We conducted a grounded theory literature review (GTLR) of 38 primary empirical studies that included AI practitioners' views on ethics in AI and analysed them to derive five categories: practitioner awareness, perception, need, challenge, and approach. These are underpinned by multiple codes and concepts that we explain with evidence from the included studies. We present a taxonomy of ethics in AI from practitioners' viewpoints to assist AI practitioners in identifying and understanding the different aspects of AI ethics. The taxonomy provides a landscape view of the key aspects that concern AI practitioners when it comes to ethics in AI. We also share an agenda for future research studies and recommendations for practitioners, managers, and organisations to help in their efforts to better consider and implement ethics in AI.

Feedback in creativity support tools can help crowdworkers to improve their ideations. However, current feedback methods require human assessment from facilitators or peers. This is not scalable to large crowds. We propose Interpretable Directed Diversity to automatically predict ideation quality and diversity scores, and provide AI explanations - Attribution, Contrastive Attribution, and Counterfactual Suggestions - to feedback on why ideations were scored (low), and how to get higher scores. These explanations provide multi-faceted feedback as users iteratively improve their ideations. We conducted formative and controlled user studies to understand the usage and usefulness of explanations to improve ideation diversity and quality. Users appreciated that explanation feedback helped focus their efforts and provided directions for improvement. This resulted in explanations improving diversity compared to no feedback or feedback with scores only. Hence, our approach opens opportunities for explainable AI towards scalable and rich feedback for iterative crowd ideation and creativity support tools.

On Quantifying Literals in Boolean Logic and Its Applications to Explainable AI

Adnan Darwiche, Pierre Marquis

Quantified Boolean logic results from adding operators to Boolean logic for existentially and universally quantifying variables. This extends the reach of Boolean logic by enabling a variety of applications that have been explored over the decades. The existential quantification of literals (variable states) and its applications have also been studied in the literature. In this paper, we complement this by studying universal literal quantification and its applications, particularly to explainable AI. We also provide a novel semantics for quantification, discuss the interplay between variable/literal and existential/universal quantification. We further identify some classes of Boolean formulas and circuits on which quantification can be done efficiently. Literal quantification is more fine-grained than variable quantification as the latter can be defined in terms of the former. This leads to a refinement of quantified Boolean logic with literal quantification as its primitive.

Algorithmic and data bias are gaining attention as a pressing issue in popular press - and rightly so. However, beyond these calls to action, standard processes and tools for practitioners do not readily exist to assess and address unfair algorithmic and data biases. The literature is relatively scattered and the needed interdisciplinary approach means that very different communities are working on the topic. We here provide a number of challenges encountered in assessing and addressing algorithmic and data bias in practice. We describe an early approach that attempts to translate the literature into processes for (production) teams wanting to assess both intended data and algorithm characteristics and unintended, unfair biases.

ShieldGemma 2: Robust and Tractable Image Content Moderation

Wenjun Zeng, Dana Kurniawan, Ryan Mullins +14

We introduce ShieldGemma 2, a 4B parameter image content moderation model built on Gemma 3. This model provides robust safety risk predictions across the following key harm categories: Sexually Explicit, Violence \& Gore, and Dangerous Content for synthetic images (e.g. output of any image generation model) and natural images (e.g. any image input to a Vision-Language Model). We evaluated on both internal and external benchmarks to demonstrate state-of-the-art performance compared to LlavaGuard \citep{helff2024llavaguard}, GPT-4o mini \citep{hurst2024gpt}, and the base Gemma 3 model \citep{gemma_2025} based on our policies. Additionally, we present a novel adversarial data generation pipeline which enables a controlled, diverse, and robust image generation. ShieldGemma 2 provides an open image moderation tool to advance multimodal safety and responsible AI development.

Industry actors in the United States have gained extensive influence in conversations about the regulation of general-purpose artificial intelligence (AI) systems. Although industry participation is an important part of the policy process, it can also cause regulatory capture, whereby industry co-opts regulatory regimes to prioritize private over public welfare. Capture of AI policy by AI developers and deployers could hinder such regulatory goals as ensuring the safety, fairness, beneficence, transparency, or innovation of general-purpose AI systems. In this paper, we first introduce different models of regulatory capture from the social science literature. We then present results from interviews with 17 AI policy experts on what policy outcomes could compose regulatory capture in US AI policy, which AI industry actors are influencing the policy process, and whether and how AI industry actors attempt to achieve outcomes of regulatory capture. Experts were primarily concerned with capture leading to a lack of AI regulation, weak regulation, or regulation that over-emphasizes certain policy goals over others. Experts most commonly identified agenda-setting (15 of 17 interviews), advocacy (13), academic capture (10), information management (9), cultural capture through status (7), and media capture (7) as channels for industry influence. To mitigate these particular forms of industry influence, we recommend systemic changes in developing technical expertise in government and civil society, independent funding streams for the AI ecosystem, increased transparency and ethics requirements, greater civil society access to policy, and various procedural safeguards.

