Each year, large language models get more powerful, and we’re racing to find new ways to use them. In the world of access to justice, this new research shows a growing maturity. Moving beyond tentative “what can it do” explorations, the latest academic research is trying out bold new efforts at reshaping how justice works and help for self-represented litigants happens.
Our lab helped organize the fifth edition of an academic AI and access to justice workshop in Turin, Italy as part of the JURIX conference on legal knowledge systems. A few highlights included:
- A tool in the Los Angeles court to improve the mandated review of defaults in consumer debt cases, preventing litigants from losing their court cases without a chance to be heard. Another paper, a follow-up to the workshop at ICAIL in Chicago this summer, focused on applying AI to resentencing in California.
- Voice-first, telephone-tied, AI to close the digital access gap in the predominantly rural African country of Niger. Another paper focused on usability and adoption risks in refugee camps in Tanzania.
- A study of differential gender bias across AI models in assigning parenting rights in Czechia. Claude Haiku 4.5, perhaps surprisingly, showed the strongest gender bias.
- A preliminary study of how judges talk about AI in the European Court of Human Rights, revealing a focus on risk and threats to state power.
- A new taxonomy of legal tasks to help make sure we’re all talking about the same thing when we benchmark and talk about what legal AI can help with.
- Several papers talked about applying AI to legal outcome predictions, potentially improving settlement outcomes. A research study in Chile showed that this approach has limits; offering litigants the extra information about a likely outcome did not significantly change their willingness to settle their cases. This tied to a paper in the main conference showing a similar result in Brazilian legal courts.
- One paper introduced and evaluated the ability of open-source AI agents to carry out legal help tasks such as finding legal help, filling in forms, and gathering legal information, raising questions about the future value of dedicated tools across these tasks.
In our broader world, we have to acknowledge that the initial buzz and excitement over new large language models is joined by a growing concern. Some folks are getting sick of hearing about generative AI’s wonders, and instead focusing on AI’s environmental impact, hallucinations, access to technology, and job replacement.
It’s no surprise that legal aid attorneys are among the fastest adopters of generative AI. Unlike private businesses, legal aid attorneys do not need to connect work to billable hours.
Yet as researchers, we face a tension: the ability of LLMs to replace jobs both creates new ways to close the gap for the millions of Americans without access to an attorney’s help and makes it harder for ordinary people to earn money through work. We live in the lacuna between job displacement and a hoped-for reduction in unmet legal need. To the extent that LLMs contain embedded biases, we risk perpetuating them without careful design. To the extent that we release systems that aren’t carefully validated for accuracy, we risk feeding self-represented litigants misinformation.
As this broader discussion is going on, we have to recognize that the general public, more and more, is relying on ChatGPT when they encounter a legal problem without the intermediation of any of our carefully built and tested systems. An access to justice maximalist approach, therefore, has to reckon with the risk of not moving as well as the risk of going too quickly with providing AI-assisted legal help.
My view: proceed carefully, with testing and clear understanding of the risks of failures, but embracing the best and strongest case for using generative AI whenever possible.
During the workshop, researchers confronted these fears head on, in the papers highlighted above and in the full list described below.
The AI for Access to Justice workshop joined up at JURIX 2025 with two other workshops: AI for Dispute Resolution and AI for Data Access and Justice. We shared a single pool of papers and an eager audience. Like past editions of this workshop, our topic has drawn an unusual level of interest and selectivity. We accepted 19 papers and rejected 6; academic workshops often only attract a handful of papers. This shows strong demand for applied work in the field and perhaps also the interdisciplinary nature of the organizing team.
Here is the full list of papers presented in the workshop:
1. Understanding Rights Through AI: The Role of Legal Chatbots in Access to Justice argues that AI-powered chatbots can democratize access to justice by serving as low-cost entry points for vulnerable populations, particularly girls and women in African regions, to understand and assert their legal rights outside of formal court settings.
2. The Private Family Forecast: a predictive method for an effective and informed access to justice proposes an AI-driven predictive method for family law intended to provide clients with statistical information about likely judicial outcomes, thereby supporting informed decision-making and fostering consensual negotiated solutions to overcome financial, temporal, and emotional barriers to justice.
3. Artificial intelligence enabled justice tools for refugees in Tanzania: a feasibility study from Nyaragusu camp performs a feasibility analysis to determine the factors—including perceived ease of use, usefulness, trust, security, awareness, and influence—that affect the intention to adopt mobile justice services in a refugee camp setting.
4. What Legal Help Teams and Consumers Actually Do: A Legal Help Task Taxonomy to Accelerate Research and Innovation introduces a comprehensive, structured Legal Help Task Taxonomy, encompassing seven high-level categories of activities performed by legal aid staff, court navigators, and consumers to create a shared vocabulary for accelerating rigorous research, clear technology design, and interoperability in the access-to-justice field.
