By Margaret Hagan, Executive Director of the Legal Design Lab
At this year’s International Conference on Artificial Intelligence and Law (ICAIL 2025) in Chicago, we co-hosted the AI for Access to Justice (AI4A2J) workshop—a full-day gathering of researchers, technologists, legal practitioners, and policy experts, all working to responsibly harness artificial intelligence to improve public access to justice.
The workshop was co-organized by an international team: myself (Margaret Hagan) from Stanford Legal Design Lab, Quinten Steenhuis of Suffolk University Law School/LIT Lab; Hannes Westermann of Maastricht University; Marc Lauritsen, Capstone Practice Systems; and Jaromir Savelka of Carnegie Mellon University. Together, we brought together 22 papers from contributors across the globe, representing deep work from Brazil, Czechia, Singapore, the UK, Canada, Italy, Finland, Australia, Taiwan, India, and the United States.
A Truly Global Conversation
What stood out most was the breadth of global participation and the specificity of solutions offered. Rather than high-level speculation, nearly every presentation shared tangible, grounded proposals or findings: tools developed and deployed, evaluative frameworks created, and real user experiences captured.
Whether it was a legal aid chatbot deployed in British Columbia, a framework for human-centered AI development from India, or benchmark models to evaluate AI-generated legal work product in Brazil, the contributions showcased the power of bottom-up experimentation and user-centered development.
A Diversity of Roles and Perspectives
Participants included legal researchers, practicing attorneys, judges, technologists, policy designers, and evaluation experts. The diversity of professional backgrounds allowed for robust discussion across multiple dimensions of justice system transformation. Each participant brought a unique lens—whether from working directly with vulnerable litigants, building AI systems, or establishing ethical and regulatory frameworks for new technologies.
Importantly, the workshop centered interdisciplinary collaboration. It wasn’t just legal professionals theorizing about AI, or technologists proposing disconnected tools. Instead, we heard from hybrid teams conducting qualitative user research, sharing open-source datasets, running field pilots, and conducting responsible evaluations of AI interventions in real-world settings.
Emerging Themes Across the Day
Across four themed panels, several core themes emerged:
- Human-Centered AI for Legal Aid and Self-Help
Projects focused on building AI copilots and tools to support legal aid organizations and self-represented litigants. Presenters shared tools to help tenants facing eviction, systems to automate form filling with contextual guidance, and bots that assist in court navigation. Importantly, these tools were being built in partnership with legal aid teams and directly with users, with ongoing evaluations of quality, safety, and impact. - Legal Writing, Research, and Data Tools
A second group of projects explored how AI could help professionals and SRLs write legal documents, draft arguments, and find relevant precedent more efficiently. These systems included explainable outcome predictors for custody disputes, multilingual legal writing assistants, and knowledge graphs built from court filings. Many papers detailed methods for aligning AI output with local legal contexts, language needs, and cultural sensitivity. - Systems-Level Innovation and AI Infrastructure
A third set of papers zoomed out to the system level. Projects explored how AI could enable better triage and referral systems, standardized data pipelines, and early intervention mechanisms (e.g., detecting legal risk from text messages or scanned notices). We also heard from teams building open-source infrastructure for courts, public defenders, and justice tech startups to use. - Ethics, Evaluation, and Responsible Design
Finally, the workshop closed with discussions of AI benchmarks, regulatory models, and ethical frameworks to guide the development and deployment of legal AI tools. How do we measure the accuracy, fairness, and usefulness of a generative AI system when giving legal guidance? What does it mean to provide “good enough” help when full representation isn’t possible? Multiple projects proposed evaluation toolkits, participatory design processes, and accountability models for institutions adopting these tools.
Building on Past Work and Sharing New Ideas
Many workshop presenters built directly on prior research, tools, and evaluation methods developed through the Legal Design Lab and our broader community. We were especially excited to see our Lab’s Legal Q&A Evaluation Rubrics, originally developed to benchmark the quality of automated legal information, being adopted in People’s Law School’s (in British Columbia) Beagle+ project as they deploy and test a user-facing AI chatbot to answer people’s common legal questions.
Another compelling example came from Georgetown University, where our previous Visual Legal design work product, patterns and communication tools are now inspiring a new AI-powered visual design creator built by Brian Rhindress. Their tool helps legal aid organizations and court staff visually and textually explain legal processes to self-represented litigants—leveraging human-centered design and large language models to generate tailored explanations and visuals. A group can take their text materials and convert it into human-centered visual designs, using LLMs + examples (including those from our Lab and other university/design labs).
We’re excited to see these threads of design and evaluation from previous Stanford Legal Design Lab work continuing to evolve across jurisdictions.
