On October 17 and 18, 2024 Stanford Legal Design Lab hosted the first-ever AI and Access to Justice Summit.

The Summit’s primary goal was to build strong relationships and a national, coordinated roadmap of how AI can responsibly be deployed and held accountable to close the justice gap.

Who was at the Summit?

Two law firm sponsors, K&L Gates and DLA Piper, supported the Summit through travel scholarships, program costs, and strategic guidance.

The main group of invitees were frontline legal help providers at legal aid groups, law help website teams, and the courts. We know they are key players in deciding what kinds of AI should and could be impactful for closing the justice gap. They’ll also be key partners in developing, piloting, and evaluating new AI solutions.

Key supporters and regional leaders from bar foundations, philanthropies, and pro bono groups were also invited. Their knowledge about funding, scaling, past initiatives, and spreading projects from one organization and region to others was key to the Summit.

Technology developers also came, both from big technology companies like Google and Microsoft and legal technology companies like Josef, Thomson Reuters, Briefpoint, and Paladin. Some of these groups already have AI tools for legal services, but not all of them have focused in on access to justice use cases.

In addition, we invited researchers who are also developing strategies for responsible, privacy-forward, efficient ways of developing specialized AI solutions that could help people in the justice sphere, and also learn from how AI is being deployed in parallel fields like in medicine or mental health.

Finally, we had participants who work in regulation and policy-making at state bars, to talk about policy, ethics, and balancing innovation with consumer protection. The ‘rules of the road’ about what kinds of AI can be built and deployed, and what standards they need to follow, are essential for clarity and predictability among developers.

What Happened at the Summit?

The Summit was a 2-day event, split intentionally into 5 sections:

• Hands-On AI Training: Examples and Research to upskill legal professionals. There were demo’s, explainers, and strategies about what AI solutions are already in use or possible for legal services. Big tech, legal tech, and computer science researchers presented participants with hands-on, practical, detailed tour of AI tools, examples, and protocols that can be useful in developing new solutions to close the justice gap.
• Big Vision: Margaret Hagan and Richard Susskind opened up the 2nd day with a challenge: where does the access to justice community want to be in 2030 when it comes to AI and the justice gap? How can individual organizations collaborate, build common infrastructure, and learn from each other to reach our big-picture goals?
• AI+A2J as of 2024: In the morning of the second day, two panels presented on what is already happening in AI and Access to Justice — including an inventory of current pilots, demo’s of some early legal aid chatbots, regulators’ guidelines, and innovation sandboxes. This can help the group all understand the early-stage developments and policies.
• Design & Development of New Initiatives. In the afternoon of the second day, we led breakout design workshops on specific use cases: housing law, immigration law, legal aid intake, and document preparation. The diverse stakeholders worked together using our AI Legal Design workbook to scope out a proposal for a new solution — whether that might mean building new technology or adapting off-the-shelf tech to the needs.
• Support & Collaboration. In the final session, we heard from a panel who could talk through support: financial support, pro bono partnership support, technology company licensing and architecture support, and other ways to build more new interdisciplinary relationships that could unlock the talent, strategy, momentum, and finances necessary to make AI innovation happen. We also discussed support around evaluation so that there could be more data and more feeling of safety in deploying these new tools.

Takeaways from the Summit

The Summit built strong relationships & common understanding among technologists, providers, researchers, and supporters. Our hope is that we can run the Summit annually, to track progress in tackling the justice gap with AI and to observe what progress has been made, year-to-year. It is also to see the development of these relationships, collaborations, and scaling of impact.

In addition, some key points emerged from the training, panels, workshops, and down-time discussions.

Common Infrastructure for AI Development

Though many AI pilots are going to have be local to a specific organization in a specific region, the national (or international) justice community can be working on common resources that can serve as infrastructure to support AI for justice.

• Common AI Trainings: Regional leaders, who are newly being hired by state bars and bar foundations to train and explore how AI can fit with legal services, should be working together to develop common training, common resources, and common best practices.
• Project Repository: National organizations and networks should be thinking about a common repository of projects. This inventory could track what tech provider is being used, what benchmark is being used for evaluation, what AI model is being deployed, what date it was fine-tuned on, and if and how others could replicate it.
• Rules of the Road Trainings. National organizations and local regulators could give more guidance to leadership like legal aid executive directors about what is allowed or not allowed, what is risky or safe, or other clarification that can help more leadership be brave and knowledgeable about how to deploy AI responsibly. When is an AI project sufficiently tested to be released to the public? How should the team be maintaining and tracking an AI project, to ensure it’s mitigating risk sufficiently?
• Public Education. Technology companies, regulators, and frontline providers need to be talking more about how to make sure that the AI that is already out there (like ChatGPT, Gemini, and Claude) is reliable, has enough guardrails, and is consumer-safe. More research needs to be done on how to encourage strategic caution among the public, so they can use the AI safely and avoid user mistakes with it (like overreliance or misunderstanding).
• Regulators<->Frontline Providers. More frontline legal help providers need to be in conversation with regulators (like bar associations, attorneys general, or other state/federal agencies) to talk about their perspective on if and how AI can be useful in closing the justice gap. Their perspective on risks, consumer harms, opportunities, and needs from regulators can ensure that rules are being set to maximize positive impact and minimize consumer harm & technology chilling.
• Bar Foundation Collaboration. Statewide funders (especially bar foundations) can be talking to each other about their funding, scaling, and AI strategies. Well-resourced bar foundations can share how they are distributing money, what kinds of projects they’re incentivizing, how they are holding the projects accountable, and what local resources or protocols they could share with others.

