The goal of tutorials is to broaden the perspective of our interdisciplinary community, addressing practical, technical, policy, regulatory, ethical, or societal issues related to FAccT. We solicited three types of tutorials: Translation Tutorials to foster dialogue between disciplines, Implications Tutorials to describe effects of algorithmic systems in society, and Practice Tutorials focused on a specific tool or frameworks.
This tutorial introduces the Rashomon Effect: the phenomenon that many substantively different models can achieve similarly high performance on the same dataset. In high-stakes settings, this matters because models that are equally accurate may still differ in interpretability, fairness, variable importance, and even predictions. Model selection therefore cannot be treated as a purely technical question of choosing the most accurate model. It also raises questions about explanation, accountability, and which criteria should guide deployment. Aimed at an interdisciplinary FAccT audience, the tutorial translates recent technical work on the Rashomon set into terms useful for researchers, auditors, policymakers, and practitioners. We explain why multiplicity arises, how Rashomon sets can be characterized and explored in practice for interpretable model classes, and what this perspective can and cannot offer for accountability in high-stakes AI. The tutorial includes a hands-on activity where participants examine a set of near-optimal models and reflect on how different selection criteria lead to different choices.
Recent years have seen rapid progress in both fairness and explainability, yet these two areas are still too often treated separately. This tutorial argues that explainability should not be viewed only as a transparency mechanism, but also as a practical tool for fairness analysis in NLP systems. In particular, explainability can help identify when a model relies on problematic signals, reveal why disparities arise across groups, and support the design of mitigation strategies, especially in realistic settings where sensitive attributes are unavailable. The tutorial introduces two complementary families of methods: attribution-based approaches, which highlight the input features driving a prediction, and concept-based approaches, which uncover higher-level representations and internal concepts learned by the model. We show how these methods can be integrated into a fairness workflow to detect biased behaviors, analyze the concepts associated with sensitive information, and better understand the causes of model errors. The practical component of the tutorial is centered on Interpreto, an open-source library for Hugging Face models that supports both attribution-based and concept-based explainability. Through two case studies in classification and generation, participants will see how explainability can be used to investigate real fairness issues in NLP. The tutorial thus provides both a conceptual framework and reusable tools for connecting fairness evaluation with model interpretation in concrete practice.
This tutorial gives a detailed account of how the concept of ‘gaming’−a touchstone of computer science−has been adopted into digital administration and the upshots for public law. Increasingly, when applicants ask governments for basic information about the rules and algorithms that will be used in automated decision-making, they are told that they cannot access these details for fear that they will use them to ‘game the system’. The tutorial will reveal how this shifting approach to administrative decision-making, one that views hearing processes as games to be won rather than as eligibility assessment mechanisms, is a subtle but significant transformation. The tutorial will explore these effects and theorize their implications for public law and for fair decisions more broadly. The speakers will draw on empirical evidence gathered through their qualitative studies of, and legal practice involving, digitalized immigration administration in the United Kingdom and Canada to illustrate how the risk of gaming increasingly impedes access to information relevant to administrative hearings.
Guardrails – mechanisms that filter, flag, or constrain Large Language Model (LLM) inputs and outputs – have become a core component of the AI safety stack, shaping how AI systems behave in chatbots, social media platforms, and public-facing services. Yet while the research community devotes significant attention to evaluating the outputs and capabilities of LLMs themselves, the guardrail models that govern what users actually see and experience have received less scrutiny. We argue that evaluating guardrails is as important as evaluating the LLMs they protect. Historically, guardrails have been difficult to examine: often they performed as proprietary classifiers embedded in content moderation systems, visible to users only through refusals or minimally explained filtering decisions. The recent proliferation of open-source guardrail models – with accessible architectures, chain-of-thought reasoning traces, and open safety taxonomies – has created an opportunity for independent evaluation. This tutorial invites FAccT researchers to critically examine guardrails across languages, domains, and deployment contexts. Drawing on our multilingual, context-aware evaluation of guardrails in a humanitarian setting and on the design of Mozilla AI’s open-source any- guardrail framework, we show how guardrails can be systematically and contextually evaluated; through an interactive tutorial activity, we show that others can do this too in their own context and language of interest. We also identify key research frontiers in LLM guardrails design, including dynamic safety taxonomies, agentic capabilities such as tool use and retrieval for fact-checking and trustworthiness, multimodal coverage, and the creation of feedback loops between evaluation results and guardrail improvement.
