Roman builders raised an arch in a set order: first a temporary wooden frame, the centering, to carry the stones; last, the keystone at the crown. Then the stones locked together; the frame came down. Remove the frame before the keystone is set, and the arch falls.

The legal profession has formed lawyers the same way. The frame was apprenticeship: a junior lawyer drafts, a senior lawyer redlines, and the junior absorbs why. The keystone is judgment: the capacity to stand behind advice and supervise others. Generative AI is now taking the frame down. That is no loss in itself; much of it was drudgery. But the first drafts juniors learned from are increasingly written by machines, and the profession has not yet settled on what will hold new lawyers up while their judgment forms.

The Keystone Initiative at Stanford CodeX exists to solve that problem: to form lawyers who can be trusted to supervise AI’s legal work.

Judgment is not knowing the rules

As Holmes wrote in The Common Law: “The life of the law has not been logic: it has been experience.” H.L.A. Hart asked whether a ban on vehicles in the park covers bicycles. Roller skates? Someone must decide; deciding well is judgment. Polanyi explained why it is hard to teach: “we can know more than we can tell.”

An expert reads a draft and knows a clause is wrong before she can say why. Often she simply fixes it. The redline, the delta between one draft and the next, is one place that judgment leaves a record.

Why this is society’s problem

Law runs on answerability. Under Rule 11(b), by presenting a filing to the court, a lawyer certifies that it is based on an “inquiry reasonable under the circumstances.” In Mata v. Avianca, lawyers submitted a brief citing non-existent cases generated by ChatGPT, and the court reminded them of their “gatekeeping role.” More than 2,000 judicial decisions involving AI hallucinations or related fabricated or misrepresented material have since been catalogued worldwide.

Bar and bench put the safeguard in the same place: a competent human. ABA Formal Opinion 512 says AI tools “cannot replace the judgment and experience necessary for lawyers to competently advise clients.” Chief Justice Roberts wrote that legal determinations “often involve gray areas that still require application of human judgment.” Each of these authorities places responsibility on a human capable of oversight. None says where that capability will come from.

The stakes reach past the profession:

• Clients often can’t check the work. They rely on the license. If the apprenticeship behind it hollows out, nothing announces it, and the clients least able to notice pay first.
• Law changes through people who notice an inherited rule no longer fits. Tools trained largely on past practice can reinforce it; they do not bear responsibility for deciding when it should change.
• Judgment could be rationed, or opened up. If apprenticeships grow rare, judgment becomes a perk of the right firm. Yet in a randomized study, GPT-4 produced large speed gains, and where it improved quality, the lowest-skilled participants saw the largest gains. Measure the path from competent to expert, and judgment could reach people the old system shut out.

The irony of automation

In 1983 Lisanne Bainbridge named the “ironies of automation.” Machines take the routine work, yet humans must still catch their failures with skills that decay without practice. “The more advanced a control system is,” she wrote, “so the more crucial may be the contribution of the human operator.” Law’s version is sharper: the better the tools, the rarer and subtler their errors, and the more expertise needed to catch them.

The apprenticeship functioned, in part, like a public good paid for privately. Through billable hours, clients helped fund the training of lawyers who went on to serve others. As machines write first drafts more cheaply, clients have less reason to pay for training-by-drafting, and individual firms have less incentive to subsidize it alone. That collective-action problem needs a neutral institution.

What Keystone will do

Much of the record of expert judgment already exists in tracked changes and version histories across firms. Any firm can study its own, but patterns that reflect shared professional judgment rather than one partner’s habits are more likely to emerge across many experts and matters. That points toward pooling across firms, which requires a protocol they trust.

So Keystone begins by developing the Redline Standard: a vendor-neutral protocol, written with the profession at the table, for de-identifying and safely contributing redlines to a Stanford-held research corpus. The standard will govern confidentiality, timing, permissions and permitted use. Its design principles are that each firm retains control of its contributions, no commercial party receives rights the firms do not grant, and nothing is asked of any firm before the standard exists.

We start with patent drafting where edits offer a repeatable domain for study, and most U.S. nonprovisional utility applications publish about 18 months from the earliest filing date for which benefit is claimed. Publication creates a point at which the filed text is public, though draft-level confidentiality still requires the protocol above. If the method works there, other practice areas can follow.

On that foundation we can build a practice simulator: trainees work realistic synthetic matters, free to fail, measured against expert revision, the aim being that new lawyers reach client work only after clearing a demonstrated bar. The premise, that practice against expert redlines builds judgment faster than unstructured experience, is a hypothesis; a university is where to test it rather than assert it.

This fall, a working group will meet at the Stanford d.school to sharpen that hypothesis, followed by a convening at Stanford Law School to begin drafting the standard. Google has committed as the initiative’s anchor partner.

Setting the keystone

The frame comes down either way. The question is whether the profession sets the keystone first, or learns later what it failed to hold up.

The legitimacy of legal practice rests in part on a promise: behind legal work stands a person who understood it and can answer for it. Keeping that promise will take the profession acting together on neutral ground. With Roland Vogl, CodeX’s executive director, we are building a coalition of law firms, legal departments, law schools, technology partners, and others in the legal ecosystem. Learn more at the project page.

Ian C. Schick, Ph.D., Esq., is a Non-Residential Fellow and leads the Keystone Initiative at Stanford CodeX. He is also co-founder and CEO of Paximal, a legal technology company that receives no special access under Keystone’s governance.