Give the Devil the Floor

There is a moment that happens in nearly every mediation session. A party drops their position and starts telling the truth.

Maybe it is a business partner who finally admits the lawsuit is not really about the contract. It is about feeling betrayed by someone he considered a friend. Maybe it is a divorcing spouse who stops citing custody statutes and says, quietly, that she is terrified of being alone. The room changes. Something that was stuck becomes movable.

No algorithm caught that. A human being, paying close attention, did.

I have been an early adopter of technology for most of my career. When AI tools started appearing in professional conversations, my instinct was not to resist them. It was to figure out what they were actually good for.

What I found was more useful than I expected, and more limited.

AI is good at the work that surrounds a session. It can read a stack of intake submissions, identify the core issues, and surface the actual points of disagreement faster than most practitioners can under a tight schedule. That means walking in with more mental bandwidth available for what actually requires attention. It is also useful for drafting. Once parties reach agreement, turning those terms into clean settlement language is largely mechanical, and a solid first draft beats starting from nothing.

One of the best uses I have found is asking AI to play devil’s advocate against my own preparation.

Most people bring AI into their process looking for confirmation. They want their argument organized, their talking points sharpened. What they are actually doing is asking a very capable tool to agree with them more efficiently. The better move is to hand it the other side.

Before a session, I will sometimes type out a question I am planning to ask, or a reframe I am considering, and ask the AI how a resistant party might receive it. Not how a reasonable person would receive it. How someone who is defensive, who has been living inside their version of events for months, who doesn’t yet trust the process, might hear it.

What comes back is usually instructive. A question I thought was neutral turns out to carry an implicit assumption. A reframe I thought was clarifying reads as dismissive to someone who feels wronged. I adjust before I am in the room, which is the only time adjusting is still cost-free.

You are not looking for the objection a calm person would raise. You are looking for the objection someone makes when they are not feeling heard. Those are different objections, and AI is surprisingly good at surfacing them.

What it can’t do is tell you how the other party will feel when you say it out loud, in a room, after an hour of difficult conversation. It can help you prepare the language. It can’t read the moment. That part is still yours.

A client once asked for $225,000 in a situation where the combined assets were under $75,000. What they actually wanted was $10,000 and an acknowledgment that their ex had misspent their joint money. The inflated number was not greed. It was someone who felt wronged and didn’t yet have precise enough language for the injury. No model is trained to hear that.

That gap doesn’t live in documents. It lives in the room, in body language, in what someone doesn’t say, in the quality of a silence after a difficult question. It can’t be uploaded. It can’t be summarized. It requires another human being willing to stay steady in the presence of someone else’s distress, calm enough to notice what is actually happening, and skilled enough to ask the one question that moves things forward.

There is a practical issue that doesn’t get enough attention.

When a mediator uploads session materials to an AI platform, those materials don’t stay in the room. Depending on the platform and its terms of service, they may be used to train future models, stored on servers outside the mediator’s control, or accessible to third parties no one agreed to involve. Most parties have no idea this is happening. They are told the session is confidential. They are not told where that confidentiality ends.

The risk is not that the platform publishes someone’s name. Details accumulate. A business dispute involving a specific dollar amount, a particular industry, and a timeline that anyone in that field would recognize. A family matter with a custody arrangement that narrows the pool of possible parties to a handful. Individually, none of those details identify anyone. Together, they can.

The minimum standard should be anonymization before anything goes into an AI tool. Names replaced with placeholders. Identifying details generalized. Anything that would allow a reader with some context to reconstruct who is in the room should come out before the document leaves the mediator’s hands.

This is not a reason to avoid AI. It is a reason to treat it the way you would treat any other disclosure. What happens here stays here. That obligation doesn’t end when you open your laptop.

The mediators who will use AI well are the ones who are clearest about what it can’t do. Not to protect professional turf. Because they understand what the work actually requires.

AI can give you the strongest version of the argument you will face. It can’t tell you how the other party will feel when they make it. It can help you prepare for the session. It can’t read the room once you are inside it. It can’t notice that what is being asked is not what is needed, and hold that observation until the moment is right to surface it.

That is the work. The tools that help me do it better are welcome. The ones that would replace that moment are solving the wrong problem.

#Mediation #ConflictResolution #DevilsAdvocate #ADR #DisputeResolution #AIandMediation #MediationMonday

Steven Corn is a certified mediator practicing in Maryland and elsewhere focusing on cases that involve divorce, family, civil and music industry issues.