The single strongest driver of fabricated citations is the distance between what your brief needs to say and what the authorities actually hold. It arrives four ways — a position the law is against, a corner of practice with barely any law, a citation count the law can’t fill, and a template slot with nothing to put in it — and every one of them ends with an obliging model manufacturing the support.
Probability map. Many people are under the mistaken impression that an LLM is some sort of hypercomplex thinking machine. Really it’s just an incredibly huge probability map of words, patterns, and phrases followed by other words, patterns, and phrases. Unless it’s attached to a database of authoritative sources, nothing connects what it generates back to reality.
Completion pressure. Every model runs under a standing directive that strongly favors providing an answer rather than tell you it can’t. Hand it a gap where a real authority should be, and it fills the gap instead of flagging it.
Latent contradiction. Your LLM prompt can look fine on its face, but against the backdrop of law, non-obvious contradictions can be hiding in wait. That conflict is invisible to you, and sometimes even to the model itself. Something has to give, and completion pressure decides what: it honors the request and manufactures the rest.
A model that drafts legal documents is trained to complete the task in front of it. Hand it a favorable set of facts and a conclusion to reach, and its job — as it has learned the job — is to produce a persuasive brief that gets there. If genuine authority exists, it uses it. If it doesn’t, the pressure to finish the assignment doesn’t go away.
That single gap — between the support your brief needs and the support the law actually provides — shows up in four ordinary ways. The law is against you. There is barely any law. You asked for more citations than the law can fill. Or your template opened a slot where no authority belongs. They look like four different drafting problems. They are one problem, and the model resolves each of them the same way: it manufactures the missing support.
The further your desired outcome sits from settled law, the more the model has to invent to close the distance: a case that stands for exactly your proposition, a quotation that says exactly what you need, a court that ruled exactly your way.
It might result in dissent reported as the majority, a case that never existed, or even a quotation that never appeared. You asked for a brief that argues a position the law doesn’t support, so the model probabilistically fabricates support.
The single strongest driver of fabricated citations is the distance between what your brief needs to say and what the authorities actually hold. When real support doesn’t exist, an obliging model manufactures it.
This is the version everyone recognizes once it’s named, and the version nobody thinks they’re doing. You are not asking for a fake case. You are asking for the best available argument for a position — which is the job. The model treats “the best available argument” as “an argument that wins,” and where the authorities won’t get there, it supplies authorities that will.
The same cause, a few ways it turns up in ordinary practice. Each is routine work you’d never flag as risky — which is exactly how the cause slips in unnoticed.
When the LLM is prompted to come up with a legal theory that doesn’t exist, it can confidently invent one — along with the falsified authority to back it up.
The law is full of exceptions to nearly every rule, carve-outs wherever a doctrine bites — a fair-use defense, a good-faith exception, an implied waiver, some narrow situation that fits your facts and little else. The LLM sees this pattern everywhere, so it’s mathematically easy for an LLM to manufacture a case that recognizes the exception you need, even if no such case exists.
The model can invent a case that distinguishes the bad facts, or a quotation that says the rule doesn’t apply to your situation — even if no such case or quote exists.
Fabrication clusters exactly where authority is scarce — a novel question, a small jurisdiction, an unpublished corner of practice. The model carries a strong assumption that a governing case must exist, and where reality is emptiest, it interpolates one into being.
A model’s picture of the law is built from what it saw in training. For common questions in busy jurisdictions it saw a great deal. For a question of first impression, a rule that turns on one small state’s intermediate court, a local court’s own procedure, or a specialized administrative corner, it saw little or nothing.
Scarcity doesn’t make the model cautious; it makes the model creative. Its prior expectation is that law on point exists — briefs are supposed to have supporting cases — so when it can’t retrieve a real one, it generates a case shaped like the answer the question seems to call for. The thinner the real authority, the higher the fabrication rate.
That’s the counterintuitive part: the areas where you most need a reliable answer, because you can’t easily check it yourself, are the areas where the model is least reliable and most confident. Novelty and confidence rise together.
Completion pressure in LLMs can manufacture a citation for your novel theory, creating direct support for a case that “recognized” it, a quotation that “supports” it, a court that “adopted” it.
The reality of any frontier — a new technology with no case law yet, an issue of first impression in your circuit, a question the appellate courts simply haven’t answered.
The draft supplies a controlling case anyway — invented — that resolves the open question cleanly.
Anywhere the reporters run thin — a territorial or tribal court, a small state’s intermediate appeals, an old unpublished line no database fully indexed.
The model cites a state appellate opinion that doesn’t exist, with a reporter cite that resolves to something else.
Specialties thin on reported law invite it — a rare tax provision, a niche regulatory doctrine, a sub-field where only a handful of opinions exist.
>_Draft a brief arguing that the statute creates a private right of action for our client. Make the argument as strong as you can and back every point with controlling authority.
Nothing in it says “invent a case.” But “back every point with controlling authority,” aimed at a holding the law never reached, is the latent contradiction — and completion pressure settles it by manufacturing the authority.
Volume is a multiplier. The moment a request calls for more supporting cases than the model genuinely has on point, it stops retrieving and starts generating — and the fabrications cluster in the back half of a long brief.
Ask for “a memo with a couple of leading cases” and a model usually reaches for real, well-known authority. Ask for “forty cases supporting this,” a thirty-page brief, or “a case for every element,” and the arithmetic changes. Real on-point authority is finite. The quota is not.
