AI in Mediation. the Tool Is Not the Process: Using the IBA Guidelines to Evaluate Risk in Mediation Practice

Artificial intelligence has become, in one way or another, a part of many dispute resolution practices. Counsel use AI to prepare mediation briefs, assess litigation risk, test settlement ranges, or draft suggested terms. Parties use it to understand the process or evaluate options. Some mediators may use it to organize information, draft correspondence, test language, or reflect on process choices.

The discussion about AI in mediation has also become more urgent and comprehensive. Much of it properly focuses on confidentiality, neutrality, party autonomy, disclosure, competence, and human judgment. For working mediators, however, the next step is practical: distinguishing between lower-risk and higher-risk uses of AI and deciding what safeguards are needed in ordinary mediation practice.

My view is that whether mediators are “for” or “against” AI is not a useful question. It is whether a particular use of AI is consistent with the process we have been asked to conduct.

In some parts of a mediator’s work, AI is potentially quite useful. It can summarize long documents, compare position statements, organize chronologies, identify apparent points of agreement and disagreement, suggest agenda structures, generate possible questions, and help reframe inflammatory language. It can also assist with generic role-plays, training simulations, plain-language explanations, and first drafts of routine communications. For busy mediators, this is not insignificant. Preparation is often the invisible work of mediation, and, in complex disputes, the volume of material can be substantial. Anyone who mediates regularly knows that preparation is not just reading the brief. It is trying to understand the best approach to mediation for each particular dispute.

But we must remember that mediation is not simply information processing. Mediators know that some of the most important information in the room is not in the documents. It is in tone, timing, hesitation, anger, fatigue, mistrust, and the particular way a party says, “That is not what happened.” It is in what counsel does not say. It is in the gap between a stated position and an underlying concern. It is in the practical, emotional, relational, and reputational dimensions of conflict.

One of the most important question facing mediators is whether AI can be used without weakening the qualities that make mediation valuable to disputants: confidentiality, neutrality, party self-determination, balanced participation, trust, and judgment.

Start with the IBA Guidelines

The International Bar Association Mediation Committee has issued Guidelines on the use of generative artificial intelligence in mediation. They are helpful because they do see AI in binary terms as either prohibited or harmless. They recognize the potential benefits of AI while grounding its use in mediation values.

The Guidelines are broad. They contemplate AI use by mediators, parties, representatives, and institutions, from administrative support and document synthesis to risk analysis, settlement drafting, mediator selection, and institutional administration.

In many ways the most important contribution of the guidelines to the AI debate is the development of a safeguard framework. The IBA focuses on party autonomy, privacy and confidentiality, neutrality, impartiality, independence, and balanced process. It also includes a sample AI usage statement that may be adapted where disclosure of AI use is required or helpful. In other words, the Guidelines do not ask only, “What can AI do?” They ask, “What does AI use to do to the process?” and this is the question mediators should also be asking.

At the time of writing, I am not aware of any Canadian ADR organization having issued mediation-specific guidelines on the use of generative AI. Existing Canadian materials appear to consist mainly of general professional standards, online mediation guidance, arbitration-focused commentary, educational programming, and broader AI resource lists. That makes the IBA Guidelines a useful reference point, while still requiring Canadian mediators to apply them through Canadian privacy law, mediation agreements, institutional rules, professional standards, and the circumstances of the particular case.

A Practical Risk Matrix

A useful way to start is by separating the proposed uses of AI in mediation. The risk changes depending on whether the tool is being used for generic learning, administrative drafting, confidential document review, live-session transcription, risk modelling, option generation, or settlement drafting. A mediator does not need the same safeguards for every use but does need a well thought out and organized way to decide which safeguards are appropriate. Using a generic role-play in training is one thing. Uploading the parties’ actual mediation briefs is another.

Obviously, not all AI use carries the same risk. The risk depends on the tool, the information entered, the purpose of the use, the level of review, the sophistication of the participants, and whether the use affects the mediation process itself. A simple risk matrix may help mediators think about the issue more clearly. The point is not to create a fixed taxonomy, but to encourage mediators to identify the specific use before deciding on the safeguard.