Private governance approaches - standards, methods and incentives in place outside of governmental regulatory frameworks - work in parallel to regulation to help industries coalesce around responsible and safe practices. Private governance includes mechanisms like assurance and certification, insurance and liability (including private litigation), due-diligence frameworks, venture investment, professionalization, and standardized procurement processes. These are common in other sectors and industries, like food safety, healthcare or aviation (Lin, 2014), but remain nascent in AI. Many who have gestured at the importance of, for example, insurance in advancing AI governance do not have a complete understanding of how insurers may make key determinations of risk factors, or how private governance is distinct from industry self-regulation. There is a clear need for disambiguation around these mechanisms, how they work in practice, and what they might achieve. In this tutorial, we will collaboratively explore the dimensions of assurance, liability, and insurance, and how they might interplay to support public interest AI governance goals. The tutorial will create a dialogue between sociotechnical, policy/legal, and private governance experts through a series of lightning talks, a panel discussion, and audience discussion, to advance ecosystem capacity. Tutorial participants will gain perspectives on the role of extra-governmental mechanisms that can impact the development and deployment of AI systems.
AI systems now allocate credit, inform bail decisions, screen job applicants, and triage patients—at population scale, often invisibly. Countries responded unevenly: through data subject rights, risk classification schemes, transparency mandates, and fiduciary duty proposals. Each response embeds a theory of the AI–person relationship that generates characteristic strengths and structural blind spots. Researchers and system designers who engage with fairness, accountability, and transparency without understanding those legal theories work in a regulatory vacuum, missing both the constraints their systems face and the genuine gaps their technical work could address. This tutorial translates four dominant legal models of AI governance. Drawing on comparative analysis across fifteen jurisdictions and regulatory instruments—from the EU AI Act to the Colorado Artificial Intelligence Act, from the GDPR to the UK Algorithmic Transparency Recording Standard—it provides participants with a map of what existing law can and cannot demand from technical systems, and where the most consequential regulatory gaps remain.
A growing number of journalists are investigating the impacts of AI across the whole supply chain, from automated decision-making systems in the public sector, to data centers, to issues of labor. These investigations touch upon topics of AI Accountability that academics also tackle, with complementary perspectives and methods. Successful collaboration with journalists is often defined as secondary coverage of existing research, but there are deeper models of collaboration that can generate more attention and impact. However, it can be hard to know how to work with investigative journalists in practice. Where to begin? What are the mutual benefits of working together? What are common challenges? How to ensure collaborations go well, and conciliate the different timings between journalistic investigations and academic research? This tutorial, delivered by investigative journalists who regularly collaborate with academics and civil society and based on concrete examples from their own AI accountability investigations, is geared towards members of the FAccT community who would like to work in coalition with journalists on AI accountability issues, from being public partners on stories to expert allies behind the scenes. At the end of the tutorial, participants will know how investigative journalists operate; how to build trusted relationships with them; and how to collaborate at different stages of an investigation in a way that’s mutually beneficial, ethical and impactful.
There is a ready-made set of statistical tools for quantifying the uncertainty of the kinds of metrics used to measure and characterize fairness in machine learning, such as confidence intervals for ratios of binomial proportions (called ``risk ratios'' or ``relative risk'' in biostatistics). Currently, none of the major software toolkits for auditing machine learning models include uncertainty quantification. In this tutorial, we will summarize the existing methods, show relevant statistical packages and present sample code for using them and producing visualizations. This is a straightforward intervention that will help improve the rigor of the quantitative side of `algorithmic auditing' work.
Multimodal vision-language models are increasingly deployed in high-stakes settings, which makes understanding the biases they encode both a technical and ethical priority. This tutorial addresses that challenge directly, using counterfactual fairness as its core evaluative framework. The concept of counterfactual fairness is familiar in machine learning: it captures the intuitive idea that if decisions should not be made on the basis of protected characteristics (race, gender, disability, age, and others) then changing those attributes in the input to a model should not change the output. This tutorial focuses on the question of how we can develop and use counterfactual image datasets to assess bias in large vision-language models (LVLMs). Our discussion will begin with the process of image dataset creation and the trade-offs of collecting versus generating images. We will summarize which protected characteristics have been studied (and overlooked) in the literature, and provide examples of numerical, closed-class, and open-ended generation tasks and corresponding bias metrics. Furthermore, we will address some of the more contested questions around counterfactual fairness, such as whether mapping social constructs like race and gender onto discrete counterfactual classes risks reinforcing the very categories we seek to interrogate, and whether the use of synthetic images introduces new sources of bias to the analysis. The session will culminate in a hands-on session so that attendees can test and critique the methodology for themselves. This tutorial is aimed at researchers and practitioners in all disciplines — no prior background knowledge is required.
