When Arbitrators Use AI: LaPaglia v. Valve and the Boundaries of Adjudication — A TRW Law Firm Guide (Dhaka • London • Dubai)

Executive takeaways

• AI can help with logistics; it cannot replace judgment. Using AI to organise records, generate outlines, or check citations is very different from allowing a tool to shape factual findings or legal conclusions.
• Transparency is the safety valve. Disclosing intended AI use (and its limits) early in the case—and inviting party comment—greatly reduces challenge risk.
• The U.S. test is “exceeding powers” and due process. The argument in LaPaglia v. Valve frames heavy AI reliance as an arbitrator exceeding the parties’ agreement to receive a reasoned decision by a human neutral, and as a due-process problem if facts outside the record are introduced without party input.
• Soft-law guidance is converging. Leading guidelines emphasise non-delegation of decision-making, source verification, and party consultation.
• Plan now. Add a short “AI Protocol” to Procedural Order No. 1 (PO1) and to terms of appointment. This protects award integrity and enforceability.

For a wider overview of how we design efficient, enforceable procedures, see International Arbitration.

1) The case that sparked the debate: LaPaglia v. Valve—what’s at stake?

A consumer claimant petitioned a U.S. federal court to vacate an AAA award, alleging the sole arbitrator “outsourced his adjudicative role” to AI. The petition points to (i) alleged comments about using a chatbot in other contexts, (ii) a compressed award timeline, and (iii) purported “AI tells” in the text, along with uncited facts not in the record. The challenge proceeds under the Federal Arbitration Act (FAA) §10(a)(3) (misconduct) and §10(a)(4) (exceeding powers).

Why this matters beyond the U.S.: even though national vacatur standards differ, most modern arbitration regimes revolve around party consent, equality of arms, and the tribunal’s duty to deliberate and give reasons. If a party can plausibly show that a non-human system substituted for deliberation—or that unvetted, extra-record material crept into an award—the award’s legitimacy is at risk anywhere.

2) Where AI adds value in arbitration (and where it doesn’t)

Helpful, low-risk uses (with human control):

• Organising the record: deduplication, exhibit indexing, chronology building.
• Language support: drafting scaffolds, translation aids, plain-language rewrites for non-decisional text (e.g., procedural summaries) subject to counsel review.
• Template hygiene: tables of contents, citation formatting, defined terms lists.

High-risk uses (or outright no-go without disclosure and party input):

• Evidence synthesis and credibility findings. AI should not determine what the record “proves” or who is credible.
• Legal reasoning and holdings. Tools may help outline, but the tribunal’s own analysis must drive the outcome.
• Fact material from outside the record. Any third-party retrieval or inference must be disclosed and opened to comment.
• Undisclosed reliance on systems that hallucinate sources or mischaracterise authorities.

3) The emerging consensus principles

Across the most-cited guidance documents, the core principles align:

1. Non-delegation: Arbitrators must not delegate their adjudicative role to AI.
2. Verification: Any AI-assisted output used in the process must be checked against the record and reliable sources.
3. Transparency & party consultation: If AI tools will be used beyond purely administrative tasks, tell the parties and give them a chance to be heard.
4. No extra-record surprises: If a tool surfaces facts or authorities outside the record, flag them and invite comment before they influence the award.
5. Responsibility: Arbitrators remain entirely responsible for the award—style, substance, and accuracy.

4) How LaPaglia maps onto challenge grounds

(a) Exceeding powers (FAA §10(a)(4) and analogues)

Parties typically agree to a human neutral who will deliberate and issue a reasoned decision. If AI is used such that it replaces deliberation or authorship on substantive issues, a challenger may argue the arbitrator exceeded the mandate. The closer AI use gets to decisional outsourcing, the stronger the challenge.

(b) Misconduct / due process (FAA §10(a)(3) and international analogues)

If an award contains facts not in the record (or drawn from undisclosed AI outputs) without giving parties a chance to respond, a court could find fundamental unfairness. Internationally, similar concerns arise under “unable to present one’s case” and “public policy” refusal grounds at the enforcement stage.

(c) Evidentiary hurdles for challengers

Proving AI reliance can be difficult. “AI-detector” opinions are notoriously unreliable. Stronger indicators include process evidence (disclosures, emails, tool usage logs), textual anomalies coupled with extra-record content, or internal inconsistencies that cannot be reconciled with the record. Ultimately, courts ask: did the arbitrator decide, or did a tool?