The Need for Empirical Grounding and Regulatory Innovation
A major takeaway from the group discussions was the urgent need for new empirical research on how people actually interact with legal AI tools—what kinds of explanations they want, what kinds of help they trust, and what types of disclosures and safeguards are meaningful. Rather than assuming that strict unauthorized practice of law (UPL) rules will protect consumers, several papers challenged us to develop smarter, more nuanced models of consumer protection, ones grounded in real user behavior and real-world harms and benefits.
This opens the door for a new generation of research—not just about what AI can do, but about what regulatory frameworks and professional norms will ensure the tools truly serve the public good.
Highlights
There were many exciting contributions among the 22 presentations. Here is a short overview, and I encourage you to explore all the draft papers.
Tracking and Improving AI Tools with Real-World Usage Data: The Beagle+ Experiment in British Columbia
One of the standout implementations shared came from British Columbia’s People’s Law School’s Beagle+ project. This legal chatbot, launched in early 2024, builds on years of legal aid innovation to offer natural-language assistance to users navigating everyday legal questions. What makes Beagle+ especially powerful is its integrated feedback and monitoring system: each interaction is logged with Langfuse, recording inputs, outputs, system prompts, retrieval sources, and more.
The team uses this real-world usage data to monitor system accuracy, cost, latency, and user empowerment over time—allowing iterative improvements that directly respond to user behavior.
They also presented experiments in generative legal editing, exploring the chatbot’s ability to diagnose or correct contract clauses, with promising results. Yet, the team emphasized that no AI tool is perfect out of the box—for now, human review and thoughtful system design remain essential for safe deployment.
Helping Workers Navigate Employment Disputes in the UK: AI-Powered ODR Recommenders
Glory Ogbonda and Sarah Nason presented a pioneering tool from the UK designed to help workers triage their employment disputes and find the right online dispute resolution (ODR) system. Funded by the Solicitors Regulation Authority, this research uncovered what users in employment disputes really want: not just legal signposting, but a guided journey. Their proposed ODR-matching system uses RAG (retrieval-augmented generation) to give users an intuitive flow: first collecting plain-language descriptions of their workplace conflict, then offering legal framing, suggested next steps, and profiles of potential legal tools.
User testing revealed a tension between formal legal accuracy and the empathy and clarity that users crave. The project underscores a core dilemma in legal AI: how to balance actionable, user-centered advice with the guardrails of legal ethics and system limits.
Empowering Litigants through Education and AI-Augmented Practice: The Cybernetic Legal Approach
Zoe Dolan (working with Aiden the AI) shared insights from her hands-on project working directly with self-represented litigants in an appeals clinic. Dolan + Aiden trained cohorts of participants to use a custom-configured GPT-based tool, enhanced with rules, court guides, and tone instructions. Participants learned how to prompt the tool effectively, verify responses, and use it to file real motions and navigate the courts.
The project foregrounds empowerment, rather than outcome, as the key success metric—helping users avoid defaults, feel agency, and move confidently through procedures. Notably, Dolan found that many SRLs had developed their own sophisticated AI usage patterns, often outpacing legal professionals in strategic prompting and tool adoption. The project points to a future where legal literacy includes both procedural knowledge and AI fluency.
AI-Driven Early Intervention in Family and Housing Law in Chicago
Chlece Walker-Neal presented AJusticeLink, a preventative justice project from Chicago focused on identifying legal and psycho-legal risk through SMS messages. The tool analyzes texts that users send to friends, family, and others, detecting language that signals legal issues—such as risk of eviction or custody disputes—and assigning an urgency score. Based on this, users are linked to appropriate legal services. The project aims to intervene before a crisis reaches the courthouse, helping families address issues upstream. This early-warning approach exemplifies a shift in justice innovation: from reactive court services to proactive legal health interventions.
PONK: Helping Czech Litigants Write Better Legal Texts with Structured AI Guidance
The PONK project out of Czechia presented a tool for improving client-oriented legal writing by using structured AI and rule-based enhancements. Drawing on a dataset of over 250 annotated legal documents, the system helps convert raw legal text into clearer argumentation following a Fact–Rule–Conclusion (FRC) structure. This project is part of a broader movement to bring explainable AI into legal drafting and aims to serve both litigants and legal aid professionals by making documents more structured, persuasive, and usable across a wider audience. It showcases how small linguistic and structural refinements, guided by AI, can produce outsized impact in real-world justice communications.
Can AI Help Fix Hard-to-Use Court Forms and Text-Heavy Guides? A Visual Design AI Prototype
Brian Rhindress presented a provocative question: can AI be trained to reformat legal documents and court forms into something more visually accessible? And could the backlog of training materials in PDFs, docs, and text-heavy powerpoints be converted into something more akin to comic books, visuals, fliers, and other engaging materials?
