Don't Murder Your Credibility With AI

In Roald Dahl’s creepy short story “Lamb to the Slaughter,” a wife uses a frozen leg of lamb to kill her husband before cooking and feeding the weapon to the police officers investigating the crime. The brilliance of that story is that a leg of lamb isn’t ordinarily an instrument of death. It’s dinner. And only becomes lethal because of how the wife chooses to use it.

Generative AI is the leg of lamb of modern arbitration practice. Used properly, it can nourish a case by summarizing a mountain of exhibits, drafting a first-pass chronology, tightening prose, translating witness statements, flagging inconsistencies between expert reports, and completing a host of other time consuming (and expensive) tasks.  Used badly, it can bludgeon credibility.  

Most litigators are familiar with court sanctions, including disqualification, for citations to cases hallucinated by AI.  If a federal judge will disqualify counsel in a court case, imagine the impact of such a mistake in arbitration, where the checks and balances of a jury and appellate review don’t exist and, and the rules of procedure and evidence are relaxed. Without those institutional filters, the arbitrator’s judgment about who is credible is the primary filter. Destroy it, and the results may devastate your case. (A similar caution applies to any arbitrator who may be tempted to let AI do the heavy lifting in deciding a case. But that’s a subject for another day.)

A few arbitral institutions have published guidance on the proper use of AI in arbitrations: (1) “SVAMC Guidelines on the Use of AI in Arbitration” (Silicon Valley Arbitration & Mediation Center, April 2024) — a principle-based framework stressing confidentiality, human oversight, and non-delegation; (2) “SCC Guide to the use of artificial intelligence in cases administered under the SCC rules (SCC Arbitration Institute, October 2024) — confidentiality, quality, integrity, and no delegation of decision-making to AI; (3) “AAA/ICDR Guidance in Arbitrators’ Use of AI Tools 2025”  (American Arbitration Association - International Center for Dispute Resolution, March 1, 2025) — focusing on arbitrators’ use of AI and setting guidelines to preserve accuracy, fairness, independent decision-making, and transparency; and (4) “CIArb Guideline on the Use of AI in Arbitration” (The Chartered Institute of Arbitrators, September 2025) — focused on procedural guidelines for the use of AI in arbitrations. As might be expected, these guidelines address the issues presented by AI use in broad strokes. However, certain recurring themes are unmistakable: transparency, confidentiality, verification, and human responsibility.

The upshot: AI can be used in arbitration so long as the user is transparent and verifies the results. The following is a non-exhaustive list of “Dos and Don’ts” to help achieve this result:

DO keep a human firmly on the hook. AI may draft but the attorney must verify and own the result. Every citation, every fact, every quote gets checked against the primary source. No exceptions.

DO use retrieval-augmented prompting. Feed the model an actual document and ask it to work from that — not from its memory. This is the single most effective way to cut hallucinations. However . . .

DO protect confidentiality. Never paste confidential, privileged, or sensitive material into a public tool. Use approved, vetted, closed systems. Redact and anonymize where appropriate.

DO keep a prompt log that includes the tool's name and version, a short description of its use, and the complete prompt and output. Keep those records.

DO check the ground rules first. Confirm what the tribunal's procedural order, the institutional rules, and any applicable law say about AI before you use any AI tool.

DO disclose when it matters. Use is not always disclosable in the abstract — but if it affects evidence or procedural decisions, get ahead of it. It is best to address AI use and the rules that will apply early on.

DO stay competent. AI literacy is now part of professional diligence. If you don't understand what the tool does with your data, you're not ready to use it.

DON’T submit unverified output. It’s a mistake to serve unverified evidence, authorities, or analysis to the tribunal – hoping nobody notices. Someone will.

DON’T put confidential material into public models. Data-handling practices are often opaque, and what goes in may not stay private.

DON’T assume that authority cited by an AI exists or stands for the proposition cited.  Always verify.

DON’T create a record you can’t reproduce. If you can’t show what you fed into the tool and what it returned, you can’t defend the output.

DON’T hide it. Non-disclosure is itself a credibility risk — and in some circumstances, a sanctionable one.

In the Dahl story, the leg of lamb was never the problem. Someone improperly used it as a weapon, then fed it to the people whose job was to find it. The lamb was neutral. The choice wasn't.