Fabricated Authorities: The Check Before AI Work Leaves the Building

AI writing tools produce invented citations. Not occasionally, not because the model is broken, and not in ways that will be fixed by the next version. Understanding why this happens changes how you design your review process.

Why Fabrication Is a Property, Not a Bug

A language model does not retrieve documents. It predicts text. Given the opening of a legal memo, it predicts what words come next, based on patterns learned from enormous quantities of human writing. Those patterns include: cases have names, volume numbers, page numbers, and jurisdictions. Reports have authors, dates, and publishers. Statutes have titles and section numbers.

When the model produces a citation, it is producing the shape of a citation, assembled from learned patterns. If a real case exists that matches the shape it produces, that is a coincidence. If the name, volume, reporter, page, and year all match a real decision, that is a lucky coincidence. The model has no mechanism to check. It does not access a database. It does not know what exists.

This matters because the failure is not correctable at the generation stage. A better prompt does not fix it. A more expensive model reduces the rate but does not eliminate it. The only fix is a human checking the output against a real source before the output leaves your hands.

The legal profession learned this in public. In 2023, filed documents in US courts contained invented case citations generated by AI tools. The cases did not exist. The citations were detailed and plausible. Lawyers were sanctioned. The citations were not sloppy, which is the part worth sitting with.

Why Fabricated Material Reads Better Than Real Material

This is the counterintuitive part, and it explains why "it looked right" is such a common failure signature.

Real cases are messy. The holding you need is buried in paragraph thirty-seven. The language is technical, sometimes archaic, sometimes contradictory within the same opinion. Real quotations are inconvenient. They say almost what you need but not quite.

Fabricated material has none of these problems. The AI produces a citation that fits the argument perfectly, because the argument came first and the citation was generated to match it. The language is clean. The holding is precise. The year is plausible. The court is the right court for the jurisdiction.

This means fabricated citations feel more useful than real ones. They slot into a draft without awkwardness. They match the surrounding prose in register and clarity. The temptation to trust them is built into how they read.

Any verification habit you design has to account for this. The check cannot be "does this look right?" because looking right is exactly what a fabricated citation does best.

The Specific Verification Step

For anything cited or quoted in a professional document, the check is the same. It has three parts.

First: locate the source independently.

Do not search for the citation by copying it into a search engine. That is a search for evidence that the citation exists, and you may find secondary material that repeats the invented citation. Go to a primary source. For cases, that means a legal database such as Westlaw, LexisNexis, Fastcase, or the relevant official court database. For statutes, use the official legislative text. For reports and studies, go to the publisher's website or the original document.

Search for the case name without the volume and page number first. If the case exists, you will find it. If it does not appear, the citation is likely fabricated.

Second: verify the specific detail the document relies on.

If the document quotes language from the case, find that language in the source. Read the surrounding paragraphs. Confirm the quote is accurate and that it means what the document says it means in context. Summaries of holdings are frequently wrong even when the case is real. The AI may cite a real case for a proposition the case does not actually support.

This step is slower than the first, but it is the step that matters. A real case can be cited for the wrong proposition. That error is subtler than a fabricated citation and equally damaging.

Third: record what you checked.

Keep a simple log: the citation as it appeared in the draft, the source you used, the date you checked, and whether it verified. This takes thirty seconds per citation. If a question arises later about what was reviewed, you have an answer. If your firm has a quality control process, the log is evidence that the process ran.

"It Looked Right" Is the Exact Failure Signature

Every reported case of a fabricated citation passing review involved someone who found the citation credible. That is the failure mode. The person who reviewed it made a judgment based on plausibility rather than verification.

Plausibility is not the same as accuracy. In professional work, plausibility is not enough.

The implication is structural. You cannot train people to look more carefully at citations, because the problem is not that the citations look careless. You can train people to run the verification step regardless of how plausible the citation appears. The step is not triggered by suspicion. It is triggered by the fact that AI was involved in the draft.

This is a process question, not a judgment question. The check should be as automatic as spell-check. It runs because it runs, not because something looked wrong.

Making the Check Fast Enough That People Actually Do It

The main reason verification does not happen is friction. If checking a citation requires opening three systems, navigating a complex database, and writing a report, people will skip it under deadline pressure.

Here is a practical approach that reduces friction.

Keep the verification log as a simple spreadsheet or shared document. Columns: citation, source checked, date, result (verified / not found / found but proposition wrong), initials. Two minutes total per citation if the source is accessible.

Create a short list of the primary sources your work draws on most frequently and bookmark the direct search pages. If you cite a particular court regularly, bookmark that court's opinion search. Reduce the number of clicks between "I need to check this" and "I am searching."

Set a firm rule: AI-assisted drafts carry a note in the file that the verification step is required before issue. This makes the check a named step, not an assumed one. A named step is harder to skip.

For longer documents, assign the verification pass to a specific person rather than assuming the drafter will do it. The drafter is close to the work and more likely to accept plausibility as confirmation. A second person is more likely to run the check mechanically.

A Worked Example: Verifying One Paragraph

Here is a paragraph an AI tool might produce.

The duty of care in professional negligence cases was clarified in Harrington v. Calloway, 412 F.3d 881 (7th Cir. 2005), where the court held that a professional who relies on third-party information without independent verification cannot claim that reliance as a defence where the error was foreseeable. The principle has been applied broadly in subsequent decisions.

Verification pass:

Step one. Search Westlaw (or the 7th Circuit's official opinions database) for "Harrington v. Calloway." Result: no case found. Search for "Harrington" and "Calloway" in 7th Circuit cases in 2005. No result. The citation does not appear to correspond to a real decision.

Step two. Because the case is not found, step two does not apply. The proposition cannot be verified because the source does not exist.

Step three. Log entry: "Harrington v. Calloway, 412 F.3d 881 (7th Cir. 2005): not found in Westlaw or 7th Circuit database. Citation likely fabricated. Removed from draft. [Date] [Initials]."

The paragraph is removed or replaced with a verified authority before the document is sent.

The whole pass took four minutes. The risk removed was significant.

A Note on Policy

If your firm uses AI tools, the verification requirement belongs in a written policy, not in informal expectations. Informal expectations do not survive deadline pressure or staff changes.

A written AI usage policy makes the verification step a stated requirement, clarifies which tools are approved for which tasks, and gives staff a clear reference when questions arise. If you are working on a policy for a legal or professional services firm, this guidance from Plainstart covers the structure: AI policy for law firms.

The Short Version

AI tools fabricate citations because they predict text, not because they retrieve documents. Fabricated citations read well because they are built to fit the argument. "It looked right" is not a defence and not a check. The check is: find the source independently, verify the specific language or proposition, and log what you did. The check runs on every AI-assisted draft, not just on the ones that feel suspicious.

This article is general process guidance, not professional or legal advice. What specific obligations apply in your setting depends on your engagement terms, your professional body, and the jurisdiction you work in. Your own adviser is the right person to assess that.