5) Practical playbook for tribunals: an “AI Protocol” for PO1

Clause 1 — Purpose & scope
“The Tribunal may use digital tools to manage logistics and drafting. The Tribunal shall not delegate any part of its decision-making to any AI system.”

Clause 2 — Transparency
“The Tribunal will inform Parties if it intends to use AI tools beyond administrative functions (e.g., drafting scaffolds, translation aids). Any such use shall be explained in general terms.”

Clause 3 — Record integrity
“The Tribunal will not rely on AI-generated information outside the evidentiary record without first disclosing the material and allowing Parties to comment.”

Clause 4 — Verification
“The Tribunal will independently verify any AI-assisted text included in orders or awards against the record and authoritative sources. References must be accurate and checkable.”

Clause 5 — Confidentiality & security
“Any AI tool used by the Tribunal shall be operated in a manner that does not upload case data to public or third-party training sets. Use is restricted to secure environments with no model training on case content.”

Clause 6 — Party use of AI
“Parties shall disclose if AI systems materially assist in drafting evidentiary submissions (e.g., expert analytics). They must verify accuracy, cite sources, and certify no confidential data was exposed to public models.”

Clause 7 — Responsibility
“The Tribunal assumes full responsibility for all procedural rulings and the reasoning and conclusions of any award.”

This dovetails with our broader procedure-first approach; see International Arbitration.

6) Counsel strategy: safeguarding the record (and the award)

• Ask early. At the first case management conference, request that the tribunal state on the record how (if at all) it intends to use AI.
• Lock in the protocol. Propose the PO1 language above, plus data-security and confidentiality terms for any tool usage.
• Police the record. If an order or draft award references uncited facts or non-existent sources, promptly request clarification and an opportunity to be heard.
• For party submissions:
• Use closed/enterprise tools only; ban uploads to public models.
• Human-verify all outputs; keep a source log.
• Disclose tool assistance where it touches substantive analytics (e.g., expert modelling).
• If challenging an award: build proof around process defects (lack of disclosure, extra-record reliance), not around AI-detector screenshots.

7) What arbitral institutions and appointing authorities can do now

• Issue standing guidance that reaffirms non-delegation and verification.
• Update terms of appointment to include basic AI and data-security undertakings.
• Provide safe toolkits (secure drafting and evidence environments) so tribunals need not improvise with public systems.
• Encourage simple disclosure templates—e.g., “The tribunal used tool X for formatting only; no external sources were consulted beyond the record.”

8) Ethics, confidentiality, and data protection

Even where AI use is procedurally acceptable, arbitrators and parties must honour confidentiality and data-protection obligations. That means:

• No ingestion of confidential case data into public or retrainable models.
• Prefer on-premise or enterprise AI with no training on user inputs.
• Redact personal data and special category data where possible; maintain transfer logs when cases span Dhaka–London–Dubai infrastructures.
• If in doubt, don’t upload—run tools locally or within a vetted hosting environment.

For our end-to-end approach to data-secure arbitrations, see International Arbitration.

9) Likely future trend lines

• Soft law hardens into practice. Disclosure/verification norms will become standard in PO1 and terms of appointment.
• Tooling bifurcates. Tribunals and institutions move to private, auditable AI for admin tasks; public chatbots fade from professional use.
• Challenge jurisprudence evolves. Courts will focus on process integrity (non-delegation, party participation, accuracy), not on per se bans of technology.
• Training and accreditation. Expect arbitrator rosters to require AI-competency and data-security modules as table stakes.

10) A concise checklist you can adopt tomorrow

For tribunals

• Add an AI paragraph to your terms of appointment.
• Disclose intended use; invite comment.
• Keep AI away from fact-finding and legal conclusions.
• Verify every proposition against the record before it hits the award.
• Use secure, no-training environments only.

For parties

• Propose an AI Protocol in PO1.
• Human-verify any AI-assisted text; keep source logs.
• Monitor for extra-record references; object early.
• If challenging, focus on non-delegation and due process, not detectors.

How TRW Law Firm can help (Dhaka • London • Dubai)

We design technology-aware arbitration procedures that preserve fairness and enforceability:

• Drafting PO1 AI Protocols and updating terms of appointment.
• Advising tribunals and parties on defensible AI use and data security.
• Stress-testing awards for process vulnerabilities before issuance.
• Handling set-aside/recognition strategies where AI use is alleged.

Explore our broader approach at International Arbitration and speak with our cross-office team about building an AI-resilient procedure for your next case.