AI for Mediators: Summaries and Agreements That Stay Neutral

A mediator can use AI for session notes and agreement drafts without compromising neutrality by keeping three rules: every party agrees in advance, only your own anonymised notes go in, and every output is checked for balance before anyone sees it. The tool formats and tidies. It never characterises a party or adds a term.

What ranks for this topic today is mostly policy statements and think-pieces about AI and dispute resolution. Those help with principles, and they leave a working mediator without a routine. Here is one: a session-note checklist for neutrality, a summary prompt built to stay balanced, rules for drafting an agreement from terms the parties have already reached, and consent wording you can adapt.

Follow me on Instagram@sagnikteaches

Why neutrality is the main risk, ahead of confidentiality

Most advice about AI in sensitive professions starts with data protection, and that matters in mediation too. But the distinctive danger is subtler. A mediator's value rests on being impartial in fact and in appearance. Language models are fluent writers with habits that undermine this: they choose vivid verbs for one speaker and bland ones for another, they give the more articulate party more space, they smooth over disagreement into a tidy narrative, and they drift towards recommending outcomes. A summary that reads "Ms A raised concerns, whereas Mr B insisted" has already taken a side, and no one wrote that on purpose.

Connect on LinkedInSagnik Bhattacharya

The widely used Model Standards of Conduct for Mediators, revised in 2005 and approved that year by the AAA, the ABA and the ACR, set out nine standards. Six bear directly on AI use: self-determination, impartiality, conflicts of interest, competence, confidentiality and quality of the process. Standard II asks a mediator to conduct a mediation impartially and to "avoid conduct that gives the appearance of partiality", which is the test a lopsided summary fails. The text also says the standards do not have the force of law unless a court or other regulatory authority adopts them, and your own rules may come from a court roster, a professional body or a statute; those control. Use the headings as a mental checklist whichever code you follow: does this use of AI protect the parties' right to decide for themselves, keep me impartial in fact and appearance, preserve confidentiality, and stay within my competence?

Subscribe on YouTube@sagnikteaches

That last one deserves a plain statement. Competence includes understanding the tool well enough to know when it is wrong. If you cannot explain to a party what the tool did with their words, do not use it on their case.

Consent comes first, and it comes from every party

Mediation depends on trust in the process. If one party learns after the fact that their words went through a tool they did not know about, the damage is hard to repair, even if nothing leaked. So disclose at the start, in the agreement to mediate, and get each party's agreement individually. Here is wording you can adapt; have a lawyer or your programme's guidance check it against the rules that apply to you.

Use of AI tools. The mediator may use an AI writing tool to format the mediator's own
notes, to prepare neutral summaries of points the parties have stated, and to tidy
the wording of terms the parties have agreed. The mediator will not enter party names,
recordings, transcripts, documents supplied in confidence, or anything shared in a
private session. The mediator reviews all output for accuracy and balance. Any party may
decline the use of AI tools at any time without any effect on the mediation.
Each party initials below to confirm they have read and accept this clause.

Four decisions sit inside that clause. It limits the tool to your notes and agreed terms, so parties know exactly what is touched. It excludes private sessions (caucus material is the most sensitive thing you hold). It states review, because the responsibility is yours. And it gives any party a veto, which costs you little and protects the perception of fairness a great deal. If one party declines, you do that case by hand. That is a minor inconvenience compared with the alternative.

For the contract-wording side in other settings, how to mention AI use in client contracts and proposals covers how professionals phrase this without alarming clients.

What goes in the tool, and what never does

Draw the line before you open a window. On the safe side is your own typed notes with names replaced by "Party A" and "Party B", and the final text of terms the parties have agreed. On the other side are recordings, transcripts, documents a party gave you in confidence, and anything from a private session. For confidentiality settings, on free consumer accounts content may be used to improve models unless you switch that off, while ChatGPT Business and Claude Team do not train on business content by default, according to both pricing pages in November 2026. List prices are $20 per seat a month billed annually or $25 monthly with a two-seat minimum. Confirm on the vendor's page, and read what to check in an AI tool's privacy policy and terms for the questions that matter.