Once the request outruns what the model actually knows, every additional citation has to come from somewhere. It comes from the model’s sense of what a citation looks like — a plausible party name, a plausible reporter volume, a plausible year — assembled into a string that reads exactly like the real ones beside it.
This is why long, citation-dense filings hallucinate more than short ones, and why the invented cites tend to appear late, after the genuine authority is spent. The brief looks most thoroughly researched exactly where it is most fabricated.
It’s the reflex of every thorough advocate — “give me a citation for every point,” a full survey of the case law, a brief that leaves no supporting authority unmentioned. The first many are real, but fabrication creeps in when the model runs out of genuine authority to cite or it just starts pattern matching.
Any elements test invites it — the four factors of a preliminary injunction, the elements of fraud, any multi-prong standard where each prong wants its own case. When you ask for a separate case for each element, the model finds the real ones and invents the rest.
The instinct to shore up a weak spot — “add a couple more cites here,” “is there anything else on this,” one more authority to make the paragraph land. It doesn’t even have to be thin law — it’s permuting a probability map of words, and a dead end means you get a plausible-looking citation that doesn’t exist.
It’s the reflex of every thorough advocate — “give me everything on point,” a full survey of the case law, a brief that leaves no supporting authority unmentioned. The first many cites are real, but then the model keeps producing fakes because they simply match the word patterns.
>_Write a comprehensive brief on this issue and cite at least fifteen cases that support our position.
The number is the trap. When real on-point authority runs out before fifteen, completion pressure keeps producing citations rather than stop short of the count you set.
A rigid brief skeleton with a citation slot behind every heading tells the model something dangerous: that a supporting authority exists for each one. When one doesn’t, the slot still gets filled.
Fill-in-the-blank drafting — a standard motion skeleton, a form brief with “[cite]” behind each point heading, an outline that pairs every argument with a supporting case — is efficient precisely because it’s structured. That same structure carries an implicit instruction: every slot has an answer.
A model working through a template treats an empty citation slot the way it treats any other blank: as something to complete. It will not leave the slot empty and note “no authority found.” It fills it, because the template’s shape says a citation belongs there, and the surrounding real citations make an invented one look native.
The failure is quiet because the output looks disciplined. A neatly structured brief with a citation under every heading reads as well-researched. The structure is real; some of the authority filling it is not.
Every firm runs on them — a house motion-to-dismiss shell, a summary-judgment template, a brief bank where each heading comes pre-wired with a citation slot. The model filled the slot rather than leave it empty, even if the fill is fake.
Drafting grids that pair every cause-of-action element with a “supporting authority” field — the tidy table that assumes a case exists for every row.
The “see, e.g.” and “accord” slots that expect three to five parallels — the string cite that looks thin with only one case behind it.
## Argument I — [point heading]
Supporting authority: [case, with pincite]
## Argument II — [point heading]
Supporting authority: [case, with pincite]
## Argument III — [point heading]
Supporting authority: [case, with pincite]Every “[case]” slot tells the model an authority belongs there. For the heading with none on point, the empty slot is the latent contradiction — and it gets filled rather than left blank.
Adding a citation directive (“only cite real cases”) mathematically cannot ever fix the problem. Humans read a citation directive and think the LLM must comply. But every LLM has completion pressure built into its architecture, a mathematical property of the model itself, that will readily override such a directive.
The defense isn’t a cleverer prompt — it’s a check that runs after the draft exists. Verbatim reads the finished brief and reports, for every authority it cites, whether the cite resolves to a real case, statute, or regulation, and whether the quoted language actually appears at the pin cite. A citation invented to prop up a weak position has nowhere to hide: it doesn’t resolve, and the report says so before the brief leaves your desk.
What you do with that is a legal judgment that stays with you — strengthen the argument, find real support, or reconsider the position. The point is that you learn the support is imaginary from a report, not from a show-cause order.
Verbatim’s answer to scarcity is simple: if the cited authority isn’t among the real cases, statutes, and regulations it checks against, the report says it isn’t there — it never fills the gap for you. Where a model interpolates a case into a thin area, the verifier returns an empty result instead of a fabricated one. “No authority found” is an answer you can act on; an invented citation is not.
Its coverage spans federal and state case law, the U.S. Code and all fifty state codes, and federal regulations — so a real but obscure authority resolves and links to its source, while a fabricated one simply doesn’t. That distinction is the whole point.
Verbatim doesn’t care whether a brief cites five authorities or five hundred; every one gets the same two questions — is it real, and does the quote appear where it’s cited. A report on a citation-dense brief simply surfaces more findings, and the fabricated back half lights up the same as the front. The length that hides the problem from a tired reviewer doesn’t hide it from the report.
The count you can actually stand behind is the count Verbatim verifies. If the law supports a dozen cases, padding to forty just means twenty-eight findings to resolve — better on your screen than in a filing.
Verbatim treats every filled slot as a claim to be checked, not decoration. It resolves each citation in the assembled brief and confirms the quoted language against the source, so a slot filled to satisfy a template — rather than because a real case belongs there — surfaces as a finding. The empty slot the model refused to leave empty becomes visible again.
A template makes drafting fast; a verification pass makes the fast draft safe to file. The two are complementary, and both belong before the caption page ever reaches a clerk.
Verbatim reads a finished brief and reports, for every authority it cites, whether the cite is real and whether the quoted language actually appears at the pin cite — so a fabrication surfaces on your screen, not in a show-cause order. Bring a brief and we’ll walk you through the report.