Fabricated vs. Misgrounded: The Two Ways Legal AI Gets Citations Wrong

Yes, AI invents case law, in two distinct ways

Fabrication and misgrounding are not one problem with two names. They have different causes, leave different traces, and require different checks.

AI does invent case law, and treating it as a single problem is the first mistake. There are two separate failure modes with different causes and different fixes. Fabrication produces a citation with no real case behind it. Misgrounding produces a citation to a genuine case that the AI has quietly misread. Stanford RegLab found that leading legal AI research tools, including retrieval-grounded ones, hallucinate on "1 in 6 or more" queries ^[1]. Damien Charlotin's public database now catalogs more than 1,300 court proceedings flagging suspected AI hallucinations, updated daily ^[2]. Both numbers describe a moving target, not a solved one. Because the two modes look identical on the page, a reviewer who only scans for names that "look wrong" will catch fabrications while waving misgrounded citations straight through to a signed filing.

The reason the single-problem framing fails is that the two modes come apart at the level of cause. A fabrication is the model filling a hole where it has no retrieved authority to work from, so it generates something that reads like a citation and resolves to nothing. A misgrounding is the model working from real material and describing it wrong. Those are different mistakes made under different conditions, and they respond to different countermeasures. Anything you do to make fabrication rarer, such as forcing the model to cite only from a retrieved set, does very little to stop the model from mischaracterizing a document that was correctly retrieved. Once you hold the two apart, the rest of the review follows: you stop asking the single question "is this citation good" and start asking three narrower ones that each mode answers differently.

Fabricated citations: the case does not exist

A fabricated citation is the failure most people picture: the AI invents a case that was never decided. The party names are made up, the reporter volume and page are made up, and sometimes the court and year are made up too. Nothing resolves to a real opinion because there is no opinion. These are the citations behind the most widely reported sanctions, and they are also, paradoxically, the easier mode to catch. A fabricated case fails the simplest possible test: existence. When you try to pull the opinion from a reliable reporter or database, there is nothing to retrieve. Fabrication tends to spike when a model is asked about a narrow or obscure area of law where its training data is thin, so it fills the gap with a plausible-sounding invention. The reviewer's job here is mechanical but essential: confirm that every cited authority actually exists before anyone relies on it.

What makes fabrication tractable is that its tell is binary. Either the opinion resolves or it does not, and there is no honest middle ground where an invented case partially exists. That does not make it trivial to catch, because a fabricated citation is engineered by the model to look ordinary. The party names sound like party names. The reporter abbreviation is a real reporter. The pin cite falls in a plausible range. Nothing on the surface flags the entry as manufactured, which is exactly why the eyeball test is not a substitute for retrieval. The habit that closes this gap is unglamorous and non-negotiable: every authority gets pulled, not skimmed for plausibility. A citation that cannot be retrieved from a reliable source is treated as fabricated until proven otherwise, no matter how convincing it reads.

It helps to be precise about why this mode is the "easy" one, because the word can breed complacency. Fabrication is easy to catch in the sense that the test is unambiguous and fast to run. It is not easy to catch in the sense of being safe to assume away. The reported sanctions almost all involve fabricated authorities that a single retrieval attempt would have exposed, which tells you the problem is rarely the difficulty of the check and almost always the fact that no one ran it. Fabrication is caught by discipline, not by cleverness.

Misgrounded citations: a real case, the wrong proposition

A misgrounded citation is more dangerous because it survives a surface check. The case is real. The citation format is correct. You can pull the opinion and hold it in your hand. The problem is what the AI claims the case says. It attributes a holding the court never reached, inserts a quotation that appears nowhere in the text, or cites the case for a proposition it does not support, sometimes for the exact opposite of what it holds. Because the citation resolves cleanly, existence checks pass and the error slides through. This is the mode that grounding was supposed to solve and only partially does. Catching a misgrounded citation requires reading the opinion and matching each quotation and proposition against what the case actually stands for, which is slower, harder, and easier to skip under deadline.

