90 incidents on record · 2026 Headlights Incident reports by Ellie Harris · Melbourne
10 new this week Library last updated 13 July 2026
← The incident library
HD-INC-025
Legal services · Australia · 2025 · Hallucination & fabrication

An immigration lawyer filed seventeen cases that did not exist, and the quotes from the tribunal were invented too

By Ellie Harris · Filed 31 January 2025

Alleged: Applicant's legal representative developed or deployed the AI system implicated in this incident. Details are drawn from public reports; parties are presumed innocent of any wrongdoing not established by an official finding.

An immigration lawyer filed seventeen cases that did not exist, and the quotes from the tribunal were invented too

What happened

In Valu v Minister for Immigration and Multicultural Affairs (No 2) [2025] FedCFamC2G 95, the applicant’s legal representative filed an amended application and an outline of submissions in the Federal Circuit and Family Court of Australia. The documents cited a series of Australian authorities in support of the applicant’s case, including passages presented as direct quotes from a decision of the Administrative Appeals Tribunal.

The cases could not be verified by the Court, because they did not exist. The quotes attributed to the Tribunal did not exist either. When the matter was examined, the legal representative acknowledged that he had used an AI tool, ChatGPT, to identify Australian case law, and that the tool had returned non-existent decisions and fabricated quotations that he had then put before the Court without checking them against any real source. The fabricated material amounted to seventeen non-existent case citations, together with fabricated quotations attributed to the Tribunal across eight separate passages, all identified by the Minister’s lawyers before the hearing.

The legal representative accepted that providing false case citations and quotes attributed to the Tribunal was a breach of his duty to his client and to the Court, and that his conduct fell short of the standard expected of a legal practitioner. Judge Rania Skaros referred him to the Office of the New South Wales Legal Services Commissioner. The case became one of the landmark Australian examples of generative AI hallucinations entering court documents, and it sits alongside a growing line of decisions in which practitioners have been sanctioned or referred for filing AI-invented authority.

What an auditable version would have shown

The thing that makes hallucinated citations so corrosive is that they are designed to look correct. A fabricated case carries a plausible name, a plausible court, a plausible year, and a plausible quotation. The Court caught these because a citation is checkable against an authoritative source, the actual law report, and they failed that check. The failure was that nobody ran the check before the documents were filed.

An auditable drafting workflow would treat every citation as something that has to be confirmed against a real legal database before a document can leave the practitioner’s hands, and would keep a record of the result. For each authority cited, that record would carry the citation as drafted, the database queried, and the outcome of the lookup, found with a matching report or not found. Run over the Valu submissions, it would have produced a column of seventeen not-found results before anything was filed. The relevant fact for a court or regulator afterwards is not only that the citations were false, but that no such check ever stood between the model’s output and the filed document.

Where the gap was

The gap was the absence of any verification routed to a trusted source, and any record that verification was done.

A CitationVerifier checks every citation in a document against a real legal, academic, or technical database before the document can ship, and refuses to pass a citation it cannot find. A VerificationGate generalises the same discipline: a claim the model proposes is routed to a trusted source for confirmation, never back to the model that produced it. The model proposes; a database disposes. Either control, applied to the Valu submissions, would have caught seventeen citations that no law report contained. A ConductRecord then preserves the result, so the difference between checked and unchecked is part of the file rather than a matter of later admission.

What happened instead is the default failure mode of generative AI in expert work: the model produced fluent, authoritative-looking output, and a professional trusted it because it looked like the real thing, with no independent check in between.

What governance should have looked like

The legal profession does not need to ban AI drafting tools to avoid the Valu outcome. It needs the verification step to be non-optional and recorded.

Every citation in a document bound for a court should be confirmed against a real source before filing, automatically, as a condition of the document being finalised, not as a discretionary check a busy practitioner might skip. The confirmation should be recorded, so that a practitioner can show the court that every authority was verified, and so that a firm or chambers can see at a glance whether the discipline is being followed. Courts in multiple Australian jurisdictions are now issuing practice rules requiring exactly this kind of verification and disclosure for AI-assisted documents. The practitioners who build the check into their workflow, and can prove it ran, will not be the ones being referred to the regulator. The ones who rely on the output looking right will be.

The reference implementation of CitationVerifier, VerificationGate, and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. Anyone can read every line and verify the signatures. The repository is public now.

Sources

The mailing list

Fresh incident reports every week. One email to match.

We add new incidents to the library regularly, and send a single short email each week with what's new. The library stays free and open; this is just how you keep up with it.

No tracking. Unsubscribe in one click.

The record

An auditable system would have produced a signed, tamper-evident record the moment this happened: what the system did, the version that did it, the basis it acted on, and the action taken, and Applicant's legal representative could have produced it on demand.

This is the record the system as deployed did not produce in a signed, auditable form.

What this teaches
Capture what happened when it happens
What the system did, the version that did it, the basis it acted on, and the action taken, recorded at the moment, not reconstructed after.
Sign it, so no one has to trust the record-keeper
A tamper-evident entry. Edit it later and the signature breaks. The record does not ask for the benefit of the doubt.
Make it verifiable by anyone
A court, a regulator, a customer's lawyer can check the record themselves, without taking the company, or us, at our word.

Headlights summarises publicly reported AI incidents. All summaries are independently written, attributed to their original sources, and intended for research and educational purposes. Allegations are identified as such until established through official findings.

Last reviewed June 2026. This report is based on the sources listed above and reflects information available at the time of review; later developments may not be captured. Where a person is described as charged with or alleged to have done something, that allegation is unproven unless a conviction or a court or regulatory finding is stated. Headlights publishes journalism and commentary, not legal advice.

Want to write back?

Direct to my inbox.

ellie@useheadlights.com →