The parallel in therapy is instructive. Can therapists use ChatGPT for session notes? reaches a similar line: general chatbots and identifiable session content do not mix, and anonymised, self-written notes are the workable zone. Mediation adds a twist, because you have two or more parties who each hold a confidentiality interest.

The session-note checklist for neutrality

This is the core of the routine. After every AI-assisted summary, run these ten checks before the document leaves your desk. Each takes seconds. Together they catch the usual bias in the output. The worked answers beside each are for an illustrative workplace dispute between two colleagues, here called Party A and Party B.

Checks 1 to 3 are mechanical, and you can paste both halves into a word counter. Checks 4 to 6 need a careful read. Checks 7 and 8 are about accuracy and need your notes beside the draft. Check 10 is the one that matters most, and it is a human act: imagine each party reading it aloud in the next session.

A summary prompt built to stay balanced

The prompt does much of the work, so write it as constraints. Use your own notes, anonymised, in a fixed structure: each party's stated points as bullets, in your words.

You are formatting a mediator's own notes into a neutral session summary. Use only the notes below.
Rules:
1. Give Party A and Party B equal space; each section within 15% of the other's word count.
2. Use only the verbs "said" or "described". No adjectives about any person.
3. Do not explain motives or causes. Do not recommend or hint at outcomes.
4. Where both parties raised the same topic, group under that topic and list A then B;
   alternate which party comes first across topics.
5. Do not add any fact, date or figure that is not in the notes.
6. End with a list headed "Points still open" taken only from the notes.
Notes: [anonymised notes]

Illustrative output for a workplace dispute: under the topic "Handover of the client account", Party A "said the handover note arrived two days after the deadline and described the delay as affecting a client call." Party B "said the note was started on time and described a system outage on the final day." Under "Team meetings", the order reverses: Party B first. Points still open: "Whether the handover template should change. Who attends Monday meetings."

What you would fix. The first draft usually passes rule 2 but slips on rule 1, with one party's section 30 words longer because their notes were longer. Fix it by trimming the longer one, never by padding the shorter. Next, look for words that carry a judgement even though they are not adjectives: "delay" is arguably neutral, "failed to deliver" is not. Replace anything loaded with the party's own phrase from your notes. Finally, if the model wrote "Both parties appear keen to resolve this", delete it. That is the model's reading of mood, not anything the parties said.

A quick sum for check 1

Say Party A's section is 142 words and Party B's is 131. The gap is 11 words, 11 divided by 142 is about 7.7 per cent. That is inside 15 per cent, so it passes. If B's section had been 105 words, the gap would be 37, about 26 per cent, and you would look at what the longer section contains that the shorter one lacks. Sometimes the answer is that A really did raise more points, in which case a note says so. Equal space is a prompt to look, not an absolute rule.

Drafting an agreement from terms the parties reached

The agreement stage is the riskiest for neutrality. Whoever holds the pen shapes the agreement, and a model that "improves" wording can shift obligations. The rule is narrow: AI may format and tidy, and it may not originate. It sees only the terms the parties actually agreed, as you recorded them, and it adds nothing.

Below are terms agreed by the parties, recorded by the mediator. Format them into a numbered
memorandum of understanding. Rules: use the mediator's wording wherever possible; change
only grammar and layout; do not add terms, definitions, deadlines, remedies or
boilerplate; do not resolve any ambiguity: if a term is unclear, list it under
"Unclear: parties to confirm". Keep "Party A" and "Party B".
Agreed terms: [mediator's notes of the terms]

Illustrative output: five numbered terms formatted cleanly, such as "1. Party A will send the weekly client summary by Thursday 12:00." and a short "Unclear: parties to confirm" list containing "Term 4 says 'promptly': parties to confirm a timeframe." That unclear list is the most valuable part. Models are good at spotting vagueness, and a settlement that says "promptly" invites the next dispute.

