The Hallucination Tax: AI Citations and Arbitral Efficiency

Parties leave the courthouse for mediation and arbitration in pursuit of a single promise: resolution that is quicker, less costly, and more final than public litigation. That promise holds only if the papers placed before the neutral can be taken at face value. A well-built submission narrows a dispute; it does not widen it.

Generative artificial intelligence threatens that narrowing in a specific and preventable way. A small but conspicuous share of filings now arrive seasoned with invented authority: decisions that were never handed down, statutes with no enacting body, and record cites that lead nowhere. These are not freak occurrences. They are the foreseeable result of asking a language model to do the work of a verified legal database.

For dispute resolution the reframe is straightforward. A fabricated citation is not only the filing lawyer's problem. It is a cost forced onto the opposing party, onto the neutral, and onto the proceeding itself. Locating that cost, and identifying who bears the duty to prevent it, now belongs to competent advocacy and to competent neutral practice in equal measure.

When invented authority surfaces, the reflex is to fault the instrument. The lawyer protests that the system is opaque, that it produced something false without flagging it, that the slip belonged to the machine rather than the advocate. That account has worn thin, because it misstates how the technology behaves.

A general-purpose large language model is a probabilistic text generator, not an index of verified law. Asked for a case on a point, it does not retrieve an authority; it assembles the string of words most likely to resemble one. A citation that looks correct is precisely what the system is engineered to deliver, regardless of whether the decision behind it ever existed. The behavior is a feature of the design, not a defect in a given session.

That mechanical fact maps cleanly onto the duty of competence. Model Rule of Professional Conduct 1.1, read with its commentary on technology, expects a lawyer to grasp the benefits and the hazards of the tools in use. The obligation to confirm the record cannot be delegated downward to a junior lawyer, and it certainly cannot be handed to a chatbot. When fabricated authority reaches a filing, the lapse is one of professional judgment rather than of code.

The plainest harm is economic, and it lands hardest on the party that did nothing wrong. In arbitration the tribunal bills by the hour, so each hour spent confirming that a cited decision does not exist is an hour charged to the parties. One fabricated citation begins a sequence of waste.

Opposing counsel first spends time chasing an authority that cannot be found, then drafts correspondence raising the discrepancy with the tribunal. The neutral, owing an independent duty to apply the law correctly, must personally verify that the cited rule is fictitious. The matter then breeds collateral disputes over conduct and remedy that crowd out the contract or privacy question the parties actually came to settle.

Courts have begun to attach consequences. In Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), the court sanctioned counsel who filed a brief resting on nonexistent decisions and, when pressed, produced fabricated text purporting to be those opinions. Federal courts have since required offending lawyers to bear fees the opposing side incurred running down invented citations. In a loser-pays arbitration, the party that files hallucinated authority is, in practical effect, agreeing to fund its opponent's cleanup.

Beyond hours and dollars, invented authority corrodes the asset that makes a private forum efficient: the neutral's readiness to accept a submission at face value. A neutral who finds one fictitious decision in a brief cannot responsibly presume the remainder is sound. The benefit of the doubt is gone.

Once that confidence breaks, the neutral must read everything from that party skeptically, re-checking factual and legal assertions that would ordinarily have passed without challenge. The proceeding slows, the award takes longer to compose, and the billed hours rise. The advocate has not merely damaged one citation; the advocate has repriced their own credibility for the rest of the matter.

Judicial commentary has been blunt about the underlying point: it is not the tribunal's task to comb a record separating genuine citations from invented ones. Arbitrators are engaged service providers, and asking them to sort the reliable from the fabricated rarely flatters the party that created the sorting. Trust, once spent, is costly to rebuild.

Arbitration carries a stake litigation does not: the award must survive enforcement, frequently under the New York Convention. If a tribunal relies, even inadvertently, on a fictitious legal principle introduced by counsel, the resulting award acquires a vulnerability.

A losing party may contend that the award offends public policy, or that it could not fairly present its case because it was made to argue against law that does not exist. The advocate who slips invented authority into the record may therefore embed a defect inside the client's own victory, handing the loser a credible basis to resist enforcement and drawing the client into extended post-award proceedings.

Appellate treatment underscores the seriousness. In Park v. Kim, 91 F.4th 610 (2d Cir. 2024), the court addressed an attorney who relied on a chatbot to generate a brief containing a nonexistent citation, and referred the conduct for disciplinary consideration. A tribunal with authority to manage its proceedings holds the functional equivalents: it can strike an offending submission or draw adverse inferences serious enough to hollow out the claim.

None of this argues against AI in dispute resolution. Tools that use retrieval-augmented generation to ground answers in verified primary law are already speeding legitimate research. They multiply professional judgment; they do not replace it. The meaningful line runs between assistance and abdication.

For the neutral, the task is to police that line without recoiling from the technology. That means setting expectations at the outset, treating an unverifiable citation as a competence problem rather than a novelty, allocating the cost of cleanup to the party that caused it where the rules allow, and holding the sharper remedies, striking submissions or drawing adverse inferences, for genuine intransigence. The aim is to preserve the efficiency the parties bargained for, not to penalize modern tools.

For advocates, a modest protocol closes the exposure. Keep general-purpose chatbots to brainstorming and non-legal drafting, never to producing citations. Reserve authority work for legal-grade tools that link to primary law. Require the drafting lawyer to open each cited source and confirm that the decision exists, the pin cite is accurate, and the proposition is genuinely supported. Maintain a short verification log, and run the table of authorities through an established database before filing, so that anything unrecognized stops the submission until a person confirms it. The maxim is old and still sufficient: trust, but verify. In this setting, anything less drifts toward malpractice.