AI for Good

Lawyers as builders.

Open source. Open records. Free tools. Free law.

The Premise

English is the new programming language. Lawyers have been writing in it for two hundred years.

Most days the problem isn't bad law. It's a missing tool. A rule that works fine but takes a spreadsheet nobody has. A guideline anyone could follow if anyone could compute it. Lawyers see these problems first and know them best. Until now, seeing the problem and building the fix were different careers. They aren't anymore. If you can write a precise sentence, you can build. That is the whole job description.

The cost of building collapsed while nobody was looking. What used to take a firm, a budget, and a vendor now takes a careful sentence and a free weekend.

That is not what you are being sold. You are being sold a subscription, priced against the fear that you will fall behind, renewed every year on the same fear. A tool you cannot open, cannot check, and cannot fix is not a tool you own. It's rent.

There is a loud, expensive fight going on over who wins artificial intelligence. Let them have it. While everyone is busy fighting, the rest of us can go build the things that should have been built years ago, for problems that should not still exist.

The line about English is David Heinemeier Hansson's, from his conversation with Lex Fridman. We agree with him.

Two tools. One small practice.

Both built by one lawyer, on nights and weekends, in plain English. Both free. Both open. Both in use.

Red Binder Project

A family emergency plan, before the emergency.

Born from work with detained individuals. When a parent is picked up, a trusted person needs to step in that day. Childcare. The house. Who can talk to the school. Who can talk to the lawyer. The Red Binder walks a family through it in English and Spanish and produces the documents.

Everything generates in the browser. Nothing is stored. Nothing is transmitted. The code is public so anyone can check that claim instead of taking our word for it.

Free · Bilingual · Open source · No accounts, no servers, no records kept

Child Support Calculators

A formula buried in a statute becomes an answer anyone can compute.

Federal law says courts start from the guideline number. The guideline number lives in a worksheet most people never see and a spreadsheet most people can't run. These calculators run the state formula in the browser, show every step of the math, and generate the filing.

Seven states so far: Tennessee, Mississippi, Arkansas, Alabama, Louisiana, Georgia, Florida. When a schedule is out of date the tool says so, and we fix it.

Free · MIT licensed · Shows its work · Seven states and counting

Neither one needed a grant, a committee, a vendor, or a computer science degree. Both needed a lawyer who had seen the problem and a free weekend.

There are a thousand more of these. That is the point.

It's all in the approach

Inside-out design. The person closest to the problem designs the tool.

Study the problem.
Break it down.
Keep it simple.
Make it work.

The Standard

Trust is a build requirement. Not a disclaimer at the bottom.

Every line of math cites the rule that authorizes it. Not a bibliography at the end — a citation on the line, beside the number, so anyone can check one figure without checking all of them.

The citation rule is enforced in the build. A computed line without a citation fails the tests and does not ship.

Rules first. A model only where a rule cannot reach, and never where it can.

Verification runs against the state's own official calculator. Not against our own homework.

Twice the tools disagreed with a state and the tools were right. Louisiana was running an out-of-date support schedule. Georgia's parenting-time adjustment was off by an order of magnitude. Nobody was hiding anything. The math was just wrong, in public, for years, inside a number that decides what a parent pays.

We published both. That is the job.

DETERMINISTIC FIRST · CITED ON THE LINE · CHECKED AGAINST THE OFFICIAL TOOL · ERRORS PUBLISHED, NOT BURIED

The Program

This belongs in law school. Not as a seminar about it.

A course where a student takes one real problem, builds one free tool, and ships it. Not a paper about a tool. The tool. Along the way they learn the thing that actually matters now: what makes a machine's output trustworthy enough to rely on, and how to prove it.

Who

Law students. Practicing lawyers. Clinic directors. Professors who want a lab, not a lecture. No coding background required. A precise sentence is the prerequisite.

What you leave with

One working, public, open-source tool with your name on it. A method you can repeat on the next problem. A written record of how you verified the thing works.

What's already here

Suffolk runs the best legal technology lab in the country and maintains code that legal aid offices in several states depend on. Georgetown has run Iron Tech Lawyer since 2013. Stanford's Legal Design Lab publishes honest post-mortems of its own failures, which most of us would not. And this year a wave of law schools opened AI institutes to teach lawyers how the tools work and when to trust them.

All of that is necessary and we are glad it is happening. But notice what most of it produces: a lawyer who is very good at operating someone else's product. That is a fine thing to be and a terrible thing to only be.

Teaching a lawyer to use a tool and teaching a lawyer to make one are different courses. Almost everyone is building the first one. The vendors are delighted.

One rule we'd add

A tool ships with a named owner and a sunset date, or it does not ship.

What kills law school software is not bad code. It's May. The semester ends, the student graduates, the legal aid office goes back to the PDF, and the demo stays up just long enough to mislead somebody. A program that cannot say who maintains the tool in year three has no business putting it in front of a pro se litigant in year one.

Why Washington

In March 2024 the Washington Supreme Court adopted, in concept, a pathway to the bar that replaces the exam with 500 hours of supervised work for actual clients and a portfolio submitted with the application. The implementing rules are still being written.

If supervised tool-building counts toward those hours — and we think it should, because a court forms interview is legal services delivered — then the tool a student ships is the portfolio. That is a question for the court, and we intend to ask it properly rather than assume the answer.

Washington also licensed the country's first legal technicians, and then sunset the program. That court has done this before. It knows where the hard part is. There is no better room to make this argument in.

The first five

Boring, high-volume, rule-bound. A well-made protection order interview helps more people in a month than a clever demo helps in a year.

Eviction response, where the notice defects are specific and the answers are not.

Protection order petitions, where the forms are statewide and the experience is awful.

Child support worksheets, in the forty-three states that still don't have one.

Tribal court forms, which almost nobody has built and fewer still have built well.

Fines, fees, and license reinstatement — rule-bound, life-changing, unautomated.

What we're building

The syllabus. The verification standard. The first cohort. Conversations with a law school are underway. The Red Binder and the calculators are the worked examples, and the standard above is the thing being taught.

Bring a problem.

Lawyers. Students. Professors. Clinics. Bring the problem you see every day. We'll show you how to build the fix.

Deans, clinic directors, and faculty: taylor@tcblaw.org