Foundation Model (FM)-based agents are revolutionizing application development across various domains. However, their rapidly growing capabilities and autonomy have raised significant concerns about AI safety. Researchers are exploring better ways to design guardrails to ensure that the runtime behavior of FM-based agents remains within specific boundaries. Nevertheless, designing effective runtime guardrails is challenging due to the agents' autonomous and non-deterministic behavior. The involvement of multiple pipeline stages and agent artifacts, such as goals, plans, tools, at runtime further complicates these issues. Addressing these challenges at runtime requires multi-layered guardrails that operate effectively at various levels of the agent architecture. Therefore, in this paper, based on the results of a systematic literature review, we present a comprehensive taxonomy of runtime guardrails for FM-based agents to identify the key quality attributes for guardrails and design dimensions. Inspired by the Swiss Cheese Model, we also propose a reference architecture for designing multi-layered runtime guardrails for FM-based agents, which includes three dimensions: quality attributes, pipelines, and artifacts. The proposed taxonomy and reference architecture provide concrete and robust guidance for researchers and practitioners to build AI-safety-by-design from a software architecture perspective.

Responding to the rapid roll-out and large-scale commercialization of foundation models, large language models, and generative AI, an emerging body of work is shedding light on the myriad impacts these technologies are having across society. Such research is expansive, ranging from the production of discriminatory, fake and toxic outputs, and privacy and copyright violations, to the unjust extraction of labor and natural resources. The same has not been the case in some of the most prominent AI governance initiatives in the global north like the UK's AI Safety Summit and the G7's Hiroshima process, which have influenced much of the international dialogue around AI governance. Despite the wealth of cautionary tales and evidence of algorithmic harm, there has been an ongoing over-emphasis within the AI governance discourse on technical matters of safety and global catastrophic or existential risks. This narrowed focus has tended to draw attention away from very pressing social and ethical challenges posed by the current brute-force industrialization of AI applications. To address such a visibility gap between real-world consequences and speculative risks, this paper offers a critical framework to account for the social, political, and environmental dimensions of foundation models and generative AI. We identify 14 categories of risks and harms and map them according to their individual, social, and biospheric impacts. We argue that this novel typology offers an integrative perspective to address the most urgent negative impacts of foundation models and their downstream applications. We conclude with recommendations on how this typology could be used to inform technical and normative interventions to advance responsible AI.

On the Limitations of Compute Thresholds as a Governance Strategy

Sara Hooker

At face value, this essay is about understanding a fairly esoteric governance tool called compute thresholds. However, in order to grapple with whether these thresholds will achieve anything, we must first understand how they came to be. To do so, we need to engage with a decades-old debate at the heart of computer science progress, namely, is bigger always better? Does a certain inflection point of compute result in changes to the risk profile of a model? Hence, this essay may be of interest not only to policymakers and the wider public but also to computer scientists interested in understanding the role of compute in unlocking breakthroughs. This discussion is timely given the wide adoption of compute thresholds in both the White House Executive Orders on AI Safety (EO) and the EU AI Act to identify more risky systems. A key conclusion of this essay is that compute thresholds, as currently implemented, are shortsighted and likely to fail to mitigate risk. The relationship between compute and risk is highly uncertain and rapidly changing. Relying upon compute thresholds overestimates our ability to predict what abilities emerge at different scales. This essay ends with recommendations for a better way forward.

Sustainability commonly refers to entities, such as individuals, companies, and institutions, having a non-detrimental (or even positive) impact on the environment, society, and the economy. With sustainability becoming a synonym of acceptable and legitimate behaviour, it is being increasingly demanded and regulated. Several frameworks and standards have been proposed to measure the sustainability impact of corporations, including United Nations' sustainable development goals and the recently introduced global sustainability reporting framework, amongst others. However, the concept of corporate sustainability is complex due to the diverse and intricate nature of firm operations (i.e. geography, size, business activities, interlinks with other stakeholders). As a result, corporate sustainability assessments are plagued by subjectivity both within data that reflect corporate sustainability efforts (i.e. corporate sustainability disclosures) and the analysts evaluating them. This subjectivity can be distilled into distinct challenges, such as incompleteness, ambiguity, unreliability and sophistication on the data dimension, as well as limited resources and potential bias on the analyst dimension. Put together, subjectivity hinders effective cost attribution to entities non-compliant with prevailing sustainability expectations, potentially rendering sustainability efforts and its associated regulations futile. To this end, we argue that Explainable Natural Language Processing (XNLP) can significantly enhance corporate sustainability analysis. Specifically, linguistic understanding algorithms (lexical, semantic, syntactic), integrated with XAI capabilities (interpretability, explainability, faithfulness), can bridge gaps in analyst resources and mitigate subjectivity problems within data.

Assessments of algorithmic bias in large language models (LLMs) are generally catered to uncovering systemic discrimination based on protected characteristics such as sex and ethnicity. However, there are over 180 documented cognitive biases that pervade human reasoning and decision making that are routinely ignored when discussing the ethical complexities of AI. We demonstrate the presence of these cognitive biases in LLMs and discuss the implications of using biased reasoning under the guise of expertise. We call for stronger education, risk management, and continued research as widespread adoption of this technology increases. Finally, we close with a set of best practices for when and how to employ this technology as widespread adoption continues to grow.