When the Law Is Not Settled

Picture a Tuesday in late autumn, a little after nine at night. An associate we will call Dana is the last person left on her floor. The cleaning crew has come and gone, the hallway lights have dropped to their after-hours dimness, and the coffee at her elbow went cold an hour ago. On her screen is a brief due in the morning, and in the brief is one paragraph that rests on one legal proposition.

The question is narrow and technical. Does the discovery rule apply to limitations periods for claims under ERISA? Put more plainly: does the clock start when the injury happens, or when the plaintiff finds out about it?

Dana types the question into the research tool her firm licenses. The answer arrives in seconds. Yes, it says, and it cites authority. The authority is real. The prose is calm and assured, the way a good memo is calm and assured. She reads the cases, they say what the tool says they say, and she writes her paragraph. Then she turns off the lamp and goes home.

Two weeks later the opposition brief arrives. It cites three opinions from two other federal circuits, all decided within the last fifteen years, and all going the other way.

Dana pulls every one of them. The cases are real. The holdings are real. The division between the circuits is real.

Her tool did not invent anything. It did something quieter, and in its way more unsettling. It took a question on which the federal courts are divided and handed it to her as though the matter were closed.

Anyone who has practiced long enough will recognize what Dana feels as she reads that brief. It is not the sting of a typo. It is the slower understanding that the ground she built on was never solid, and that nothing on her screen that Tuesday night told her so.

Two ways of not knowing

There are two very different reasons a lawyer can be unsure about the law, and they call for nearly opposite responses.

The first is the ordinary kind. The answer exists; you simply have not found it yet. Courts have settled the question, and the uncertainty lives entirely with you, the researcher. More searching will cure it: another pass through the databases, a call to someone who knows the field, an afternoon in the footnotes of a treatise. Call this shallow ambiguity. It is a gap in your research.

The second kind is different in character. Here the courts themselves have reached opposite conclusions on the same legal question. The answer depends on which circuit you stand in, which judge you draw, or which method of interpretation a court decides to use. No amount of research will make this uncertainty go away, because it is not in you. It is in the law. Call this genuine ambiguity. It is a gap in the doctrine.

One of these is a problem you can solve by working harder. The other is a fact about the world you have to tell your client.

Most legal AI tools were built for the first kind. They are very good at finding cases and summarizing what those cases hold. What they lack is any way to recognize the second kind. They cannot tell the difference between "we have not found the counterargument yet" and "the courts are genuinely divided." So when a tool like that meets a real split, it does what it was built to do. It finds support, it stops looking, and it speaks with confidence.

We talk a great deal about hallucination, and we should. But a fabricated case is, in the end, a visible failure; someone can catch it. A real case, correctly cited, for one side of a question the courts have never agreed on, passes every check we know how to run. That is why the smooth answer is, in a sense, worse than the invented one.

The shapes a gray area takes

It helps to know genuine ambiguity by sight, because it comes in more than one form, and each form asks something slightly different of the lawyer. The working draft this essay grows from names five.

Words that fray at the edges. H.L.A. Hart called it "open texture." A rule uses a word like "vehicle," "residence," or "substantial similarity," and the word has clear cases at its center and contested territory at its margins. Is a skateboard a vehicle under a traffic statute? Is a houseboat a residence for property tax purposes? Is copying 30 percent of the distinctive expressive elements of a software interface "substantial similarity"? These are not holes in the case law. They are holes in language itself, and the cases that exist press against the edges in different directions.

Principles that pull against each other. Speech is protected. Privacy is protected. Sometimes they collide, and the law honors both. When that happens, courts reach for a balancing test, and they balance differently depending on the facts in front of them, the philosophy of the judge, and the doctrine of the circuit. There is no clear rule hiding somewhere, waiting to be found. The tension is the rule.

Methods that disagree. How should a statute be read? Some judges look to the text, some to the purpose behind it, some to how the words were used when they were written. Karl Llewellyn noticed long ago that the canons of construction come in dueling pairs, one pointing each way, so that either side of an argument can find a canon to stand on. The same words, read through different lenses, give different answers, and every lens is legitimate.

Courts that disagree. This is the circuit split, the form most practicing lawyers know best. The same statute, the same facts, and you win in one circuit and lose in another. The disagreement is the whole reason cert petitions exist. It is also where Dana's question lived. The law was settled in each circuit; it was simply settled differently.

Facts the law has not met before. Something new arrives: AI-generated content, deepfakes, algorithmic trading, digital tokens. Is AI-generated work more like a photograph or more like found art? Is a deepfake more like defamation or more like parody? The cases we have were decided before these facts existed, so the law must stretch by analogy, and reasonable lawyers disagree about which analogy fits.

None of these five is a failure of diligence. All five are ordinary features of a legal system made of words, principles, methods, many courts, and a world that keeps changing.

What the smooth answer costs

The difference between the two kinds of uncertainty is not academic. It changes nearly everything a lawyer does next.

It changes the conversation with the client. If the problem is shallow, your job is to research harder: find the controlling case, find the close analogy, build the chain of reasoning. If the problem is genuine, your job is different. You explain that the law is contested. You help the client understand the risk that comes with that, and you write it down. What you do not do is present one answer as though it were the only one.