Researchers working with pervasive data (e.g., public social media posts, behavioral traces, scraped web content) face research ethics questions that compliance-based ethics review processes are often poorly equipped to address. This tutorial introduces the PERVADE Data Ethics Decision Support Tool, a free, browser-based resource that guides researchers through structured ethical reflection across the full research life cycle. Grounded in empirical research about both researcher challenges and the perceptions of people whose data is being collected, the tool centers the dimensions of awareness and power, drawn in part from traditions in ethnographic research. Attendees will work hands-on with the tool using case studies, hypotheticals, and their own examples, and participate in open discussion about gaps and opportunities in research ethics and community-led development of this tool and future resources.
This practice tutorial introduces the Spanish-language Algorithmic Impact Assessment Tool developed by GobLab UAI, a structured, step-by-step instrument designed specifically for public-sector AI governance. The tool emerged from a multi-year collaboration with public agencies in Chile, integrating international best practices and extensive user-centered design with government teams. This tool was published in October 2024 and is currently in its fourth version. The tool assesses an impact level (low, moderate, high, very high) for automated decision-making systems including AI and produces an automated report with recommendations to mitigate the identified risks. Participants will apply sections of the tool to a real public sector use-case and gain first-hand experience in integrally assessing impacts during project design.
Many automated decision systems (ADS) are designed to solve prediction problems, where the goal is to learn patterns from a sample of the population and apply them to individuals from the same population. In reality, these prediction systems operationalize holistic policy interventions in deployment. Once deployed, ADS can shape impacted population outcomes through an effective policy change in how decision-makers operate. Drawing on a collaborative white paper co-authored by over 30 researchers spanning computer science, statistics, economics, law, sociology, and philosophy, this tutorial argues that responsible ADS deployment requires a paradigm shift: from optimizing prediction accuracy to formally understanding these systems as interventions with measurable downstream effects. Across three modules - model design, evaluation science, and implementation science - attendees will learn to translate between the prediction-centric vocabulary of ML and the intervention-centric vocabulary of policy and program evaluation. Concretely, this means understanding tradeoffs for targeting based on predicted risk versus causal treatment effects, going beyond benchmarking ADS to assess real-world impact through causal experimentation, and accounting for human discretion, feedback loops, and regulatory context in institutional deployments. Structured cross-community dialogue prompts will connect participants across disciplines and encourage collaboration on open research directions.
AI supply or value chains introduce technical, ethical, legal, and regulatory challenges, and—unlike their manufacturing counterparts—lack the modularity, redundancies, and transparency needed to easily correct failures. Tools the FAccT and broader ML communities have pioneered for auditing individual algorithms often falter when applied to the opaque networks of outsourced data, proprietary APIs, third-party tooling, and compositional deployments that define modern AI. This 60-minute tutorial introduces participants to AI supply chains and to the practical challenges of studying, auditing, designing, and building systems composed from multiple AI components. Through two moderated panels—one on Power, Regulation, and Law in AI Supply Chains, and one on Challenges in Studying, Building, and Tooling AI Supply Chains—we bring together researchers studying AI infrastructure, developers building composite systems, and scholars analyzing power and governance in AI ecosystems. Participants will leave able to identify barriers to governance and auditing across supply chain actors, point to grounded case studies, articulate technical barriers to studying these systems, and connect with others working in this space.
