PPROMPT WORKBOOK Build a prompt
Verified 27 July 2026

Verify before you cite

Hallucinated citations are not a hypothetical risk. There is now a public database of court cases where they made it into filings, and it has grown into the thousands.

How far a fake citation can travel

In March 2026 the California Court of Appeal published an opinion documenting a chain that is worth reading slowly. A pseudonymous author on a Reddit blog invented a case called Marriage of Twigg, complete with a plausible citation. From there it was attached as an exhibit to a client declaration. An attorney cited it in a filing. It was then included in a proposed order drafted by the opposing party's own counsel. A family court judge signed that order. The attorney cited it again on appeal, and when challenged, fabricated further detail about it.

At no point in that chain did anyone open the citation. The court sanctioned the attorney five thousand dollars and referred her to the State Bar.

The lesson is not “lawyers are careless”. Five separate people, including one adversary with every incentive to check, and a judge, all passed a fabricated authority along. A citation that looks correct is trusted by default. That default is the vulnerability, and it is not specific to law.

This is a documented pattern, not an anecdote

The first widely reported case was Mata v. Avianca in 2023, where a federal judge in New York sanctioned two attorneys five thousand dollars for a brief citing fabricated case law produced by ChatGPT. That was treated at the time as a cautionary oddity.

It was not an oddity. A public database maintained by the legal researcher Damien Charlotin now tracks court cases involving AI-hallucinated citations, and only logs instances where a court explicitly found or clearly implied reliance on fabricated material. It grew from roughly two hundred cases in July 2025 to 1,598 as of June 2026.

Three years after the problem became famous, it is accelerating rather than resolving.

Why models invent sources so convincingly

A citation is a highly patterned object. Case names, volume numbers, journal titles, years and page ranges all follow strong conventions, and reproducing that shape is exactly what a language model is good at. Producing a citation that looks right requires no knowledge that the underlying work exists. This is why fabricated references are typically more plausible-looking than fabricated prose, and why they slip past readers who would have caught a factual error in a sentence.

The failure mode is worse when the model is asked for support for a claim you have already made. You have supplied the conclusion; the most helpful-looking continuation is a source that supports it.

The working discipline

  1. Never let the model be the source

    Treat any citation it produces as a lead to check, never as a reference to reuse. If you did not open it, you did not read it, and you cannot cite it.

  2. Ask for what it used, not just what it concluded

    Require a quoted passage alongside each claim. A model that cannot produce the quote is usually telling you the source does not say that, or does not exist.

  3. Check existence separately from relevance

    These are two different failures. A real paper cited for a claim it never makes is just as wrong as an invented one, and it is far harder to spot.

  4. Watch for the confident specific

    Precise volume and page numbers feel like evidence of retrieval. They are the easiest part to generate. Specificity is not verification.

  5. Give explicit permission to fail

    Most fabrication happens because the model has no acceptable way to return nothing. Tell it that “no source found” is a correct and expected answer.

For every factual claim, provide:
1. the claim
2. a direct quotation from the source that supports it
3. where the source can be found

If you cannot supply a real quotation, write NO SOURCE FOUND and move on.
Do not reconstruct a citation from memory of the general literature.
A short answer with three verified sources is better than a long one with ten I have to check.

That last instruction matters more than it looks. Without it, a model asked for ten sources will produce ten, because producing ten is what was asked for.

Why confidence is not evidence

The mechanism behind fluent wrongness, and how to make claims traceable instead.

Read the guide →

Check your work properly

Turning an impression that the output is good into something measurable.

Read the guide →

When the content is hostile

Why material the model reads can carry instructions of its own.

Read the guide →

Sources