The danger is not just that misgrounding is harder to catch. It is that misgrounding actively rewards a shallow review. A fabrication offers no false comfort, because the retrieval attempt fails and you know immediately that something is wrong. A misgrounding does the opposite: it hands you a real opinion, in the right reporter, at the right page, and every superficial signal tells you the work checked out. The reviewer who confirms existence and stops has not just missed the error, they have collected positive evidence that feels like verification. That false confidence is the mechanism by which a misgrounded citation reaches a signed filing. The citation passed a check. It just passed the wrong one.

Misgrounding also comes in gradations, and the milder gradations are the most likely to slip through. An outright reversal, where a case is cited for the opposite of what it holds, is at least detectable to anyone who reads the opinion with the proposition in mind. The subtler failures are harder: a case cited for a broader rule than it actually announced, a holding stretched past the facts that produced it, a quotation that is almost verbatim but altered in a way that changes its scope. None of these are caught by asking whether the case exists. All of them are caught only by reading the opinion against the specific claim the brief is making with it.

Why the two modes look identical on the page

On the printed page, a fabricated citation and a misgrounded citation are indistinguishable. Both present as a normal-looking authority: a case name, a reporter, a pin cite, a parenthetical describing what the case supposedly holds. Nothing in the formatting separates the invented case from the real one that has been misread. This is why "does it look right" is not a verification method. Looking right is precisely the property both failure modes are optimized to have, because a citation that looked wrong would never survive even a careless read.

The practical consequence is that you cannot triage citations by eye and decide which ones "deserve" a full check. The citation that most deserves scrutiny is the one that looks most ordinary, because a competent-looking misgrounding is exactly what a shallow review approves. Any process that verifies only the suspicious-looking entries has inverted the risk. The entries that read cleanly are not the ones you can skip; they are the ones that require you to go past the surface and into the opinion itself. Every citation earns the same baseline treatment regardless of how trustworthy it appears.

A citation looking correct is not evidence that it is correct. Both failure modes are designed, in effect, to look correct.

Why "grounded" and RAG tools still miss one mode

Retrieval-augmented generation was meant to end hallucination by tethering the model to real documents. It reduces fabrication meaningfully, because the model is pulling from actual retrieved cases rather than inventing them. But grounding does not eliminate misgrounding. Stanford RegLab tested purpose-built, retrieval-grounded legal research tools and still measured hallucination rates above 17% for one product and above 34% for another ^[1]. The reason is structural: retrieval fetches a real case, but the model can still summarize it wrong, quote it inaccurately, or apply it to a proposition it does not support. The document is real; the interpretation is not. This is the misgrounded failure mode, and it is precisely the one that a "we use RAG" assurance does not close. Grounding narrows the gap between the model and the record. It does not guarantee the model reads the record correctly.

It is worth being exact about what retrieval does and does not fix, because the marketing tends to blur the line. Retrieval operates on the first half of the problem: it improves the odds that the case in front of the model is a real case. It does nothing about the second half: whether the model's description of that real case is accurate. Those are separate stages, and a tool can be excellent at the first while remaining unreliable at the second. When a vendor says the tool is grounded, the honest reading is that fabrication should be rarer, not that the summaries and quotations can be trusted without checking. The rates above 17% and above 34% are measurements of exactly this residual: real documents, still described wrong ^[1].

This also explains why "grounded" is not a synonym for "verified," and why no product can honestly promise output free of hallucination. Grounding is an input-side intervention. Verification is an output-side obligation. A tool can retrieve the right cases and still put a proposition in a brief that the retrieved case does not support, which means the human check on the far end is not redundant with the retrieval on the near end. They guard different failures. Treating a grounded pipeline as a substitute for reading the opinion is how a misgrounded citation gets a clean bill of health it did not earn.