This tutorial introduces efforts by legal advocates acting on behalf of victims and survivors of technology-related harms to reimagine the remedial infrastructure surrounding tech accountability, moving from individualized, piecemeal redress through private civil litigation to a comprehensive service model adapted from cross-disciplinary initiatives in population-wide harm reduction. The law provides narrow relief for people who experience the categories of harm that we document through legal intake and representation (e.g. social media, gig and platform workers, AI companion chatbots, site-based impacts of data centers, algorithmic discrimination, or benefits technology). Most legal service providers offer conventional approaches to client care, following professional standards in managing representation, without assessing or meeting the need for holistic, transdisciplinary services. When persons seeking redress are successful in obtaining legal representation and necessary social services, provision of these resources is individualized. Many people yearn for peer communities in which they can show up, speak their truth, and be believed. To address these gaps, TJL and Temple have developed a prototype access to justice intake mechanism to support victims and survivors of tech harm. This approach is inspired by trauma-informed case management, holistic approaches to criminal legal defense, and medical-legal partnerships (MLPs) to provide a “one-stop shop” point of entry into the provision of client services. Our goal is to consolidate our work in a practical tutorial session that engages the FAccT community particularly medical and social work practitioners, AI and social media accountability researchers, and citizen scientists—in thoughtful examination of central questions in designing our future work. We will examine case studies, consider a brief landscape of civil and human rights litigation over tech harms and unmet non-legal needs, and workshop recommendations on how MLP models could serve as inspiration for a structured, population-level response to the needs of victims and survivors of tech harms.
This tutorial examines how lessons from financial regulation can inform contemporary approaches to AI governance, with particular attention to what counts as persuasive regulatory evidence under principles‑based regimes. Drawing on the long‑standing shift from rules‑based to principles‑based regulation in financial services, the session explores how regulators evaluate governance arrangements, organisational processes, and substantive outcomes rather than narrow technical compliance. Financial services provide a valuable case study because they represent one of the most mature and socially consequential experiments with principles‑based oversight, especially in consumer protection. The tutorial connects this regulatory history to emerging AI governance frameworks, where high‑level principles such as fairness, transparency, and accountability increasingly shape regulatory practice. Building on interdisciplinary research in law, AI ethics, and policy, the session highlights how evidential standards can privilege certain actors and forms of knowledge while marginalising others. Through structured discussion and applied examples, participants gain tools to analyse how evidence, power, and institutional capacity shape the fairness and accountability of AI systems in regulatory practice.
Writing research articles is a different skill from producing effective technology policy products/artifacts. This tutorial is designed for researchers who are interested in having impact in the technology policy space. We will give specific examples of a wide variety of technology policy products including responses to government “request for comments” calls, proactive statements of principles, fact-based technology briefs, testimony before a legislative body, amicus briefs to a court, and many more. We will also highlight evidence of the impact that past technology policy products have had. As one concrete example, ACM’s Europe Technology Policy Committee (Europe-TPC) submitted 14 recommendations for the EU Code of Practice and 8 recommendations were incorporated and 4 partially incorporated. Producing technology policy products can lead to invitations for researchers to work directly in government for a time and we will describe some concrete examples of this. We will focus on conveying the real-world impacts that specific policy products have had, how those impacts were achieved, and how best practices can vary around the world.
While Responsible AI audits are widely accepted as critical aspects of AI governance, they are most often conducted in a context of mutual trust, responsiveness, and cooperation. However, companies and organizations may lie, hide information, or neglect some aspect of the audit to make their AI system appear safer, more ethical, or more sustainable than it actually is, as illustrated by historical scandals in more established industries, such as the "Dieselgate". Indeed, if proper care is not taken to ensure the trustworthiness of Responsible AI audits in the face of dishonesty and negligence, these audits may be misused for ethics-washing instead of ensuring accountability and preventing harm. The goal of this tutorial is to foster a cross-disciplinary dialogue around a central question: what are the possibilities of cheating Responsible AI audits and how can they be reconciled? We address this by examining audits at the AI component (e.g., a model) and system level and bridging perspectives from multiplicity, cryptography, system safety, and measurement theory.
In this dialogue/translation tutorial, we call the attention of FAccT and allied fields to the “implosion” of research around AI alignment; the sociotechnical work of constraining AI systems to follow given human objectives and values. Although this research has produced useful insights, it remains fragmented across disciplines, making it difficult to see how different interventions relate or where critical gaps persist. We develop a three-level framework, spanning model, interaction, and governance, and invite participants to collaboratively trace the techniques and practices that conventional surveys of alignment might miss. Participants will leave with a shared vocabulary for locating their own work within the broader alignment landscape, whether or not they currently use the term “alignment,” and concrete directions for cross-disciplinary collaboration on alignment research and practice.