AI for Justice Should be Going Upstream & Going Big

Richard Susskind charged the group with thinking big about AI for justice. His charges & insights inspired many of the participants throughout the Summit, particularly on two points.

Going Big. Susskind called on legal leaders and technologists not to do piecemeal AI innovation (which might well be the default pathway). Rather, he called on them to work in coordination across the country (if not the globe). The focus should be on reimagining how to use AI as a way to make a fundamental, beneficial shift in justice services. This means not just doing small optimizations or tweaks, but shifting the system to work better for users and providers.

Going Upstream. He also charged us with going upstream, figuring out more early ways to spot and get help to people. This means not just adding AI into the current legal aid or court workflow — but developing new service offerings, data links, or community partnerships. Can we prevent more legal problems by using AI before a small problem spirals into a court case or large conflict?

After Susskind’s remarks, I focused in on coordination among legal actors across the country for AI development. Compared to the last 20 years of legal technology development, are there ways to be more coordinated, and also more focused on impact and accountability?

There might be strategic leaders in different regions of the US and in different issue areas (housing, immigration, debt, family, etc) that are spreading

• best practices,
• evaluation protocols and benchmarks,
• licensing arrangements with technology companies
• bridges with the technology companies
• conversations with the regulators.

How can the Access to Justice community be more organized so that their voice can be heard as

• the rules of the road are being defined?
• technology companies are building and releasing models that the public is going to be using?
• technology vendors decide if and how they are going to enter this market, and what their pricing and licensing are going to look like?

Ideally, legal aid groups, courts, and bars will be collaborating together to build AI models, agents, and evaluations that can get a significant number of people the legal help they need to resolve their problems — and to ensure that the general, popular AI tools are doing a good job at helping people with their legal problems.

Privacy Engineering & Confidentiality Concerns

One of the main barriers to AI R&D for justice is confidentiality. Legal aid and other help providers have a duty to keep their clients’ data confidential, which restricts their ability to use past data to train models or to use current data to execute tasks through AI. In practice, many legal leaders are nervous about any new technology that requires client data — -will it lead to data leaks, client harms, regulatory actions, bad press or other concerning outcomes?

Our technology developers and researchers had cutting-edge proposals for privacy-forward AI development, that could deal with some of these concerns around confidentiality. THough these privacy engineering strategies are foreign to many lawyers, the technologists broke them down into step-by-step explanations with examples, to help more legal professionals be able to think about data protection in a systematic, engineering way.

Synthetic Data. One of the privacy-forward strategies discussed was synthetic data. With this solution, a developer doesn’t use real, confidential data to train a system. Rather, they create a parallel but fictional set of data — like a doppelganger to the original client data. It’s structurally similar to confidential client data, but it contains no real people’s information. Synthetic data is a common strategy in healthcare technology, where there is a similar emphasis on patient confidentiality.

Neel Guha explained to the participants how synthetic data works, and how they might build a synthetic dataset that is free of identifiable data and does not violate ethical duties to confidentiality. He emphasized that the more legal aid and court groups can develop datasets that are share-able to researchers and the public, the more that researchers and technologists will be attracted to working on justice-tech challenges. More synthetic datasets will both be ethically safe & beneficial to collaboration, scaling, and innovation.

Federated Model Training. Another privacy/confidentiality strategy was Federated Model Training. Google DeepMind team presented on this strategy, taking examples from the health system.

When multiple hospitals all wanted to work on the same project: training an AI model to better spot tuberculosis or other issues on lung X-rays. Each hospital wanted to train the AI model on their existing X-ray data, but they did not want to let this confidential data to leave their servers and go to a centralized server. Sharing the data would break their confidentiality requirements.

So instead, the hospitals decided to go with a Federated Model training protocol. Here, an original, first version of the AI model was taken from the centralized server and then put on each of the hospital’s localized servers. The local version of the AI model would look at that hospital’s X-ray data and train the model on them. Then they would send the model back to the centralized server and accumulate all of the learnings and trainings to make a smart model in the center. The local hospital data was never shared.

In this way, legal aid groups or courts could explore making a centralized model while still keeping each of their confidential data sources on their private, secure servers. Individual case data and confidential data stay local on the local servers, and the smart collective model lives at a centralized place and gradually gets smarter. This technique can also work for training the model over time so that the model can continue to get smart as the information and data continue to grow.

Towards the Next Year of AI for Access to Justice

The Legal Design Lab team thanks all of our participants and sponsors for a tremendous event. We learned so much and built new relationships that we look forward to deepening with more collaborations & projects.

We were excited to hear frontline providers walk away with new ideas, concrete plans for how to borrow from others’ AI pilots, and an understanding of what might be feasible. We were also excited to see new pro bono and funding relationships develop, that can unlock more resources in this space.

Stay tuned as we continue our work on AI R&D, evaluation, and community-building in the access to justice community. We look forward to working towards closing the justice gap, through technology and otherwise!