The Hallucination Tax: AI Negligence, Sanctions, and the Duty to Verify

Every few weeks another lawyer is caught filing a brief that cites cases which do not exist. The instinct is to call it a technology problem — the model misled them, the black box failed. That defense no longer holds. A generative model that invents a citation is not malfunctioning; it is doing exactly what it was built to do. The cost of that output does not land on the software. It lands on the person who signed and filed it, and it spreads across the entire proceeding. That spreading cost is what I call the hallucination tax.

Why the model fabricates — and why that matters

A large language model is a probabilistic engine, not a truth machine. When you ask it to find a case on a point of law, it does not query a database of verified authority. It predicts the sequence of words most likely to look like such a case. Recent research from the model developers themselves makes the point bluntly: training and evaluation procedures reward confident guessing over admitting uncertainty. Hallucination is therefore a feature of how these systems work, not an occasional glitch. A lawyer who uses a general-purpose model for legal research without understanding that mechanism is flying the plane without knowing which gauge reads fuel and which reads altitude.

This matters because the standard for using the tool is not whether the output looks plausible. It is whether the person filing it verified that the output is true. The model's confidence is irrelevant. Its fluency is, if anything, a liability — fabricated material reads exactly like the real thing.

Two kinds of hallucination, one duty

Hallucinated content

The familiar failure is fabricated authority in the body of a document: a phantom case, an invented statute, a quotation that appears nowhere in the cited opinion, a pin cite to a page that does not support the proposition. These are the errors that draw show-cause orders. Courts have been clear that there is nothing improper about using a reliable AI tool for assistance — the rules impose a gatekeeping role on the lawyer to ensure the accuracy of every filing. Using AI is permitted. Abandoning the verification that has always accompanied a signature is not.

Hallucinated metadata

The less-discussed failure is quieter and, in some ways, more dangerous. When AI generates or assists in generating a document, it can populate the hidden metadata fields — Author, Company, Created date, version history — with fabricated values. A team can scrub every sentence and every citation in a brief, file it confident that it is clean, and still ship a document whose properties name an author who does not exist or a creation date that predates the matter. In litigation, metadata is not a technical curiosity; it is evidence. Courts and opposing counsel rely on it to authenticate documents and establish when and by whom something was created. A hallucinated Created date that looks like backdating, or a fictitious reviewer in a privilege log, can undermine credibility on facts no one ever thought to check.

The danger compounds because forensic and eDiscovery tools report what the file contains, not whether the values are true. X-Ways, EnCase, Axiom, and Relativity will display a hallucinated hash value or timestamp with all the apparent authority of ground truth. Both layers — the surface properties any reviewer can open, and the deep fields only forensic tools surface — can carry invented data that passes undetected unless it is cross-checked against independent evidence.

The hallucination tax

A single fabricated citation does not stay contained. Opposing counsel burns hours trying to locate a case that was never written, then drafts correspondence flagging the discrepancy. The court verifies the non-existence of the authority itself. Satellite litigation follows — a show-cause order, a sanctions motion — and the actual dispute stalls while everyone litigates the conduct of the lawyer. Each of those hours is billed to someone. In 2025 one court ordered sanctioned counsel to reimburse the opposing party more than twenty-six thousand dollars for the fees spent investigating fake citations. In arbitration, where loser-pays is often the default, a party that files hallucinations is writing a blank check to its opponent.

The deeper tax is the loss of trust. A tribunal or judge who finds one fabricated case will distrust every other assertion in the brief. The benefit of the doubt is gone, and the entire submission gets read with a forensic, skeptical eye. That is a self-inflicted wound no efficiency gain justifies.

This is a competence and reliability problem

The exposure runs along three familiar tracks, and none of them are new law. The duty of competence — and its now-standard comment on keeping abreast of the benefits and risks of relevant technology — makes verification a non-delegable professional obligation. You cannot outsource your judgment to a chatbot any more than to a summer associate whose work you never read. Rule 11 puts the lawyer's signature behind the accuracy of every filing. And when AI touches an expert's work product, Rule 702 demands that the opinion reflect a reliable application of reliable methods — an expert who cannot account for what the model did, and prove it, has an exclusion problem. The technology did not change any of these standards. It raised the stakes for proving you met them.

A verification workflow that holds up

The answer is not to ban the tools. Used well — particularly retrieval-augmented systems grounded in verified primary law — they are real force multipliers. The answer is a disciplined, documented protocol that treats AI output as a draft to be proven, never a result to be trusted. For lawyers, arbitrators, and experts working with generative AI, the following workflow is a defensible baseline:

1. 01Restrict the tool to the task. Use general-purpose models for brainstorming, summarizing, and non-substantive drafting — never for generating case citations. Reserve citation work for legal-grade tools that link directly to primary law.
2. 02Click through every citation. No AI-generated cite enters a filing until the drafter has personally opened the primary source and confirmed the case exists, the pin cite is accurate, and the text actually supports the proposition asserted.
3. 03Run a table-of-authorities check before filing. Pass every cite through a traditional legal database. Any authority that fails to auto-populate or flags as unrecognized halts the filing until the source is manually retrieved or removed.
4. 04Inspect the metadata, not just the text. Before producing or filing, open the document properties and confirm the Author, Company, and Created and Modified dates are correct — and reconcile them against the deeper fields a forensic tool would surface.
5. 05Cross-check metadata against independent evidence. Because forensic tools report what the file contains rather than whether it is true, treat any timestamp, hash, or author value that conflicts with the known record as suspect until corroborated.
6. 06Keep an audit trail. Maintain a source-verification log in which a named person initials that each authority was checked and each metadata field reconciled. The log is your proof of reasonable process if the work is ever questioned.
7. 07Keep a human accountable for consequential calls. Privilege determinations, responsiveness decisions, and substantive legal conclusions are confirmed by a person, not delegated wholesale to the model.

The black box is a mirror

When a model produces garbage and that garbage is filed, the failure reflects the diligence of the filer, not the limits of the software. The hallucination tax is avoidable, and the price of avoiding it is the verification discipline that has always accompanied a lawyer's signature and an expert's opinion. Trust, but verify — anything less, in 2025 and beyond, edges toward malpractice.

If you are defending against an opponent's AI-assisted filing, questioning the authenticity of produced metadata, or building a verification protocol your firm can stand behind, scope the engagement early. Start a conversation through our home page or email the team directly to discuss a conflict check and approach.