Quebec's appeal court ordered a self-represented litigant to pay $1,000 in costs for filing four ChatGPT-fabricated court decisions

Court-origin record. The $1,000 costs order, the ChatGPT-fabrication basis and the four non-existent authorities are quoted verbatim from the Quebec Court of Appeal’s own signed reasons (2026 QCCA 1092), which are hosted on CanLII and mirrored by the Damien Charlotin AI-litigation tracker, and are corroborated firsthand by two genuinely independent newsrooms (a Canadian Press wire carried by CBC News and Law360 Canada).

The problem

This is an AI-governance honest-negative, not a deployment win: a court’s response to unverified generative-AI output filed as legal argument (source). In a family-law appeal over the division of assets following a divorce, a self-represented appellant filed a memorandum that “relies on two Supreme Court judgments and two Court judgments that do not exist” and that “ascribes to those decisions legal principles that he reproduces as quotations, even though those principles are no more real than the judgments themselves” (source). The Court noted the appellant was “not ill-equipped to deal with legal proceedings and legal argument” and “works as an IT consultant for his own firm” (source).

What was built

No system was built or deployed; the appellant used a general-purpose chatbot to research and draft his submissions (source). When questioned by the Court, “the appellant conceded that these authorities are the product of the artificial intelligence ChatGPT” (source). The Court recorded that the failure was not innocent: the appellant “discovered they did not exist during a consultation with counsel even before filing his memorandum, but he did not consider it necessary to amend his memorandum to remove the fabricated references and quotations before filing it, attributing this to an oversight on his part” (source). A Canadian Press wire carried by CBC News reports the same basis, “fake legal cases generated by ChatGPT”, and that the man learned the cases did not exist during a consultation with a lawyer but filed his arguments without removing them (source). Law360 Canada, reporting the case independently, likewise records the appellant’s “deliberate” citation “of ChatGPT-generated fake case law, including fictitious quotations” (source).

The outcome

The Court treated the conduct as a sanctionable breach, holding that “no one can assume that the result obtained is free from hallucinations produced by the engine used” (source). Citing non-existent decisions “whether generated by artificial intelligence or otherwise … amounts to a substantial breach in the conduct of the proceeding under art. 342 C.C.P.” (source). On costs, the Court held it would “reduce the amount of $5,000 claimed by the respondent and order the appellant to pay her $1,000 in legal costs as compensation for the time devoted to the matter and the work performed, in addition to bearing the costs of the appeal” (source). The dispositive order “CONDEMNS A. P. to pay the amount of $1,000 to S. K.” within 30 days of the appeal judgment (source). The Court dismissed the principal appeal, dismissed the application for leave to present new evidence, and allowed the incidental appeal in part, all with costs against the appellant (source). CBC’s report independently confirms the $1,000 costs figure and the ChatGPT basis (source), and Law360 Canada reports firsthand that “Quebec’s top court has ordered a self-represented family law appellant to pay $1,000 to his ex-wife plus her legal costs on the appeal” (source).

Corroboration and weak-source honesty

Court-origin, and carried firsthand by two genuinely independent newsrooms: the $1,000 costs order, the ChatGPT concession and the four fabricated judgments are all in the Court of Appeal’s own signed reasons (Tier 1), and both a Canadian Press wire (source) and Law360 Canada’s independent bylined report (source) corroborate the $1,000 and the ChatGPT basis firsthand. Weak-source honesty: the two independent secondaries carry different weight: the Canadian Press wire (CBC News and Global News run the same CP copy, counting as one independent secondary) prints only the facts common to a wire summary, while the Law360 Canada corroboration rests on its public headline and meta description because its article body is paywalled; neither prints the 2026 QCCA 1092 citation, the docket or art. 342 C.C.P., which rest on the Court’s own reasons (source). The green badge is the checker’s call against the public record, then the owner’s to validate, never the litigants confirming their own matter.

