My resolution: a year of AI maximalism for access to justice

Last month, I was in Turin, Italy to present a paper about Lemma Legal’s work on legal intake, covering a tool we built to help Virginia Legal Aid Society, Louisiana’s statewide legal help navigator, and Oregon State Bar’s Lawyer Referral Service. I also helped run a workshop on AI and Access to Justice. (By the way, our paper That’s So FETCH: Fashioning Ensemble Techniques for LLM Classification in Civil Legal Intake and Referral won an award as best paper at the conference. It seems there’s a strong appetite to hear about real-world applications of AI in the nonprofit legal space. )

I came away with three key ideas:

• We need to think “big” with AI. AI solutions that just solve small problems will be increasingly irrelevant to the public.
• We should think of enabling tools that help frontier models work more safely, alongside building our own custom tools.
• We need to consider harm reduction as our primary model for helping the public use AI safely.

Overall, my goal this year is to take a page from Ethan Mollick’s concept of AI maximalism. My first three years with LLMs have been overall pretty tentative and exploratory. This year, I want to see how far dreaming big can take access to justice.

Bespoke, tested solutions compete poorly with general purpose AI

At the workshop, I ran into so many interesting ideas. One theme that kept emerging was that the race to build bespoke solutions–tools with guardrails and quality control, to help keep the general public safe–keeps getting outpaced by the growing quality and capabilities of tools like ChatGPT and Gemini. We can build the best tool, but if ChatGPT is almost as good, it’s just hard to see how people will discover and use our tool at all. Especially if the guardrails and safety controls we put in our tool make it less useful or harder for the general public.

Suppose we make an amazing tool, like the lease explainer idea that keeps getting reinvented. The benefit of a tool like this is that you can load in specific rules for a specific legal jurisdiction, and maybe you can make the tool have a really nice interface to show which terms are problematic, and you can link it to jurisdiction-specific help resources to help you actually act on those rights. Great, right? Well: you are competing with ChatGPT, which also lets you…upload your lease, get an explanation, and search the web to find help resources. Google Gemini 3.0 even can create a custom interface for you to use on the fly, instead of making you read through the text of a chat thread.

Small solutions might not be compelling enough to overcome the search problem

When a layperson looks for help, they are likely to turn to tools that they know and understand. If they use ChatGPT for other parts of their life, they might not even Google for a custom solution. And if they do, they will make a quick judgment about whether the custom solution is going to do a better job than ChatGPT based on their past experience with having it solve problems for them.

We might know our tool is safer and higher quality, but on the surface, a layperson won’t know the difference. And to be honest? Unless we can keep our tool up to date, it’s not clear that our version will be safer and higher quality for that long.

Another important factor is the scope of what the tool can do. Single-purpose tools, each with their own user interface and instructions, can become overwhelming when a user is expected to pick them up to solve different steps in their journey, such as learning about the law, finding out if they are eligible for a legal solution, generating a pleading, and then filing the document.

This week, I had a handful of general-purpose tasks that I might have tried to write a script for to help me with in the past, or even have turned to a dedicated, paid web service to help me with. Instead, I used OpenAI’s commandline AI agent tool, Codex, to go forth on the Internet and do it for me. Instead of a specialized tool, I used the hammer I had in front of me. Jinzhe Tan’s paper, presented at the JURIX workshop, talks about building Legal AI Web Agents that can go forth and do something similar for legal tasks, including filling in forms.

“Big” solutions solve litigants’ full problems–not just a small part

For years, I have written about the importance of understanding a litigant’s full user journey when building an app. I like the metaphor of a map. A legal app is always an intervention into a small part of that journey, and it is your job as a developer to help someone understand what happens before, during, and after the time that they use your app.

This has always seemed sensible if for no other reason than access to justice technology never has a large enough budget to tackle the full solution. We’ve always had to dream a little small.

But what if LLMs throw that advice out of the window? What if AI enables us to build complete solutions almost all of the time? Is it still smart to keep building small tools and hope that our instructions help a litigant complete the full process?

• Instead of a small project to help improve grammar, what if we used AI to help someone draft an entire narrative?
• Instead of a triage tool that gets someone to a legal help resource, what if we used AI to tell them exactly what to do, and how to do it?
• Instead of just telling someone they are eligible for SNAP benefits, what if we collected all of their documentation for them from third party sources and filed the application for them?