Then do three checks of your own. Compare every term against your notes line by line. Look for new verbs that create stronger duties ("shall" where the parties said "will try to"). And check obligations are even-handed in form: if A's duties are written as "shall" and B's as "will endeavour to", that is a drafting asymmetry with legal effect. Practice varies on legal review. Many mediators tell parties to take independent advice on the final agreement, and the parties' lawyers often finalise the text. If your role includes drafting, follow your programme's rules, and see how family lawyers handle financial disclosure with AI for the legal side of document handling.

Three other jobs where AI is safe, and one where it is not

1. Agenda and ground rules. Draft a generic agenda and a one-page ground-rules sheet. No party data is involved. Edit for your style.
2. Follow-up letters. A short letter confirming the next session date, what each party agreed to bring, and the open points. Use the same balanced prompt with "Points still open" only.
3. Plain-language rewrites. Where a party struggles with long wording, ask for a plainer version of your own text, then compare it with the original so meaning has not shifted.
4. Not safe: assessing credibility or suggesting settlements. Asking the tool who is more reasonable, or what a fair split would be, hands judgement to a system with no accountability. A mediator facilitates; the parties decide. Do not use AI for either of these.

If you coach or supervise mediators, the same list works as a training handout. For a shorter version of the same question in counselling, safe AI uses for counsellors that never touch client notes sets out the uses that sidestep records entirely.

Before and after: rewriting a loaded summary line by line

The checklist is easier to learn from a failing example. Here is an illustrative first draft of one topic from a neighbour dispute over a shared fence, followed by what the checks flag and the repaired version.

The draft: "Party A raised a reasonable concern that the fence had been left in poor repair for months. Party B insisted the damage was caused by a storm and was defensive about being asked to pay. Both parties seem keen to settle." Run the checks. Check 4 flags "reasonable" and "defensive". Check 2 flags "insisted", which carries more force than "raised". Check 5 flags the storm as a cause the summary states as fact, when it is only B's account. Check 6 flags nothing, but "seem keen to settle" breaks check 8 because the notes record no such statement.

The repaired version: "Party A said the fence has been in poor repair for several months and described the gaps now visible. Party B said the damage happened during a storm and described the cost of repair as a concern." Same length for each, same verbs, no judgement, no invented mood. The repaired lines took about two minutes by hand, which is why the checks pay for themselves: they are short, and every one is a specific edit, not a vague instruction to "be neutral".

Separate meetings and caucus material

Many mediations include private sessions where each party speaks to the mediator alone. Treat everything from those meetings as outside any AI tool, including your own notes of them, unless the party has expressly agreed otherwise in writing for that purpose. The Model Standards are explicit here: Standard V.B says a mediator who meets anyone in private session shall not convey information from it, directly or indirectly, to any other person without the consent of the person who disclosed it. The reason is structural: private-session content is what you promised not to share, and a tool that holds it creates a route, however unlikely, for it to surface somewhere it should not. If you need help structuring private-session notes, use a blank template you made yourself and fill it by hand.

The same caution applies to anything a party hands you "off the record", such as a draft figure they might accept. Keep that in your head or on paper. A neutral summary for the joint session should contain only what each party said or agreed could be shared in joint session, and your notes should mark which is which so a stray private point never lands in the tool by accident.

A walk-through: a commercial lease dispute over two sessions

Take an illustrative dispute between a shop owner (Party A) and a landlord's representative (Party B) over repairs and rent arrears, mediated across two sessions of two and a half hours. After session one the mediator spends 20 minutes typing notes with Party A and Party B labels, 14 bullets each, and then runs the balanced prompt. The draft comes back at 410 words. Checks 1 to 3 flag a gap: A has 230 words, B 160. Reading the notes, the mediator sees A had five more points on repairs. The mediator keeps the content, adds a line under "Points still open" saying repairs raised the most items, and sends the summary to both parties together, not separately, so neither has an informational edge.

