Sanction teardown · TJ San Paulo, Brazil · 2025-12-01
José Carlos Pinto de Faria v. Prevent TWB do Brasil et al.
What happened
In TJ San Paulo, Brazil, a filing relied on an unnamed/unconfirmed AI tool to help draft legal argument. The court identified the following problems with the citations in that filing:
-
Fabricated (Case Law)Attorney reproduced an ementa/precedent at fl. 6 that the court could not locate; number was incomplete (marked by repeated 'x'); court concluded it was an AI-generated, nonexistent precedent and not a real decision.
Which AI tool
an unnamed/unconfirmed AI tool. Note: Charlotin's public database records tool attribution only where a court order, brief, or reporting on the matter states it explicitly; "unidentified" or "implied" means the record indicates AI use but does not name a specific product — we do not guess.
Outcome
Monetary Fine; Bar Referral (monetary penalty: 1 BRL.)
Additional detail
The court found that the agravante (through his counsel) reproduced a 'precedent' ementa that does not correspond to any real judgment and contained an incomplete case number (marked with repeated 'x'), concluding it was likely an AI 'alucinação' (fabricated citation) resulting from misuse of an intelligence‑assistance tool. The majority treated this as litigância de má-fé and imposed a fine of 2% of the updated value of the cause and ordered a copy sent to OAB/SP; a dissenting judge would have excused the attorney for lack of dolo and declined to impose the fine.
How Citation Safe would have caught this
Citation Safe runs three deterministic layers before a brief is filed: (1) does the citation exist against CourtListener's database of published opinions, (2) if quoted, does that exact language appear in the source, (3) does the cited case actually support the proposition it is cited for. Fabricated case citations fail Layer 1. Fabricated or misattributed quotations fail Layer 2 even when the underlying case is real. Misrepresented holdings — a real case cited for a proposition it does not support — are the target of Layer 3. None of these checks involve asking another language model whether the citation looks right; they are lookups and text-matches against the actual source, which is why a hallucinated citation has to survive a direct lookup against the authoritative source — not another model's opinion — to earn a VERIFIED stamp; our measured false-verify rate is published live at /quality.
Check a brief before you file it → · See our live false-verify rate
Source: https://www.damiencharlotin.com/documents/2861/20250001272706.pdf, via Damien Charlotin's public AI Hallucination Cases Database (CC0).