Conflation guard

Do not merge this with Specter Aviation Limited c. Laprade, 2025 QCCS 3521, a separate Quebec Superior Court decision that fined a different self-represented litigant $5,000 for ChatGPT-fabricated citations, and which is cited as a footnote inside these reasons (source). The $5,000 in this story is the amount the respondent claimed and the Court reduced, not a sanction; the operative figure is the $1,000 the Quebec Court of Appeal ordered (source).

How this was verified

Method: the critical facts (the $1,000 costs order, the “product of the artificial intelligence ChatGPT” concession, the four non-existent judgments, the “substantial breach” art. 342 C.C.P. framing and the Aug. 11, 2026 disposition) were quoted verbatim from the Quebec Court of Appeal’s own signed reasons (2026 QCCA 1092, court-origin Tier 1). This remediation pass re-checked those figures against the official reasons on CanLII, the neutral-citation host, and against the Damien Charlotin tracker’s S3 mirror, cross-checked with two independent newsrooms: a Canadian Press wire carried by CBC News (Tier 2) and Law360 Canada’s bylined report (Cristin Schmitz, Aug. 18 2026, Tier 2). The byte-tie on the mirror is closed: the S3 copy is captured at web.archive.org (2026-08-17 snapshot) with a CDX SHA-1 digest of FAMRHEZE3D6EJ7M5RQZ4HJK3GTATEHUI, byte-identical to the local PDF. Sources re-checked live 2026-08-24.

Sources

Tier legend: Tier 1 primary (court’s own signed reasons); Tier 2 strong independent press. Sources re-checked live 2026-08-24.

1. Court of Appeal of Quebec (Cour d’appel du Québec) · A. P. v. S. K., 2026 QCCA 1092 (official signed reasons on CanLII, the neutral-citation host) · 2026-08-11 · Tier 1 (court’s own reasons) · https://www.canlii.org/en/qc/qcca/doc/2026/2026qcca1092/2026qcca1092.html
2. Damien Charlotin AI-litigation tracker · same signed reasons, S3 mirror (byte-tied to Wayback Machine 2026-08-17, CDX SHA-1 FAMRHEZE3D6EJ7M5RQZ4HJK3GTATEHUI) · 2026-08-11 · Tier 1 (mirror of court’s own reasons) · https://www.damiencharlotin.com/documents/2830/AP_v._SK_Canada_11_August_2026.pdf
3. CBC News (carrying a Canadian Press wire) · “Quebec man ordered to pay $1K over use of fake AI-generated court cases” · 2026-08-16 · Tier 2 (independent secondary) · https://www.cbc.ca/news/canada/montreal/quebec-fine-ai-court-cases-9.7309184
4. Law360 Canada (LexisNexis Canada), Cristin Schmitz · “Quebec Appeal Court dings family self-rep $1K plus costs for ‘deliberate’ use of bogus ChatGPT cases” (bylined original reporting, distinct from the CP wire; article body paywalled, public headline and meta description carry the $1,000 and ChatGPT-fabrication basis firsthand) · 2026-08-18 · Tier 2 (independent secondary) · https://www.law360.ca/ca/estates/articles/2515051

Related case files

• A Missouri appeals court fined a pro se litigant $10,000 after 22 of 24 cited cases turned out to be AI fabrications, the closest parallel on the self-represented-appellant axis: an appellate court pricing a non-lawyer’s unverified AI citations.
• The Ontario Law Society Tribunal ordered a lawyer to pay $31,150 in costs after Grok-generated fake citations, the nearest Canadian case pricing the cost of AI-fabricated authorities, here against a professional rather than a self-rep.
• The Ninth Circuit fined two lawyers $2,500 each and suspended them six months over AI-hallucinated briefs, the U.S. appellate counterpart, where the court also stressed the breach lies at signing and filing, not in using AI.
• The Federal Court of Australia made an indemnity-costs order after AI-fabricated citations, a Commonwealth appellate parallel treating fabricated authorities as a costs question.