This is not a set of definitive rules. It is a way of slowing down and that is one of the most important safeguards available to mediators: pause before using the AI tool, identify the risk, and decide what protection is needed. It also helps avoid the common mistake of discussing AI as though all uses are equally risky or equally benign. They are not.

Party Autonomy and Risk Analysis

Mediation is a consensual process. The parties make the decisions. Used carefully, AI can support that decision-making through helping parties and counsel organize information, consider options, and test risk. Used carelessly without thought, it may create the appearance of authority without the substance of judgment.

This warning is particularly important when AI is used for litigation risk analysis, BATNA or WATNA assessment, settlement modelling, concession analysis, or the generation of possible resolution options. A litigation risk analysis generated by AI may look objective. A settlement range may appear data driven. A proposed compromise may seem neutral because it was produced by a machine rather than by a person. A number can change the conversation, even when the number is only as good as the assumptions behind it.

Settlement decisions are rarely based only on legal merits or expected value. They may involve cost, delay, uncertainty, business needs, family dynamics, reputation, enforcement risk, emotional exhaustion, and the desire to move on. In a commercial case, for example, a range generated from legal costs, probability of success, and possible damages may still miss the reason the file is difficult: an ongoing business relationship, reputational concerns, or a party’s need for an acknowledgement that has no obvious monetary value. These factors do not always fit neatly into a model. That does not make modelling a waste of time, but it does mean that mediators should be careful about how AI-generated analysis is used and described.

It should be treated as a prompt for discussion, not as an answer. Some common questions questions are often:

• What assumptions are built into this analysis?
• What facts are uncertain?
• What risks are missing?
• What non-monetary interests are not captured?
• How would the other side see this?
• What would change the assessment?

Mediators must be particularly cautious about presenting AI-generated analysis as though it represents a fair, correct, or objective outcome. AI may help structure a conversation, but it must not decide the destination.

Confidentiality & Privacy Come First

For mediators, confidentiality is central to the service we provide. The convenience of AI creates a real temptation. It is easy to paste a mediation brief, chronology, expert report, email conversation, or draft settlement document into a generative AI tool and ask for a summary. The easier that becomes, the more disciplined mediators must be.

Before using AI with matter-specific information, mediators should ask practical questions:

• Is the information confidential?
• Does it identify the parties directly or indirectly?
• Has it been properly anonymized?
• Could the parties still be identified from the factual context?
• What are the tool’s terms of use?
• Is the information retained or used for training?
• Where is the data stored?
• Is the tool public, proprietary, open-source, or enterprise-grade?
• Do the parties or counsel need to know or consent?

Anonymization is important, but it is not always enough. In a specialized industry, a small community, a unique dispute, or a matter involving public events, parties may be identifiable even after names are removed. Confidentiality analysis, therefore, has to include contextual identifiability, not just whether obvious names and addresses have been deleted.

The safest starting point is conservative: do not place identifiable confidential mediation information into a general-purpose AI tool unless there is a clear, informed, and defensible basis for doing so. That basis may include secure terms of use, appropriate privacy protections, informed consent where required, and a careful judgment that the proposed use is actually necessary.

Your Neutrality Is Not Necessarily Prompt Neutrality

Mediators occupy a different role from advocates. Counsel may use AI to test arguments, strengthen submissions, prepare clients, or draft persuasive language. That is advocacy. Mediators must be more careful. Our credibility depends on the parties’ confidence that we are not adopting one side’s framing as our own.

This creates a risk when AI is used to summarize materials or generate possible options—inputs shape outputs. If one party’s brief is more detailed, more polished, or more forcefully written, the resulting summary may give that party’s version disproportionate weight. If the prompt is framed in a way that assumes one side’s narrative, the output may appear neutral while embedding that assumption.