It changes the brief. On a genuinely contested question, the honest brief shows the court that you understand the split. It distinguishes the adverse authority, explains why your side has the better argument, and acknowledges the uncertainty rather than hiding it. Courts tend to respect a brief that walks through a real disagreement with care. They do not respect a brief that pretends the disagreement is not there. A tool trained simply to "answer the question" pushes in exactly the wrong direction: it picks a side and moves on.

It touches the duty of candor. The Model Rules of Professional Conduct ask a lawyer to disclose directly adverse authority in the controlling jurisdiction, to represent clients competently, and to keep them reasonably informed. A lawyer who relies on a tool that presents contested doctrine as settled, and who never learns of the other line of cases, may find all three duties in play at once. The tool's silence does not become the lawyer's excuse.

It creates real exposure. Picture a client that organizes its affairs around one circuit's rule, on advice that never mentioned the other rule, and then finds itself litigating in a court that follows the other one. When the harm arrives, the question will be whether the lawyer competently advised on a known area of doctrinal uncertainty. False confidence has a price, and it is not always the tool vendor who pays it.

What an honest answer sounds like

The fix is not more data poured into the same design. It is a different design, one built from the start to notice when courts disagree.

Imagine Dana's question answered honestly. The reply would not begin with "yes." It would sound something like this:

This question is genuinely contested. Some federal circuits hold that the discovery rule does not apply to these claims; others hold that it does. Here are the cases on each side. Here is what governs in your circuit, and here is the line of authority opposing counsel will almost certainly cite. You will need to distinguish those cases, and you may want to tell your client the question is unsettled.

That answer asks more of the system. To give it, a tool has to do five things:

1. Identify that the question has conflicting judicial authority.
2. Map which courts have said what.
3. Flag the split in plain words, rather than burying it in a list of results.
4. Present the competing positions without false neutrality, including an honest view of which side has the stronger numbers, the weightier precedent, or the recent momentum.
5. Counsel the lawyer that this is a place where the law is open and judgment is required.

There is a reason tools are not usually built this way, and it is worth saying out loud. A confident answer feels better. It feels more useful. Lawyers are busy, and a tool that gives direction is more pleasant to use than a tool that adds caveats. But that pleasure is borrowed against the future. A tool that manufactures confidence is worse than no tool at all, because it removes the very doubt that would have sent the lawyer back to the library. A tool that tells the truth about a gray area is the one that actually helps.

A map, not an oracle

Here is the heart of it. Computation cannot make the law determinate. No system, however large, can resolve a question the courts themselves have not resolved. What computation can do, and do well, is find the places where the courts have left a question open, and show them to us.

That changes what we should want from these tools. The promise is not "the machine answers legal questions." It is "the machine shows you which questions deserve your judgment." It is a humbler promise. It is also more honest, and more useful in the daily work of practice.

Consider how far that could reach. Not long ago, a junior lawyer could miss a circuit split simply by searching for the wrong combination of words. A well-built system should be able to catch it on its own, and to say something like: there are eight cases on this question; they break five to three in your favor; the recent ones lean the other way. That kind of answer does not remove uncertainty. It makes uncertainty visible, and once visible, workable.

The Institute has tried to give that map a scale. Its Grayness Score gathers the signals courts leave behind when they are unsure (an acknowledged split, a question of first impression, hedged language, a forceful dissent, a balancing test with no ranking among its factors) and combines them into a single reading that runs from clear law at one end to profound indeterminacy at the other. For a practitioner, the number is not the most important part. The most important part is that the number arrives with the cases that produced it, so a lawyer can open them and see for herself. The scholarly case for that approach is set out in Measuring Doctrinal Indeterminacy and in the paper Detecting Genuine Doctrinal Ambiguity.

Until tools like that are the norm, the questions worth asking of any research tool are simple ones:

• Does it flag circuit splits? When you ask about a question with conflicting authority, does it tell you so, or does it pick a side?
• Does it distinguish the two kinds of uncertainty? When it says the law is unclear, does it mean it has not found the right case, or that the courts are divided? It should tell you which.
• Does it show you the cases that go the other way? Even if it does not name the split, does it at least put the adverse authority in front of you, or does it bury it?
• Does its confidence mean anything? A confidence score is only worth reading if it reflects doctrinal disagreement, not merely how often a case has been cited.
• Does it sound different on contested questions? "Here is the law" and "here is a contested area where you will need to brief both sides and counsel your client" should not sound the same. If they do, the tool is doing you a disservice.

Back to the lamp

Return, for a moment, to that Tuesday night. Same cold coffee, same dim hallway, same brief due in the morning. Only this time, when Dana types her question, the answer does not begin with "yes." It begins with "the courts are divided."

Her night gets longer. She writes two paragraphs instead of one. She sends a note to the partner before she leaves. Perhaps, in the morning, someone calls the client. The brief is harder to write, and much easier to stand behind. And when the opposition arrives two weeks later, with its three opinions from two other circuits, there is nothing in it she has not already read.

The law has always had gray areas; that is not new. What is new is that we can now find them systematically, if we choose to build tools that look. The temptation will always be to use all that power to sound certain. The better use is to show a lawyer, clearly and early, where the certainty ends. In law, honesty about uncertainty is very often worth more than confidence about a settlement that never happened.