Mediators have heard some version of it for years: “My brother-in-law is a lawyer, and he says this case is worth at least $1 million.”
The brother-in-law may practice in an entirely different field. He may know only one side of the story. He may never have seen the pleadings, the controlling law, or the damaging email sitting in the other side’s exhibit binder. None of that necessarily matters. To the party, what matters is that someone who sounds knowledgeable has confirmed what the party already hoped was true.
Today, ChatGPT is the new brother-in-law. Except ChatGPT is available 24 hours a day. It will answer unlimited follow-up questions. It speaks with confidence. And it can produce pages of polished analysis explaining why the user’s position seemingly makes perfect sense.
That presents mediators with a new version of a very old problem.
Artificial intelligence can help parties and counsel prepare for negotiation. Emerging research suggests it can improve preparation, generate options, identify interests, and assist negotiators in considering alternatives. But research also suggests that AI advice can foster overreliance, create confidence without corresponding accuracy, and generate seemingly precise recommendations that depend heavily on how the question was asked.
The distinction may be this: AI can help mediation when it expands a party’s thinking. It may hurt mediation when it simply confirms what the party already believes.
A Better-Prepared Brother-in-Law
One of AI’s most promising uses is negotiation preparation.
In a 2025 Negotiation Journal report on preliminary classroom experiments at MIT, researchers used generative-AI coaching and “backtable” bots in multiparty negotiation courses. The research did not include a control group capable of isolating the technology’s effect, but students’ assessments were notable: 82% reported feeling better prepared, 77% reported improvement in identifying other parties’ interests, and 59% found it easier to express their own positions or priorities.[1]
AI may also influence how negotiations unfold. A 2025 study in the Journal of Purchasing and Supply Management conducted three experiments, including one involving professional negotiators, in which participants negotiated against a customized ChatGPT-based buyer. A competitively prompted AI obtained better economic terms and sometimes completed negotiations faster, while a collaboratively prompted AI produced greater trust, satisfaction, and desire for future interaction.[2]
Although that study examined AI as a negotiating counterpart rather than as counsel’s adviser, its lesson is relevant: AI does not bring a single negotiating style to the table. How it is instructed matters. Counsel who ask AI for “the strongest possible counteroffer” may receive very different assistance than counsel who ask it to identify both parties’ interests and develop proposals capable of satisfying them.
For mediation, those capabilities present real opportunities. Counsel can use AI to anticipate counterarguments, identify interests beneath stated positions, develop settlement structures, and test assumptions. Clients can use it to organize concerns and clarify priorities.
In that role, ChatGPT is a useful brother-in-law. Instead of merely telling the party what the case is worth, it can help the party arrive better prepared to negotiate it.
When Advice Becomes Validation
The problem begins when AI moves from helping a party think to telling the party what to think.
A 2024 Nature Human Behaviour meta-analysis examined 106 experiments involving 370 effect sizes on human-AI collaboration. Human-AI combinations generally performed better than humans alone, but worse than whichever—human or AI—performed best alone. Particularly relevant to negotiation, the researchers found performance losses on decision-making tasks but greater gains on content-creation tasks.[3]
That distinction maps neatly onto mediation. “Give me ten different ways to structure a settlement” is a creative task. “What is this case worth?” is a judgment. The first may open the negotiation. The second may close the negotiator’s mind.
Imagine plaintiff’s counsel believes a case is worth $1.5 million and asks an AI system to evaluate it using the facts counsel selects. The system returns a range of $1.2 million to $1.6 million.
Defense counsel does the same from the defense perspective and receives $250,000 to $400,000.
The actual bargaining zone has not changed. But each lawyer may now believe that a sophisticated and apparently neutral source has independently validated the lawyer’s position.
The old brother-in-law said, “You have a great case.” The new one provides a six-page memorandum explaining why.
Confidence Without Accuracy
Behavioral research gives mediators reason to take that problem seriously.
In an incentivized experiment published in Computers in Human Behavior, researchers found that participants over-relied on AI advice, including when it conflicted with contextual information and their own assessments.[4]
A separate 2022 preregistered Scientific Reports experiment produced an equally interesting result. Participants receiving algorithmic advice changed their answers 13% more frequently and reported greater confidence in their final answers, yet were 13% less likely to reach the correct solution.[5] The task was not legal negotiation, and the finding should not be stretched beyond what was studied. But the dynamic is relevant: advice can increase certainty without increasing accuracy.
Mediators already contend with overconfidence, selective interpretation of facts, and unrealistic assessments of risk. ChatGPT does not have to create a bad position to interfere with settlement. It merely has to make an existing bad position feel more authoritative.
The Machine-Generated Anchor
AI-generated numbers pose another problem: they look precise.
A 2025 PLOS ONE study asked four versions of ChatGPT to recommend opening offers in salary negotiations. Researchers generated 7,600 unique prompts and submitted each 13 times to each of four model versions—98,800 submissions per model version and 395,200 overall—and varied factors including which side of the negotiation was asking. Recommended offers changed significantly depending on the model version, whether the prompt came from the employee or employer perspective, and other information supplied in the prompt.[6]
Salary negotiations are not litigation valuations. But the study demonstrates an important point: an AI-generated number can appear objective even though it depends heavily upon inputs, framing, and model behavior. Once ChatGPT tells a client that a case is worth $1.4 million, however, $1.4 million can become an anchor.
