What is AI Citation Sanctions?

Why It Matters

The sanctions record is the clearest evidence that unverified AI output is a professional risk rather than a technology risk. Courts stopped treating fabricated citations as an exotic accident some time ago, and the penalties now land on firms and named attorneys rather than on the tools that produced the text.

The Ninth Circuit’s order is the one worth reading in full, because it states the principle that governs everything else: the rules are not violated at the point of research and drafting, but at the point of signing and filing. Whatever produced the error, the signature is where accountability attaches.

What Courts Have Held

The duty is personal. In California, an appellate court held that delegating citation verification to a paralegal does not satisfy it. The lawyer who signs owns the check.

Pro se is no excuse. New York’s Appellate Division, Second Department sanctioned a self-represented filer over an AI error and made the point explicitly.

Amounts are climbing. Illinois imposed $50,000 on a firm and $10,000 on the individual attorney for fabricated citations across post-trial submissions. New York has ordered $2,500 and $8,000 penalties. Florida fined an attorney $1,500, made him personally liable for the opposing party’s fees, and referred him to the bar.

Candour compounds it. In several matters the sanction was increased by how the error was handled afterward: claiming a fabrication was a typo, or failing to disclose that a tool produced it.

Where It Breaks

Firms respond with a policy that says citations must be checked. That restates the duty without installing a control, and the Reuters reporting three years into the sanctions era shows the volume still rising. The gap is a check that runs on every filing.

Delegation is the second failure. Verification pushed to the least senior person available is the pattern the courts have now specifically rejected, which means the check has to produce something the responsible lawyer actually reviews.

The third is tool trust. A research platform that grounds answers in a case-law database reduces fabrication but does not remove it, and the professional rule does not care which tool produced the error. Grounding changes the base rate; the signature still carries the duty.

How Flytebit Handles It

Verification is a gate rather than a guideline: each citation is checked against the source it names, and an unsupported claim is flagged to the reviewer before the document can be signed. The check runs outside the model, so it does not inherit the failure it is looking for, and the decision record shows which authorities were verified and who signed. The industry application is on our Professional Services page, and the evaluation method behind the check is in Evaluating Agentic AI Systems.