Prompting in neutral terms helps. So does varying prompts and comparing outputs. But the deeper point is that neutrality cannot be delegated to a tool. A mediator should not ask AI to decide who is right, determine the fairest result, or generate the best settlement. Better uses include asking it to identify issues, summarize competing positions, list areas of agreement and disagreement, suggest questions, or generate a range of possible options for further discussion.

The mediator’s experience and judgment must remain essential. AI may help test whether language sounds neutral, but the mediator must decide whether it is neutral in context. That context includes the history of the dispute, the relationships in the room, the information asymmetry between the parties, and the possible impact of the mediator’s words.

A Balanced Process

To be fair, mediation requires a balanced process. Each party must be given an opportunity to fully participate, express their views, understand the process, and make informed choices. AI can support that goal. It can help with language translation, plain-language summaries, information organization, accessibility, and preparation. It may help a mediator identify unresolved issues or participation barriers.

The same tools may also create an imbalance. One party may have access to sophisticated AI tools and technical support while the other does not. One party may use AI during the mediation without the other party understanding how it is being used. A represented party may rely on AI-assisted analysis in a way that pressures an unrepresented party. A mediator may unknowingly rely on outputs that reflect cultural, linguistic, or social assumptions.

The process question is therefore important: are the participants comfortable with the use of AI, do they understand what it is being used for, and can they raise concerns? In some cases, AI use may need to be disclosed. In others, disclosure may not be legally required but may still be good process.

The IBA’s sample AI usage statement helps us because it encourages users to identify the tools they are using, refer to their terms and conditions, acknowledge their data protection responsibilities, protect confidentiality, and confirm that AI outputs will not be treated as final authority. That should be kept in mind at all times because the issue is not the technology’s novelty. The issue is whether participants can still have confidence in the process.

Developing Your Policy

Mediators do not need a complex AI policy for every file. A practical protocol should begin by identifying the proposed use of AI and deciding whether confidential or identifying information is involved. If it is, the mediator should consider whether the use is necessary, whether the information can be limited or anonymized, what the tool’s privacy terms and data use policies provide, and whether disclosure or consent is required or appropriate.

The protocol should, at a minimum, address the mediation values engaged by the proposed use of AI:

• Does the use preserve neutrality?
• Does it support or undermine party autonomy?
• Could it affect balanced participation?
• Are the outputs being treated as prompts for professional judgment, or as conclusions?
• Is any factual or legal content being independently verified?

These are not simply technical details. They go to the integrity of the mediation process.

A simple clause for a mediation agreement might begin along these lines:

The mediator may use generative artificial intelligence tools for limited administrative, organizational, drafting, or professional support purposes in connection with the mediation. The mediator will take reasonable steps to protect confidentiality, including limiting and anonymizing information entered into such tools where appropriate. The mediator will not rely on AI-produced content as a final authority and will independently review any AI-assisted output before using it in the mediation process. Participants may raise any concerns about the use of AI tools in the mediation.

Please note that this is only a starting point and there are more comprehensive examples available online. Whatever precedent that is used must be adapted to the mediator’s practice, the dispute, applicable law, institutional rules, and the expectations of the parties. It may also be too narrow for some cases and unnecessary for others. Its value is not that it solves every issue. Its value lies in requiring the mediator and participants to think about AI use before a problem arises.

Be Mindful

The risk is not that AI will replace mediators. The more immediate risk is casual use. These tools may help us prepare, draft, train, and organize complex information. In some settings, they may reduce costs and improve access to dispute resolution.

But mediation is not valuable simply because it is efficient. It is valuable because it creates a structured process for difficult conversations, informed decision-making, and negotiated resolution. The process depends on trust, confidentiality, neutrality, timing, judgment, empathy, and the parties’ confidence that they remain responsible for their own decisions.

The practical task is to stop asking whether AI in mediation is good or bad in the abstract. Much more useful questions are:

• What is the proposed use?
• What is the risk?
• What safeguards are needed before the tool becomes part of the process?

That requires more than a general comfort or competence with technology. It requires attention to the specific file, the specific parties, and the specific use being proposed. The tool is not the process. The mediator’s judgment and the parties’ informed choices and consent must remain central.