AI in Mediation: Confidentiality & Disclosure Rules (2026)
When the Mediator's Assistant Is a Machine: AI, Confidentiality, and Fairness in 2026
What new court rulings, ethics opinions, and disclosure rules mean for anyone bringing generative AI into the mediation room.
Mediation has always depended on a simple promise: what is said in the room, stays in the room. Generative AI is testing that promise in ways most mediation agreements were never drafted to anticipate. Over the past two years, courts, ethics boards, and bar associations have started to answer the hardest questions — and the answers are not the same everywhere.
Confidentiality and Privilege: A Tale of Two Approaches
Mediation confidentiality is a creature of statute, not an inherent privilege, and the statutes are not uniform. The Uniform Mediation Act, adopted by many states, protects "mediation communications" from disclosure — but that protection can be waived, and courts increasingly treat a public AI tool as an outside third party rather than a participant in the process. Feed a settlement brief into a consumer chatbot in a UMA state, and you may have opened the communication to someone outside the mediation, waiving the very privilege you meant to protect.
California takes a different path. Its Evidence Code excludes mediation communications from evidence as a matter of public policy, not as a waivable privilege, and the state's high court has held there is no implied waiver even when those communications surface in open court. Sharing a document with a mediator or an AI tool retained for the mediation does not, by itself, break that protection under Evidence Code section 1119.
Federal court-referred mediation adds a third layer: 28 U.S.C. § 652 keeps those communications confidential subject to narrow exceptions, and while no reported decision yet applies this directly to AI, the reasoning courts have used for attorney-client privilege in litigation — that an AI platform is a tool, not a person bound by confidentiality — is likely to carry over.
Absent a California-style exclusionary rule, assume that anything typed into a public AI tool has been shown to a third party. Use closed, contractually-bound systems for anything prepared for mediation, and put the ground rules in writing before the session starts.
A Sample Confidentiality Update
Courts Now Want to Know When You Used AI
The 2023 sanctions in Mata v. Avianca, where a filing cited cases a chatbot had invented, reset expectations across the federal bench. Individual judges have since issued standing orders requiring lawyers to say whether generative AI touched a filing and to certify that a human verified every citation and fact.
These rules vary by court rather than following one national standard:
- N.D. Texas (Judge Starr): a certificate at filing stating whether generative AI was used, with confirmation of human review.
- E.D. Pennsylvania (Judge Baylson): affirmative disclosure naming the tool and confirming citations were checked.
- D. New Jersey (Judge Padin): identification of the specific tool, a description of which sections it touched, and certification of attorney review.
- W.D. North Carolina (Charlotte): certification that no AI was used, or that all AI-assisted content was human-verified.
- N.D. California (Magistrate Kang): clear labeling of AI-assisted exhibits, with the output still fully at the filing attorney's risk.
None of this replaces ordinary professional duties. Model Rules 1.1 and 3.3 already require lawyers to be competent and candid, and "the AI did it" has not worked as a defense to a Rule 11 sanction.
Keeping the Mediator's Judgment in the Room
Every model mediator standard rests on impartiality. That duty does not pause when a mediator turns to a settlement-analysis tool or an AI-generated case summary. Algorithms trained on human data can absorb human bias, and a tool that quietly favors one profile of party over another undermines the fairness the whole process exists to protect.
Three habits keep AI in its proper, advisory role:
- Bias-test before you rely on it. Run similarly situated hypothetical parties, varying only demographic details, and watch for outcomes that shift for no defensible reason.
- Keep a human in the loop, always. A mediator reviews every AI-generated suggestion line by line and retains full authority to override it.
- Say so, out loud. Parties should be told when AI is assisting the process, what it is and is not doing, and that the final decisions are theirs — not the algorithm's.
Enforcing the Deal Once It's Signed
A mediated settlement is enforced like any other contract: it needs a genuine meeting of the minds, the capacity to agree, and — in many states — a signed writing that says the parties intend to be bound. California's Evidence Code section 1123 is a good example of a statute built for exactly this purpose.
Challenges built on claimed pressure or incapacity rarely succeed absent real fraud, duress, or a documented inability to understand the agreement. Ordinary advocacy, even vigorous advocacy, is not coercion. The practical lesson is procedural: confirm capacity, repeat the key terms aloud before closing, and give every party — especially anyone unrepresented — a real chance to ask questions or consult someone before signing.
Two international instruments matter for cross-border agreements: the 2019 UN Singapore Convention on Mediation and the UNCITRAL Model Law, both built to make mediated settlements enforceable the way arbitral awards already are across borders.
Remote Mediation: Identity, Privacy, and Jurisdiction
Video-on for every participant has become the baseline expectation, not a courtesy — it is how a mediator confirms who is actually in the room and notices an uninvited listener. A secure, encrypted platform, a reminder that recording is off-limits, and locked or waiting-room-controlled access round out the basic security posture.
When parties sit in different states or countries, agree in advance on governing law and, where a party is in the EU, keep GDPR's data-handling expectations in mind.
Jurisdiction at a Glance: Confidentiality
| Jurisdiction / Rule | Protection | Waiver Rule |
|---|---|---|
| California (Evid. Code §§1115–1122) | Excluded by public policy, not a waivable privilege | No implied waiver; requires written agreement of all parties |
| Uniform Mediation Act states | Confidential and privileged | Written consent required; third-party disclosure likely waives |
| Federal ADR (28 U.S.C. §652) | Confidential, limited exceptions | Third-party sharing likely breaches the protection by analogy |
| Other states (varies) | Most protect mediation communications | Disclosure outside the mediation generally treated as waiver |
Where This Is Headed
- Legislatures should clarify, by statute, that mediation confidentiality covers approved AI processing.
- Courts should harmonize AI-disclosure orders around a common baseline: name the tool, certify human verification.
- Mediators and ADR providers should adopt written technology protocols covering identity checks, data handling, and plain-language explanations for parties.
- Wider adoption of the UNCITRAL Model Law and Singapore Convention would strengthen cross-border enforcement of mediated agreements.
This article summarizes general legal and policy developments as of 2026 for informational purposes and does not constitute legal advice. Confidentiality, privilege, and disclosure rules vary by jurisdiction and change quickly in this area — consult a licensed attorney about your specific situation.