My gut feeling is that the smaller the task that we have an AI do–say, a very small and safe task, like grammar checking–the less likely that someone is to pick it up for just a part of their process. Instead, general purpose tools are going to win out. By thinking big, we can help solve litigant problems in a way that is likely to increase adoption. And with the help of LLMs for the coding tasks, the dreaming-big cost is not as large as it used to be.

The best of both worlds? Can we make “safe” tools that AI can use?

Florian Schnitzhofer and Christoph G. Schuetz had a really cool poster presentation at JURIX:
Enhancing Automated Decision-Making in Administrative Law Through Digital Twins of Legislation. A Grounded Theory Approach. You might already know about Model Context Protocol: a way to have AI make use of safe, rules-based code to solve discrete problems, like a calculator for solving an equation. The new idea for me is that we could apply central ideas from rules as code, namely that we can turn legislation into software, to enhance the loosey-goosey world of generative AI. By putting a rules-as-code engine behind the Model Context Protocol, we can have an LLM like ChatGPT ask the rules engine to give an exact, 100% correct answer through an interface that the user already knows and prefers.

I think this idea has some strong potential. Not everything really fits the idea of a tool that the AI can call, though. Sometimes we want to point the AI at a trusted knowledge base, or have it do something as simple as ensuring it always gathers a complete set of facts (like the jurisdiction the person is in) before generating any tailored legal information.

Yes, we have RAG chatbots already. And yes, we have custom GPTs. But they are still pretty cumbersome to build and they suffer from the same discovery problem that other legal tech tools are increasingly going to run into, discussed above.

I’d like to see more innovation in this space, so that we can:

• Point ChatGPT and similar general purpose LLMs to a set of tools to answer specific legal questions like “do I qualify” or “how much child support do I owe?”
• Point these tools at trusted knowledgebases made for consumption by the AI for questions that we haven’t thought of.
• Offer some safe personas or pre-written rules about what questions to ask, without making the legal aid provider or court pay the extra cost of running the query through their own systems.

But our small solutions really are a lot safer!

While a layperson might pick up an AI tool and be satisfied if the output is readable and “looks right”, as AI researchers, legal aid providers, and courts, we have the ability to prove that our tools are safe and reliable. We can do that with systematic testing. The fact that laypeople might just not care can be enormously frustrating! They might pick up a tool with 80% reliability when we have a hard to find solution that would get them to 99% safety.

The risks of improper AI usage are real. Litigants can get sanctioned for citing made up cases. They can get steered with bad advice into a worse outcome than doing nothing. And they might miss a deadline, pay more than they should, or even risk physical harm.

I still struggle with this. But I think it’s clear that we can’t “just say no” to litigants who want to solve their legal problems, or “hire a lawyer” when they can’t afford to pay for one and legal aid programs are turning away 2 in 3 applicants who qualify for their help. It’s unrealistic to expect abstention from ChatGPT when the alternative is no legal help at all.

Towards a harm reduction approach for AI use by self represented litigants

I propose that as an access to justice community, we think very hard about a harm reduction approach to enabling pro se litigants in the world of free or low cost large language models. What might that mean?

I think we need to do a comprehensive survey to understand what use cases are most popular for self represented litigants. I don’t really know what those use cases are yet. Margaret Hagan and others have been doing some of that work, but I think we need to do much more.

While we wait for those data, I think we can still start to predict some use cases, and then think through the harm reduction strategies that might make them safer. And we have to think about how to roll out those solutions fast–without necessarily requiring litigants to adopt a clunky interface that we built for them from scratch. I have a few, tentative thoughts:

Where have I been thinking too small?

Working on the FETCH paper really helped me see how thinking “big” doesn’t have to be too expensive. I’ve done several small AI projects where we used one AI tool, got something (like a draft output) as far as we could, and then used human labor to finish the work.

For example: we’ve used AI to automatically add “labels” to documents for document automation. We’ve also used it to draft the series of questions that the litigant interacts with to fill-in the documents.

With FETCH, I realized that it might cost pennies to throw 3, 4, or 5 AI models at the same problem and have them work together to solve it. While a single affordable model might not get better than 85% accuracy, finding a way to combine them can bring the accuracy level up high enough to avoid the need for a human’s labor at all.

I’d like to push that idea to the limits this year to see how far it can take me.

My goals for this year

This year, I want to tackle at least a few projects (let’s commit to 2-3) where I don’t think small, but instead dream big about the biggest, best impact that AI could have if it worked perfectly. I don’t know what those problems are going to be, yet. But I look forward to solving them!

Will you join me? What one problem do you think litigants are turning to ChatGPT for now, and how can we help them solve it better and more safely without just saying “no?”