Between sessions, the parties each give small movements by email. The mediator logs them in the notes, not in a chatbot. In session two the parties reach four terms. The mediator types them, runs the agreement prompt, and receives a clean memorandum with one entry in "Unclear: parties to confirm": the repair deadline of "end of the quarter". The parties confirm a date, the mediator updates the notes by hand, and the final is reviewed by each party's adviser. In total, AI shortened note formatting from about 40 minutes to about 20 across the two sessions (illustrative) and surfaced the one ambiguity that would have caused a problem.

How a neutrality failure appears, and how you would notice

Real failures are quiet. A summary that always puts the same party first can look natural, because the party who speaks first often is the one who initiated the mediation. Over several cases, if the same side tends to come first, tidy that up. A second failure is tone transfer: if your notes about one party are warmer, perhaps because they were more courteous in the room, the model will reproduce that warmth, and the other party will feel it. Strip emotional commentary out of the notes you give the tool, or keep it out of your notes altogether.

A third is the "helpful" completion. The model adds a sentence such as "Both parties agreed to review in a month" when no one said it. Check 8 catches it. A fourth is the over-neat resolution: the model joins two separate points into a compromise-sounding line. If a sentence seems to settle something the parties have not settled, it came from the model.

Keep a short log of every correction you make, with the check number. After ten cases you will see which checks catch the most, and you can add the top two to the prompt as explicit rules.

Writing it down: a one-page AI policy for your practice

Parties, referrers and programmes increasingly ask how mediators use AI. A short policy answers them and keeps you consistent. It should name the tools and their settings, the jobs allowed, the jobs excluded, the consent step, the neutrality checks, the log, and the person responsible (you). Review it twice a year or when a vendor changes its terms. Share the policy with referrers if they ask, and share the consent clause with every party before the first session.

For a view of how the same issues land across other client-confidential professions, AI note takers for executive coaches works through a related confidentiality problem. If you would rather map your own policy with someone, that is what my AI implementation consultation covers, through the contact page, and the AI guide lists the rest of the series.

Questions mediators ask about AI and neutrality

Do all parties need to consent before I use AI on a mediation?

Treat it as yes. Neutrality means no party should feel the process was shaped by something they did not agree to. Put the AI use in the agreement to mediate, name the jobs, and let any party opt out. Check any rules from your programme, court or professional body too.

Can AI draft the settlement agreement?

It can format and tidy terms the parties have already reached, using only the wording you supply. It must not add terms, fill gaps or suggest fairer splits. Parties are usually told to take independent legal advice before signing, and a lawyer or the parties' advisers should review the final text.

Is a neutral summary prompt enough to stop bias?

No. The prompt helps, but models still pick adjectives, order points unevenly and give one side more space. Run the balance checks in this tutorial on every summary, and compare word counts and verbs for each party before anything is shared.

Should I record sessions and let AI transcribe them?

Only if every party agrees in advance and your confidentiality terms cover the recording and the tool. Many mediators avoid recording because it can change how openly people speak. Handwritten or typed notes in your own words are the safer default.

Further reads

• AI for Divorce Lawyers: Turn Intake Notes Into a Case Chronology — The lawyer's side of intake and chronologies.
• AI Chatbot for HR: A Handbook Q&A Service HR Consultants Can Sell — Another advisory practice weighing client-facing AI.
• How to Stop AI Tools Training on Your Business Data — Switch off training on the account you use.
• Is It Safe to Put Customer Data Into ChatGPT? — The basic data rules behind every consent clause.
• How to Classify Business Data Before Using AI Tools — Sort party material by sensitivity before any tool.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Model Standards of Conduct for Mediators (2005 revision approved by the AAA, ABA and ACR), read from an archived copy of the ABA's PDF; ChatGPT Business and Claude Team pricing pages, checked November 2026.