Inspired by design principles from Stanford Legal Design Lab and the U.S. Digital Service, and building off materials from the Graphic Advocacy Project, Harvard A2J Lab, Legal Design Lab and more, the project tested generative models on their ability to re-layout legal forms. While early versions showed promise for ideation and inspiration, they often suffered from inconsistent checkbox placement, odd visual hierarchy, or poor design language.
Still, the vision is compelling: a future where AI-powered layout tools assist courts in publishing more user-friendly, standardized forms—potentially across jurisdictions. Future versions may build on configuration workflows and clear design templates to reduce hallucinations and increase reliability. The idea is to lower the entry barrier for underserved communities by combining proven legal messaging with compelling visual storytelling. Rather than developing entirely new tools, teams explored how off-the-shelf systems, paired with smart examples and curated prompts, can deliver real-time, audience-tailored legal visuals.
Building Transparent, Customizable AI Systems for Sentencing and Immigration Support
Aparna Komarla from Redo.io and colleagues from OpenProBono demonstrated the power of open, configurable AI agents in the justice system. In California’s “Second Look” sentencing reviews, attorneys can use this custom-built AI system to query multi-agency incarceration datasets and assign “suitability scores” to prioritize eligible individuals who might have claims the attorneys can assist with. The innovation lies in giving attorneys—not algorithms—the power to define and adjust the weights of relevant factors, helping to maintain transparency and align tools with local values and judicial discretion.
Place Matters: How Location Affects AI Hallucination Rates in Legal Answers
Damian Curran and colleagues explored an increasingly urgent issue: do large language models (LLMs) perform differently depending on the geographic context of the legal question? Their findings say yes—and in sometimes surprising ways. For instance, while LLMs hallucinated less often on employment law queries in Sydney, their housing law performance there was riddled with errors. In contrast, models did better on average with Los Angeles queries—possibly due to the volume of U.S.-based training data. The study underscores the importance of localization in AI legal tools, especially for long-tail or low-resourced jurisdictions where statutory nuance or recent reforms may not be well represented in AI training data.
Drawing the Line Between Legal Info and Legal Advice in India’s Emerging Chatbot Landscape
Avanti Durani and Shivani Sathe presented a critical user research study from India that investigates how AI-powered legal chatbots respond to user queries—and whether they stay within the bounds of offering legal information rather than unlicensed advice. Their analysis of six tools found widespread inconsistencies in tone, disclaimers, and the framing of legal responses. Some tools subtly slipped into strategic advice or overly narrow guidance, even as disclaimers were buried or hard to find. These gaps, they argue, pose real risks for low-literacy and legally vulnerable users. Their work raises important regulatory questions: should the standard for chatbots be defined only by unauthorized practice of law rules? Or should we also integrate user preferences and the expectations of trusted community intermediaries, such as social workers or legal aid navigators?
The full collection of 22 papers is available here with links to preprint drafts as available. We encourage everyone to explore the work and reach out to the authors—many are actively seeking collaborators, reviewers, and pilot partners.
Together, these contributions mark a new chapter in access to justice research—one where AI innovation is rigorously evaluated, deeply grounded in the legal domain, and shaped by the real needs of the people and professionals who use it.
What Comes Next
The enthusiasm and rigor from this year’s submissions reaffirmed that AI for access to justice is not a hypothetical field—it’s happening now, and it’s advancing rapidly.
The ICAIL AI4A2J workshop served as a global convening point where ideas were shared not just to impress, but to be replicated, scaled, and improved upon. Multiple projects made their datasets and prototypes publicly available, inviting others to test and build on them. Several are looking to collaborate across jurisdictions and domains to study effectiveness in new environments.
Our Stanford Legal Design Lab team left the workshop energized to continue our own work on AI co-pilots for eviction defense and debt relief, and newly inspired to integrate ideas from peers across the globe. We’re especially focused on how to:
- Embed evaluation and quality standards from the start,
- Design human-AI partnerships that support (not replace) frontline legal workers,
- Spread and scale the best tools and protocols in ways that preserve trust, dignity, and legal integrity.
- Develop policies and regulation that are based in empirical data, human behavior, and actual consumer protection
Thank You
We’re deeply grateful to our co-organizers and all the presenters who contributed to making this workshop a meaningful step forward. And a special thanks to the ICAIL community, which continues to be a space where technical innovation and public interest values come together in thoughtful dialogue.
Stay tuned—our program committee is considering next steps around publications and subsequent conferences, and we hope this is just the beginning of an ongoing, cross-border conversation about how AI can truly improve access to justice.
Please also see my colleague Quinten’s write-up of his takeaways from the workshop!