The opinions had names, citations and quotations. The court said they did not exist. That is the starting point in the accessible opening of the Mata v. Avianca sanctions order: a documentary failure, followed by continued defense of the documents after the court questioned them. Court order.
This is an archive case, not a new incident. The focus here is the verification boundary described in the order, not a reconstruction of private conversations or an assessment of today’s products.
What does the court record establish?
The opening says that nonexistent judicial opinions, including fabricated quotations and citations generated by ChatGPT, were submitted to the court. It also says the respondents stood by those opinions after judicial orders called their existence into question. Those are findings described in the court’s opening paragraphs. We could not retrieve the complete order and do not reconstruct its full procedural history here. Court order.
The court drew a distinction worth preserving. Assistance from a reliable AI tool is not inherently improper. Attorneys still have a gatekeeping responsibility for the accuracy of filings. The visible record supports that distinction; it does not support a blanket conclusion that AI assistance in legal work is forbidden.
What had the product developer said?
OpenAI’s original introduction warned that ChatGPT can produce plausible but incorrect or nonsensical answers. That is the developer’s own general limitations statement. It predates this case and should not be presented as a response to the sanctions order. OpenAI’s introduction.
The warning matters to this review because plausibility is precisely the wrong stopping point for a citation check. A convincing case name is a lead to investigate, not an authority ready to submit.
Where should the paper trail have stopped?
Our recommended control is concrete: every cited authority must resolve to an original record that supports the proposition for which it is used. If the record cannot be found, the citation remains unresolved and cannot pass the review gate.
A practical review record would contain:
- The proposed citation and the exact claim it is meant to support.
- The original opinion or other authoritative record, with matching identifying details.
- The relevant passage and a check that the quotation is accurate.
- The reviewer’s disposition: supported, contradicted or unresolved.
This is our proposed prevention pattern, not a claim that the court prescribed this particular form or that a checklist alone guarantees error-free work. The final approval must depend on the evidence actually being inspected.
What should readers take from this case?
Investigate the source before trusting the surface. If a challenge arrives later, reopen the evidence check. Do not turn the generated answer itself into proof that its sources exist.
Use the citation verification guide to turn that principle into a review sequence, then connect unresolved findings to the prevention workflow. The lesson travels beyond legal citations: the record supporting a business decision has to survive inspection.
The source record
Read the original evidence and the scope of our review.
- Mata v. Avianca: Opinion and Order on SanctionsU.S. District Court, Southern District of New York · 2023-06-22 · Accessed 2026-10-11Opening paragraphs read through substantive indexed text; full order and sanctions disposition were not retrievable.
- Introducing ChatGPTOpenAI · 2022-11-30 · Accessed 2026-10-11Indexed limitations passage read; full page unavailable. This is a general product warning, not a response to this court case.