The mediator may then be working not simply against the client’s expectation, but against an expectation that has acquired the perceived imprimatur of artificial intelligence.
The Client Has Been Talking to ChatGPT
This may be particularly challenging when the AI user is the client.
A 2024 Scientific Reports paper reporting two studies with a combined 1,925 participants found that ChatGPT advice influenced decision-making across several settings and, in relevant circumstances, affected decisions similarly to expert advice.[7]
A client may therefore arrive at mediation after spending hours asking:
“Do I have a strong case?”
“What are cases like mine worth?”
“Why is their offer too low?”
“What arguments support a seven-figure recovery?”
Unlike the traditional brother-in-law, ChatGPT never gets tired of the conversation. Depending on how the questions are framed, it can repeatedly generate increasingly elaborate explanations consistent with the user’s assumptions.
Mediators may increasingly have to reality-test not only the lawyer’s evaluation and the client’s expectations, but also the invisible adviser that helped create them.
How Do You Mediate ChatGPT?
The answer is not to argue with the machine. It is to mediate the effect the machine has had on the human being sitting in the room.
First, find out whether ChatGPT is in the room. If a demand, offer, case valuation, or probability assessment seems unusually fixed, ask how it was developed and whether AI played a role. Hopefully, the party will be truthful. The point is not to criticize AI use. It is to understand the origin of the position.
Second, question the inputs rather than attack the output. What was the AI asked? What documents did it receive? Did it see the opposing evidence? Was it given the correct jurisdiction and law? Did the prompt treat disputed facts as established? A party may be impressed by the answer until the mediator begins examining the question.
Third, turn the brother-in-law into a devil’s advocate. If a party used AI to support its position, suggest asking the same system to make the strongest argument for the other side, identify the weaknesses in its own case, state what facts could materially change its valuation, and explain why its preferred settlement number might be wrong. That converts AI from a confirmation tool into a tool for expanding thought.
Finally, deconstruct the machine-generated anchor. Do not simply debate whether $1.4 million is right. Ask what assumptions produced it. Change those assumptions individually. Replace the single number with ranges, contingencies, probabilities, and missing facts. The technique is familiar reality testing. Only the source of the confidence is new.
A New Version of an Old Problem
AI is not eliminating the human dynamics of mediation. It is becoming part of them. Used well, ChatGPT can make lawyers and clients better prepared, expose interests, generate creative options, and expand the settlement conversation. Used poorly, it can become a sophisticated confirmation machine—giving an existing belief a level of polish and apparent authority the traditional brother-in-law could never provide.
Mediators do not need to become AI experts. They need to recognize when AI has become part of a party’s decision-making process and apply familiar mediation skills to an unfamiliar source of conviction.
The brother-in-law used to be waiting at home for a telephone call. Now he is sitting in everyone’s pocket.
And before the mediation ends, the most useful question may be the one neither the lawyer nor the client thought to ask him: “ChatGPT, what might I be wrong about?”
[1] Samuel Dinnar, Lawrence Susskind, Leroy Sibanda & Ololade Olaleye, Negotiation Backtable Bots: Using GenAI to Improve Multiparty Negotiation Instruction, 41 Negotiation Journal 19–65 (2025). DOI: 10.1162/ngtn_a_00015.
[2] Silke Herold, Jonas Heller, Frank Rozemeijer & Dominik Mahr, Brave New Procurement Deals: An Experimental Study of How Generative Artificial Intelligence Reshapes Buyer–Supplier Negotiations, 31 Journal of Purchasing and Supply Management 101012 (2025). DOI: 10.1016/j.pursup.2025.101012.
[3] Michelle Vaccaro, Abdullah Almaatouq & Thomas Malone, When Combinations of Humans and AI Are Useful: A Systematic Review and Meta-Analysis, 8 Nature Human Behaviour 2293–2303 (2024). DOI: 10.1038/s41562-024-02024-1.
[4] Artur Klingbeil, Cassandra Grützner & Philipp Schreck, Trust and Reliance on AI—An Experimental Study on the Extent and Costs of Overreliance on AI, 160 Computers in Human Behavior 108352 (2024). DOI: 10.1016/j.chb.2024.108352.
[5] Eric Bogert, Nina Lauharatanahirun & Aaron Schecter, Human Preferences Toward Algorithmic Advice in a Word Association Task, 12 Scientific Reports 14501 (2022). DOI: 10.1038/s41598-022-18638-2.
[6] R. Stuart Geiger, Flynn O’Sullivan, Elsie Wang & Jonathan Lo, Asking an AI for Salary Negotiation Advice Is a Matter of Concern: Controlled Experimental Perturbation of ChatGPT for Protected and Non-Protected Group Discrimination on a Contextual Task with No Clear Ground Truth Answers, 20(2) PLOS ONE e0318500 (2025). DOI: 10.1371/journal.pone.0318500.
[7] Shinnosuke Ikeda, Inconsistent Advice by ChatGPT Influences Decision Making in Various Areas, 14 Scientific Reports 15876 (2024). DOI: 10.1038/s41598-024-66821-4.