Working draft. These guides are shared for review while we prepare them for public release.
Guides and Analysis

Legal Help Commons Reports

Foundation models are already answering the public’s legal questions. This is the work of getting them to do that well across the tasks people actually need, and the shared standards, evidence, and tools that let any state build on it instead of starting alone.

Access the Report Catalog

These reports are working drafts shared for review. Enter the access code you received to browse the full catalog.


Incorrect code. Contact legaldesignlab@law.stanford.edu for access.

Reports

Guides to the questions teams are asking

Explore these guides to learn about AI strategy, development, and evaluation for legal help. Start with the first section or jump to the topic you need.

Start here

How to Plan and Fund Legal Help AI

What justice professionals are saying about AI, the real options for building or buying, and what makes the case for shared investment.

Evaluation

Telling whether AI is any good

How to measure quality, what the measurements found, and how to cover fifty tasks.

How do you tell whether a legal AI tool is any good?

Set up promptfoo and a reusable protocol like the field has done, measure any tool against the tasks you care about, and gather your own evidence before you deploy.

What you'll take away
  • How to stand up your own evaluation with promptfoo, on the tasks you care about
  • The reusable protocol teams ran for under twenty-five dollars a round
  • Why deterministic checks, not an LLM judge, carry the signal
MethodRead the guide

Does grounding the model in our own content make it better?

Across five rounds of testing, a verified list of your own resources lifted every model, while loading full articles helped only when they fit.

What you'll take away
  • What five rounds of testing showed about grounding in your own content
  • When a cheap grounded model beats a premium ungrounded one
  • When loading full articles helps, and when it backfires
EvidenceRead the guide

Can AI classify who a person is and what they need?

A 60-case eval across Claude, Gemini, and GPT-4o on tagging the legal issue, the audience, the jurisdiction, and the urgency from a person's own words. The lesson: put your own category lists in the prompt.

What you'll take away
  • Putting your own category lists in the prompt is what makes intake classification work
  • The numbers across Claude, Gemini, and GPT-4o, and where each one breaks
  • How to run it as one component, with emergency detection and the not-a-legal-issue check kept separate
IntakeRead the guide

How do we find out how good AI is across all fifty tasks?

The field has measured one task of fifty. This sets out the choices, the partners, and the work to build a real read on the rest, and how the issue-area agendas and the task taxonomy fit together.

What you'll take away
  • What the field actually knows today: one task of fifty measured
  • The three ways to build the read: a central team, distributed contribution, or task stewards
  • Whether to do yearly check-ins, and how the issue and task slices fit together
CoverageRead the guide
Tools

What you can use and build

Start with the inventory, then how to get your own content AI-ready, then the specific tools to use or rebuild.

What common tools did we build that every team can use?

You can pick up the shared vocabularies, the datasets to build and test with, and the working tools, each with an honest note on how ready it is.

What you'll take away
  • The vocabularies, datasets, and tools you can pick up now
  • An honest readiness note on each one
  • What to use instead of building it yourself
InventoryRead the guide

How do you get your website and guides ready for AI?

A practical playbook for turning your website, guides, and trainings into a knowledge base an AI can use safely: how to gather and tier content, chunk it, build in safety, and grow from search to a knowledge graph.

What you'll take away
  • The six-stage journey from a content backlog to an AI-ready knowledge base
  • How to gather content by CMS export, tier it by authority, and chunk it for the question
  • How to build safety into the content itself, and when a knowledge graph is worth it
PlaybookRead the guide

How do you work through each stage of getting content ready?

The fuller, step-by-step draft companion to Guide 17. Every stage of the journey is laid out in detail, with the goal, the steps to take, the tools to use, a picture of what the output looks like, and the signs you have done it well.

What you'll take away
  • The six-stage journey from Guide 17, worked through in full detail
  • For each stage: the goal, the steps, the tools, and what the output looks like
  • The signs a stage is done, and the maintenance cycle that never ends
Full playbookRead the guide

What does a legal help Q&A bot need to get right?

Many teams want to build one. This is a short note on what most decides whether a Q&A bot is safe and useful, with pointers to the guides that go deep.

What you'll take away
  • The four parts of a Q&A bot that outlast the model
  • How to ground answers in verified content with citations
  • How to evaluate it before it ever reaches the public
Q&A noteRead the guide

How can we automatically mask PII in documents?

Out of an expert survey and a labeled legal document set: what counts as PII in legal records, how to judge whether a masking tool actually works, and how to test one without touching a real client's file.

What you'll take away
  • What counts as PII in legal records, in three tiers from an expert survey
  • The five-part rubric for judging a masking tool, and the hard fails that override it
  • How to test a masker against a labeled document set without real client data
PIIRead the guide

Can you use the citation verifier, and build your own?

The cheapest piece of safety infrastructure in the toolkit: a short deterministic check that catches the hallucinated links and misattributed numbers an LLM judge waves through. How to run it, and how to build your own.

What you'll take away
  • What the citation verifier checks, and what it cannot
  • How to run it in your own evaluation alongside the model judge
  • How to build your own from an allowlist of your real resources
VerifierRead the guide

Can you use the issue classifier, and build your own?

A FETCH-style ensemble classifier for the LIST taxonomy, built on Quinten Steenhuis's method. Whether you can use it, how to call it, how to build your own, and how its gaps grow LIST.

What you'll take away
  • Whether you can use the classifier and the two ways to call it
  • How the FETCH ensemble method works and how to build your own
  • How the suggest-new flag feeds the growth of the LIST taxonomy
ClassifierRead the guide

How do you build a content safety wrapper?

A layer around the content that decides, for each kind of legal situation, whether to answer, what to ask first, and when to route to a human. How it works, how experts built the starter, and what is still unproven.

What you'll take away
  • The four risk tiers and what each one does to an answer
  • The per-scenario fields experts fill in, and how they reviewed them
  • How to build and test your own wrapper, and what is still unproven
SafetyRead the guide
Beyond these guides

The rest of our public work

The projects, standards, and tools the Commons has published, gathered here.