Existing model evaluation results are scattered across leaderboards, papers, and technical reports in incompatible formats. This fragmentation obscures transparency, hinders progress, and disadvantages researchers, civil society, policymakers, and industry alike, especially those who can't afford to run evaluations from scratch. Built once, shared eval infrastructure serves us all. In this tutorial, we walk through Every Eval Ever, a community-governed open source infrastructure that unifies all evaluation results under a shared metadata schema. We then present Evaluation Cards, an interface and interpretive layer for evaluation reporting designed around practitioner needs from stakeholder interviews, and show how participants can find, compare, and contribute evaluations themselves. All technical experience levels are welcome. If you can, please bring a laptop or tablet!
In recent years, the term “Global South” has increasingly been adopted in discussions on building globally inclusive AI systems and governance. However, much work remains to understand the term’s connotations, usage, and contradictions within AI ethics. In this tutorial, we will share our recent research on the term “Global South,” based on interviews with 20 scholars and practitioners in AI ethics who have engaged with global politics. We found that the term often implies harmful stereotypes of homogeneity, underdevelopment, and technological illiteracy, and can thus perpetuate an imperial gaze similar to the terms “Third World” and “developing countries.” Despite these implications, many scholars and practitioners feel pressured to use the term “Global South” due to research and funding structures that center the United States as a hub for resources and knowledge production. Rather than simply adopting another term that may carry similar stereotypes, we emphasize the need to ground AI ethics work in specific regions and power structures, use particular analytical frameworks when drawing connections across communities, and make deeper changes to build alternative funding structures. The tutorial will include a discussion in which participants reflect on their own experiences with the power dynamics of the term “Global South” and explore interventions that challenge these dynamics.
Research in the FAccT community often engages directly with questions of real-world impact in sociotechnical systems. These findings are typically communicated through academic publications at FAccT or similar venues, which are integral to advancing interdisciplinary knowledge and allow researchers to share and build on each other’s work. At the same time, this work is often not accessible to affected communities, policymakers, and practitioners who could act on these findings, limiting its potential to inform policy and practice. As a result, even policy-relevant research may have limited influence on real-world decision-making without additional forms of engagement with these critical stakeholders.
This tutorial introduces policy briefs as one practical mechanism for translating research findings into formats that can reach these audiences to bridge this gap. Policy briefs synthesize evidence for pressing policy-relevant issues and support decision-making by offering clear options and recommendations. These briefs are written in a very specific style, which is counterintuitive for academics. Instead of leading with the scientific method, hypothesis formulation and assessment, and uncertainty (as we do in academic papers), policy briefs require confident, upfront, and concise descriptions of policy solutions.
The tutorial is structured as an interactive, hands-on session. By the end of the session, participants will have a concrete starting point for a policy brief and a clearer understanding of how to translate FAccT research into forms that can inform policy and practice. More broadly, this tutorial aims to build capacity within the community to engage more effectively with policy processes while maintaining the rigor and values that define FAccT research. https://facct26policytutorial.github.io/"Consumer AI products are deployed in care domains such as parenting, mental health, eldercare. The stakes of a wrong answer are not embarrassment but harm. The dominant paradigm treats safety as a guardrail problem: filter inputs, constrain outputs, refer users to professionals. This tutorial argues guardrail-based safety does not survive the real distribution of users in care domains, and proposes a multi-tier architecture in which trust, safety, and escalation are first-class design concerns under deterministic constraints.
Juni Parenting App, an AI companion for parents of infants, deploys this architecture. Juni treats safety as a stack, not a layer. A graded content trust hierarchy governs which sources shape which responses. A deterministic orchestrator sits in front of the language model and enforces an invariant preventing lower-trust content from upgrading higher-trust pathways. A three-tier escalation policy distinguishes responses bypassing the language model, those generated under retrieval constraints, and those acknowledging wellness signals and surface support resources. Each tier handles a class of failure that a single guardrail cannot. The architecture is evaluated against a comprehensive functional test framework.
The tutorial walks through each tier with live examples, then surfaces a class of failure that no architecture alone can resolve: the system functions correctly but the resources it refers users to clinicians, warmlines, emergency services, are themselves inaccessible to populations, including immigrant parents, parents without primary care access, and parents navigating healthcare in a second language. We frame this gap “Escalation Without Infrastructure” as a structural problem the field has under-acknowledged.
Participants will leave with a working vocabulary for evaluating consumer AI in care domains: how its architecture treats trust, how it constrains generation, how it escalates, and where its responsibilities end. It applies wherever LLMs touch care decisions: from mental health chatbots to eldercare assistants to health-adjacent AI for newborn care.