The audit trail that protects the lawyer who signed. Veritas checks every citation, confirms the cases say what your brief claims, and leaves you a defensible record for every AI-assisted brief, before it leaves the firm.
The number above is our best estimate of monetary sanctions, cost orders, and fee-shifting awards levied on U.S. attorneys for filings containing AI-generated citations that could not be verified. It climbs from the conservative floor (publicly-reported direct sanctions only) to the defensible ceiling (including published cost orders and diverted court time).
Every entry below is from the public record: a sanctioned attorney, a disqualified counsel, a published warning. The professional rules have not changed: a lawyer is responsible for every citation in a filing they sign, regardless of who or what produced it.
Mata v. Avianca, Inc. Two attorneys submitted a brief citing six ChatGPT-generated cases that did not exist in any reporter. The judge wrote that he might not have sanctioned them had they come clean, but they didn't.
Mata v. Avianca, 678 F. Supp. 3d 443 (S.D.N.Y. 2023)↗A Florida attorney was suspended from practicing in the Middle District of Florida after filing pleadings citing cases the court determined did not exist.
In re Neusom · 1-year suspension, March 8, 2024↗A California appellate court fined an attorney for filing an opening brief in which 21 of 23 quoted authorities could not be located in any reporter. The court published the opinion as a warning to the bar.
California 2nd District Court of Appeal, Sept. 2025↗Johnson v. Dunn. An unverified AI-generated citation in a large firm's filing led the court to disqualify the offending attorneys, publish the opinion in the Federal Supplement, and direct the clerk to notify bar regulators in every state where they were licensed.
Johnson v. Dunn, No. 2:21-cv-1701 (N.D. Ala., July 23, 2025)↗A federal court in Oregon ordered a lawyer to pay $109,700 in sanctions and costs for filing AI-generated errors, reported as one of the largest such penalties to date.
Reported by NPR, April 2026↗Public court records across U.S. and international dockets have logged more than 1,200 decisions involving unverified AI-generated citations, roughly 800 of them in U.S. courts as of April 2026. The count climbs every week.
Veritas AI sanctions tracker↗We are not a drafting tool. We are not an assistant. We sit between whatever tool your firm already uses and the filing your name will appear on, and we make sure the work is defensible.
We check every case you cite against the official reporter. Is the case real? Does any language you quoted actually appear in the opinion? And does the case actually support the point you're using it for? Three questions, answered for every citation.
Every check is logged with its source and the time it ran, in a record that can't be edited after the fact. The finished report is sealed to your original file, so if anyone changes the file later, it shows. You end up with proof of what was checked, and when.
Plain, careful verdicts you can hand to a partner, a client, or a malpractice carrier. When a citation can't be confirmed, the report says "not located in reporter." It never overstates. The language is measured on purpose, so the report holds up to scrutiny itself.
“No brief, pleading, motion, or any other paper filed in any court should contain any citations, whether provided by generative AI or any other source, that the attorney responsible for submitting the pleading has not personally read and verified.”
Every firm's filings, citations, and verifications are separated at the database layer with PostgreSQL row-level security. The application connects under a restricted role that cannot return another firm's rows, even if a single query were to omit its filter. Isolation is enforced by the database itself, not by application code alone.
Every connection is secured with TLS. Stored data is encrypted at rest with AES-256.
Every scan, verification, and account action is written to an immutable, per-firm audit log. An administrator can export the full trail at any time for the matter file or a procurement review.
Each Verification Certificate carries a SHA-256 hash of its contents and a timestamp, so a filed artifact can be shown to be unaltered.
Access is gated by a managed identity provider with organization-scoped membership. Single sign-on is available on enterprise plans.
Veritas verifies citations and quoted language against public court reporters and public-record sources. The verification posture is built on public data, and the application runs in United States regions.
We publish a control here only once it is live in production. Veritas is not yet SOC 2 certified and does not currently offer on-premise deployment or EU data residency. Enterprise customers receive current security documentation for procurement review.
Veritas is not a research tool. Veritas is not a writing tool. Veritas runs after the brief is drafted and before it is filed.
Veritas pulls every citation out of your draft, checks each one against the official reporters and public court records, marks where each one stands, and produces a Verification Certificate. The certificate lists what was checked, how each citation came back, the records it was checked against, a shareable link, and a tamper-proof seal with the date and time.
The certificate is suitable for the matter file. The certificate is suitable for a malpractice underwriter. The certificate is suitable for the partner who will sign the brief.
The wording is careful on purpose. When a citation can’t be confirmed, the certificate says “not located in reporter.” When a quote doesn’t match the opinion, it says “differs from reporter.” Veritas never overstates what it found. The recommendation is direct; the verdict is honest.
Plans for solo practitioners, small teams, and enterprise firms, with unlimited filings on every plan.
Every scan confirms two things for each case you cite: that the case is real (it exists in the official reporter at the volume and page cited) and that any language you put in quotation marks actually appears in the opinion. That catches a made-up or mis-quoted citation, and it runs on every filing, unlimited. A deep scan goes further and reads the opinion to confirm the case supports the point you're using it for.
Veritas never guesses and never accuses. When a case can't be located, the report says "not located in reporter" and recommends manual review — it does not say "fabricated." The careful, hedged language is deliberate, so the report holds up to scrutiny itself and protects the attorney who relies on it.
Veritas is a verification layer, not a research tool, so it doesn't replace or plug into Westlaw or Lexis. It checks your finished filing against the public court record, independently of whatever you used to draft it. That independence is what makes the verification certificate trustworthy.
A citation in the WL or LEXIS format usually points to an unpublished opinion that lives only in a subscription database, with no public reporter citation. Veritas flags these for manual review rather than guessing. You can confirm them in your own Westlaw or Lexis account, or upload the opinion and Veritas will verify the quoted language against it.
Veritas verifies citations in publicly filed court documents and is built on a public-record verification posture. Data is encrypted in transit and at rest, and every check is written to an immutable audit log. Criminal Justice Information — rap sheets, NCIC/III returns, defendant identifiers, Social Security numbers — is detected and blocked at upload, and the original filing is not retained after its citations are parsed.
No. The work that could pass a citation is deterministic comparison against the source, not a model's opinion about your brief. We look each cited case up against an owned corpus of public court reporters, and we match your quoted text character-by-character against the actual opinion. A model is used in only three narrow places, and in all three it can only make the result more cautious, never less: extraction recall that catches citations whose format our patterns missed so they get checked too but can never pass one, a fail-closed filter on treatment that can rule a flag out as a false alarm but never vouch for a citation, and an optional proposition analysis grounded in the retrieved opinion. See the methodology page for the full walkthrough.
No. We strip bracket characters and keep the inner text, so a standard editorial bracket like “[t]he” matches the source cleanly. For an ellipsis, we recover the omitted opinion text and show it in a side-by-side redline so you can see exactly what was cut. Legitimate editing passes. At the same time, substantive text placed inside brackets still has to appear in the opinion, so the check holds the inner text to the source.
A single near-miss token, for example a changed number, a swapped name, or an altered negation, is held below the verified threshold and routed to a review warning rather than reading as verified or as a failure. You see the specific span and a prompt to compare it against the source. A single altered word can be the most consequential change in a brief, so we surface it for your eyes rather than guessing what was intended.
Pilots begin with a single practice group and a real matter. Bring a brief. We'll show you what a defensibility report looks like for your own work.