Bad law: the real, accurate citation that still sinks a filing

There is a third problem that sits next to the two hallucination modes and is often confused with them: bad law. A bad-law citation is real and accurately described, but the case has been overruled, vacated, or superseded. The AI read the opinion correctly. The quotation is genuine. The proposition is one the case actually stood for on the day it was decided. The trouble is that the case no longer states good law, and relying on it fails a filing just as surely as citing something invented. Bad law is not a hallucination in the strict sense, because there is no misreading involved, but it belongs in the same review because the consequence at the courthouse is the same.

Bad law matters here because it exposes a gap that the two hallucination checks leave open. Existence resolution confirms the case is real, and it will be. Quotation and proposition matching confirms the description is accurate, and it will be. Both checks pass, and the citation is still defective, because neither check asks the question that matters for currency: is this still the law. That is a distinct question with a distinct tool. Folding bad law into the review is what prevents a citation from clearing two honest checks and failing on the one nobody ran.

Matching the right check to each failure

Each failure mode maps to a specific verification method, and using the wrong one leaves a gap. Existence resolution catches fabrication: attempt to retrieve every cited authority, and anything that does not resolve is invented. Quotation and proposition matching catches misgrounding: pull the real opinion and confirm each quote appears in it and each proposition is actually supported. A citator catches a third, adjacent problem, bad law: a citation that is real and accurately described but has been overruled, vacated, or superseded. Bad law is not a hallucination in the strict sense, because the AI read the case correctly, but it fails a filing just as effectively. Running only one of these checks leaves the others open. A complete review confirms three things about every authority: that it exists, that it is quoted and applied accurately, and that it remains good law.

The reason the mapping has to be explicit is that each check is blind to the failures the others catch. Existence resolution says nothing about whether a real case was described correctly. Quotation matching says nothing about whether a correctly described case is still valid. A citator says nothing about whether a valid case was invented, because it assumes the case is real to begin with. The checks are not interchangeable and they are not redundant. Each one closes a door the others leave open, which is why a review that runs a single check, however carefully, is only ever a fraction complete.

A three-part review that covers every authority

Because the checks are non-overlapping, the only complete review is one that runs all three against every citation, not a triaged subset. The order is practical rather than doctrinal: it is efficient to fail fast on existence before spending time reading an opinion that turns out not to exist. But no step is optional and no citation is exempt. The clean-looking entry gets the same three-part treatment as the suspicious one, precisely because looking clean is the property that both fabrication and misgrounding are optimized to have.

1. Resolve existenceAttempt to retrieve every cited authority from a reliable source. Anything that does not resolve is treated as fabricated until proven otherwise, regardless of how ordinary it reads.
2. Match quotation and propositionFor each authority that does exist, read the opinion and confirm that every quotation appears in the text and every proposition is actually supported, including the subtler failures where a holding is stretched past the facts that produced it.
3. Confirm it is still good lawRun each surviving citation through a citator to catch cases that are real and accurately described but have been overruled, vacated, or superseded.

What this means for signing a filing

The reason to separate these modes is not taxonomy for its own sake. It is that the person whose name goes on the filing is accountable for all three failures, and a review built to catch only the dramatic one leaves the quieter ones live. Fabrication produces the headlines, but misgrounding and bad law fail a filing just as reliably and are far easier to miss, because they arrive wearing the costume of a valid citation. A review discipline that stops at "the case is real" is calibrated to the least dangerous of the three problems.

None of this requires trusting a vendor's assurance about how the underlying model works, and it should not. Grounding is an input-side improvement measured, in the one study cited here, as still leaving real documents described wrong at rates above 17% and above 34% ^[1]. The output-side obligation does not move: the authorities in a brief have to be confirmed to exist, confirmed to say what the brief claims, and confirmed to still be good law, by someone who checked rather than assumed. The honest promise a tool can make is that it will make those checks faster and more consistent. No tool can honestly promise that the checks are no longer necessary. This article is general information about failure modes and verification, not legal advice, and it does not substitute for a lawyer's independent review of every authority before filing.