5. Artificial intelligence and access to justice in Chile: Evaluating its potential role in conciliation procedures empirically analyzes environmental damage claims in Chile, finding that the mandatory conciliation phase is excessively long and often unsuccessful, and proposes using targeted AI-based judicial outcome prediction tools to enhance the efficiency of negotiation and conciliation.
6. Packaging Thematic Analysis as an AI Workflow for Legal Research introduces a system that operationalizes qualitative thematic analysis for legal research, particularly for case law, as a reproducible, instruction-responsive, and agentic workflow with human checkpoints to reduce the time required to derive workable themes across different languages and jurisdictions.
7. Designing Clarity with AI: Improving the Usability of Case Law Databases of International Courts investigates how AI and user experience (UX) design principles can improve the accessibility and usability of case law databases for international courts, such as the CJEU and ECtHR, which currently suffer from outdated interfaces and limited search functionality.
8. Gender Bias in LLMs: Preliminary Evidence from Shared Parenting Scenario in Czech Family Law presents preliminary results demonstrating that leading Large Language Models (LLMs) can exhibit gender bias by allocating disproportionately higher shared parenting ratios to the mother when the father holds a comparable negative risk factor in a Czech family law divorce scenario.
9. AI-Powered Resentencing: Bridging California’s Second-Chance Gap Through Natural Language and Vector-Based Legal Analytics introduces an interpretable, open-source AI platform that uses natural language and vector-based legal analytics to systematically identify eligible incarcerated individuals for resentencing under California reforms like the Racial Justice Act, thereby accelerating case evaluation and bridging the implementation gap.
10. AI Assistance for Court Review of Default Judgments demonstrates that an AI-powered Default Assistant, which uses Retrieval-Augmented Generation (RAG) to provide cited, requirement-level recommendations, significantly increases the accuracy and efficiency of human reviewers in verifying statutory compliance during high-volume debt collection default judgment reviews.
11. CaseConnect: Cross-Lingual Legal Case Retrieval with Semantic Embeddings and Structure-Aware Segmentation presents a legal information retrieval system using a hybrid methodology of legal expertise and automated processing with multilingual semantic embeddings to enable efficient cross-jurisdictional comparison and retrieval of similar cases from heterogeneous legal systems.
12. Glitter: Visualizing Lexical Surprisal for Readability in Administrative Texts proposes a visualization framework that uses language models to measure lexical surprisal (information entropy) in administrative and bureaucratic texts, aiming to estimate and improve their readability and clarity for the public.
13. AI and Judicial Transformation: An Empirical Comparative-Quantitative Analysis of Predictive Tools in EU Labour Law analyzes the digital maturity and attitudes toward predictive justice in labour law litigation across six EU Member States and proposes developing a legal chatbot that functions as a regulatory instrument to guide workers and employers toward the most suitable dispute resolution pathway (negotiation, mediation, or litigation).
14. From Textual Simplification to Epistemic Justice: Large Language Models and Digital Dispute Resolution conceptually argues that barriers to access to justice are fundamentally epistemic and proposes a normative-technical framework for LLM-mediated legal systems based on situated linguistic adaptation, explainability, and participatory infrastructure to support user interpretive agency in digital dispute resolution.
15. PLDF – A Private Legal Declarative Document Generation Framework introduces the Private Legal Declarative Document Generation Framework (PLDF), a serverless, browser-based, declarative framework designed to parse Docassemble interviews and generate documents while maintaining strong privacy guarantees for small-scale access-to-justice applications.
16. How the ECtHR Frames Artificial Intelligence: A Distant Reading Analysis conducts a computational text analysis of ECtHR judgments referring to AI, revealing that judges predominantly frame the technology through themes of risk, threat, and state power, rather than focusing on innovation or efficiency benefits.
17. LegalWebAgent: Empowering Access to Justice via LLM-Based Web Agents introduces the LegalWebAgent framework, which uses multimodal large language models to power a web agent capable of autonomously navigating legal websites, interacting with online forms, and synthesizing information to overcome practical barriers in accessing justice.
18. CourtPressGER: A German Court Decision to Press Release Summarization Dataset introduces CourtPressGER, a 6.4k dataset of German court decisions and corresponding human-drafted press releases, serving as a benchmark to train and evaluate Large Language Models in generating accurate, readable, citizen-oriented summaries of legal texts.
19. A Voice-First AI Service for People-Centred Justice in Niger reports on the “Hadin kai–bani fondo” pilot, a retrieval-bounded, voice-first AI service delivered via Interactive Voice Response (IVR) that provides nationally validated guidance on land disputes to augment customary leaders and expand access to people-centred justice in a low-connectivity, legally plural environment.
The workshop day was followed by the full conference, with a few dozen more papers on a range of legal AI topics, including many relevant to the access to justice themes of the workshop. Those papers, too, are available open access. My own paper, That’s So FETCH: Fashioning Ensemble Techniques for Classification in Civil Legal Intake and Referral, received the Best Paper award during the conference.
