Lawyers are generating legal briefs and memos with AI at astounding rates. But AI has hallucinations — cases, statutes, and authorities that look real but are completely fake. And now attorneys have suffered sanctions and suspensions for it. Verbatim reads a brief and produces a report that says, for every authority cited, whether the cite is real and whether the quoted language is actually in the cited opinion.
Fabricated authorities are no longer a curiosity — they're a bar‑discipline problem. Judges across federal and state benches now demand AI‑use disclosures on filings, issue show‑cause orders sua sponte when a cite doesn't check out, and refer attorneys to their state bars. The list of sanctioned lawyers grows every month — and careers are altered over citations that took ten seconds to make up.
Interim suspension; the retained attorney was separately fined $1,500 and ordered to self‑report her to disciplinary authorities.
A prosecutor suspended six months over fabricated cites in a murder appeal — the new‑trial order was vacated and remanded.
AmLaw 5, 150‑year pedigree — apologizing to a federal judge for an AI hallucination.
30‑day suspension, one‑year probation, and 10 hours of technology CLE.
$59,235 in the opposing party’s fees, plus $15,000 to the clerk — for a frivolous appeal built on nonexistent authorities.
$10,000 on the attorney and firm, $3,000 on co‑counsel; orders served on two state bars.
≈$31K to the defense for fees and costs — a “collective debacle,” “tantamount to bad faith.”
Anthropic’s own outside counsel — a Claude‑invented citation that survived a manual review.
The first published US discipline of its kind — a year‑and‑a‑day suspension.
The case the whole profession points to — $5,000 sanction and a referral to bar disciplinary authorities.
Do you want your name on this list? Do you want opposing counsel awarded sanctions for finding your fabricated AI citations? Every disaster started with at least one unverified cite that made it into a filing.
See Sanctions ListCases. Federal and state statutes. Federal and state regulations. Agency decisions. Full cites, short cites, Id., supra. The brief that mentions a case in passing four pages after the full cite still gets caught, resolved, and checked.
Each citation is checked against the corpus. Real cites resolve to a link straight into the source opinion or section. Parallel cites collapse to the same case. Statutes and regulations resolve to the version in effect on the filing date. Fabrications are flagged, named, and surfaced.
Every quoted span is checked against the cited opinion at (or near) the pin cite — and against the cited subsection for statutes and regulations. Words inserted that aren't in the source, paraphrases passed off as quotes, quotes attributed to the wrong page — all caught.
Most of what a brief cites is fine — and the report stays quiet about it. Its job is to pull your eye straight to what needs work: the misquotes, the version drift, and above all the fabricated authorities that don't exist at all. Over 40 different checks are run on every single citation to deep dive on authenticity. Findings are ranked by severity, so the critical, falsified material is the first thing you see — and the last thing left to fix before you file.
The cite is real and formatted correctly, and the quoted words appear directly in the authority, with pin if necessary — however it's cited, full, short form, Id., or supra — with nothing to flag. Verbatim is the gold standard.
The cite is real and resolves to the authority, but either it uses bracket substitutions, it's a quote with immaterial variations that come up due to document formatting, or it's paraphrased/summarized accurately rather than cited directly.
The cite is real and its support is sound, but there's a non‑critical issue in the citation's formatting or immaterial metadata — might annoy a judge, clerk, or other counsel for sloppiness, but not likely to draw sanctions.
Something material needs to be reviewed: the quote drifts from the source, the quote is found but the cite is wrong, or the quote is non‑reviewable because it refers to a procedural filing outside of context. Review the warning and address accordingly.
The citation doesn't exist. It's fake. This is very likely sanctions material. Even worse, if your brief or memo depends on it, you may need a new legal theory.
Drop the draft brief in. Receive a marked‑up report flagging every unverified cite and every misquoted passage — before the judge sees it, before opposing counsel checks it, before a disciplinary committee rules on it.
The brief bounced through three AI tools and four humans. The partner shouldn't have to trust the chain — the report shows what's verified, what's not, and what to fix, with a link to every authority.
Workflows ideally run Verbatim before filing. But certainly before approving payment, confirm that the brief outside counsel filed actually cites what it claims to cite. Catch the fabricated authority your $1,200/hour partner missed.
Send every AI‑generated document through Verbatim before it reaches the user. Show hallucinated cites in your UI before the user does anything with them. Per‑tenant API keys and quotas.
Bulk‑verify filings before they reach chambers. Spot‑check the authorities cited across a whole docket — or across a CLE program's materials — in one pass. The same verifier, applied at scale.
Prosecutors, public defenders, and agency lawyers file constantly and answer to bar discipline like everyone else — the Georgia AI‑cite suspension was a prosecutor. Clear every brief before it goes out, without adding a research attorney to the office.
Every verified cite in the report carries a link straight into the source authority at the pin cite. Open the opinion. Open the section. Read the surrounding text. The verifier doesn't just say it's real — it shows you where.
The U.S. Supreme Court, every federal circuit, every federal district that publishes, and every state supreme and intermediate appellate court. Parallel cites collapsed to one canonical opinion so any reporter resolves the same case.
Every section, every subsection, every historical edition retained so a brief dated 2014 verifies against the 2014 version — not whatever the current text happens to say.
The Code of Federal Regulations at every annual edition for historical lookup, the live state of the regulations for current cites, and Federal Register entries for proposed and final rules. State administrative codes where the state makes them available.
Labor, tax, immigration, securities, communications, energy, environmental, and trademark/patent administrative opinions. Same pipeline: cite extracted, existence checked, quote matched against the agency's own opinion.
A fabricated law review article reads as real as a fabricated case — the journal exists, the form is clean. Verbatim checks whether the article is real and whether the quoted language is actually in it, the same way it checks every other authority.
Daily delta ingestion. New opinions become verifiable within 24 hours of publication. Every report stamps the corpus version; old snapshots retained so any report can be re‑run against the corpus state it originally ran against.
Other tools tell you whether a case is still good law. Verbatim tells you whether your case is a real case — and whether the words inside the quotation marks are actually in it. That alone covers the failure mode getting attorneys suspended this year.
Pass raw text, upload a PDF or DOCX, or point at a document already in your DocPost vault.
Caller supplies jurisdiction and document_date, or Verbatim infers them from the caption and metadata — flagging every assumption on the finding it shaped.
Every verification produces a structured report you can drop straight into your workflow — JSON for your pipeline, or a filing‑ready PDF or annotated DOCX to review and hand off.
An advisory layer that answers “has this case been overruled, abrogated, or substantially criticised?” for every verified cite. Any authority it surfaces is run back through the deterministic verifier so a hallucinated overruling case can't sneak in. Explicitly advisory; explicitly not a Shepard's or KeyCite replacement.
Verifying not just quotes but characterisations: does the brief's summary of the case match what the case actually held? When a string cite asserts that Smith; Jones; Doe all stand for the same proposition, do they? Whose paraphrase fairly captures the authority and whose stretches it? An advisory pass over the unquoted prose that surrounds each cite.
See Verbatim work on a sample brief — yours, or one we'll bring. We'll walk through the report, the severity bands, and the link from every verified cite back to the source authority. From there, the same engine slots into your pre‑filing workflow or your AI pipeline.