Artificial Intelligence – “to be or not to be” at mediations?

As a commercial mediator, I often hear lawyers avidly discussing the hot topic of the use of AI in legal practice. How desirable is it for solicitors and barristers to use AI at mediations, and indeed mediators themselves? If AI is used, should or must it be disclosed?

Summary

In the human driven process of mediation, AI has a supporting role. Key drawbacks are issues of confidentiality, hallucinations and sycophancy.

AI can be beneficially used in preparation for a mediation, analysing settlement criteria and ranges, and potentially as a tool to overcome impasse as a last resort.

Currently, the use of AI at mediations is limited, and in my opinion rightly so, given that mediation involves the parties being able to “see the wood for the trees”. The proliferation of information that AI generates can be counterproductive to keeping the big picture in mind. 

Given the increasing prevalence of AI in business it is important that both lawyers and mediators appreciate and keep up to speed with how AI can be used in their professions. Mediators should ask how much AI is being used at a mediation, since there is no duty to disclose this. 

These points are considered in more detail below.

Confidentiality

First things first. One of the cornerstone principles of mediation is confidentiality. The principle overriding risk of putting any documents relating to a mediation into AI software is breach of confidentiality because the generally available open source LLMs use what others have input to train their models. So, what is uploaded into for example a ChatGPT search might be used to answer another person’s search and thereby be disclosed. Only closed/confidential/ private AI systems should be used in relation to a mediation. Even then there is a risk that unauthorised users accessing the lawyers’ or mediator’s AI accounts can become privy to confidential information uploaded there. This warning comes as a large caveat to all that follows.

Hallucinations

There has been a recent series of cases where solicitors’ firms have been admonished by the courts and on occasions referred / self -referred to the SRA for using AI to prepare submissions that have not been checked. This included referring to computer generated “hallucinations” such as a case cited that does not exist, or a rule that had been completely incorrectly quoted. The good news is that as AI technology is becoming more developed, hallucination rates are decreasing. However, the message remains clear: AI results need careful checking before use. 

Sycophancy

An even more concerning aspect of using AI at a mediation is sycophancy. A recent study found that large language AI models agreed with users in relation to interpersonal issues a staggering 49% more than a human would. This is apparently currently a difficult tendency to correct, either by software developers or users.  Parties often appear pretty convinced their case and behaviour are “right” at mediation; a computer reinforcing that belief will only stoke the flames of their conviction, and the dispute further. Mediation requires challenge, balance and an open mind. Asking AI for advice on negotiation, communication styles and how to lead a conversation therefore comes with a very large health warning.

Preparation

When it comes to preparing the mediation bundle, AI might well be useful when preparing chronologies, lists of issues and case summaries. Currently there is no need to disclose the use of AI in the preparation of these documents whether for mediation or trial. It is looking likely that this will continue. The Civil Justice Council Consultation on the use of AI is out for response and currently proposes there be no need to disclose the use of AI for these types of documents although it may become necessary for other court documents. The Civil Mediation Council, in its response to the Consultation, says it would be user friendly if there were consistency of approach as to how AI may be used for trial and mediation. So, watch this space for how guidance and regulation may develop in this area. 

When it comes to the mediator’s preparation, an AI agent could be restricted to a specific bundle of documents. This would allow the mediator to search the material using natural language and identify sources more quickly. The search results would be confined purely to the documents under consideration. AI could also be used to assist the mediator to write questions to ask the parties at the mediation. A mediator should not of course rely solely on the AI generated list – it could be a starting point or an adjunct to a list they have prepared. AI could be used to summarise the bundle. In one survey, AI was 77.2% accurate on summarisation and 94.8% accurate on document Q & A. However, while summaries can accelerate review, they can also reduce nuance. The thing about mediation, is that sometimes the seemingly unimportant facts or evidence on paper prove critical during the mediation session. I personally would never rely on an AI summary, although it might be a useful additional tool to reading the bundle.

Position Statements

An important document for the mediator is the Mediation Statement/ Position Statement, which outlines among other matters the party’s perspective on the dispute, issues to be focussed on at the mediation, and pertinent background matters. It is intended to be a persuasive document and is sometimes (cynically) said to be the only document the client in the opposing party will read with real care and attention. It is important therefore that the Mediation Statement is authentic and is written in the style of the team delivering it. Using AI to generate it is undesirable as it will probably create a tone that does not reflect the author, unless very heavily edited. 

Younger generation lawyers may try to counter this aspect of authenticity by seeking to improve the tone of voice in AI generated material using AI custom skills or agents trained on previous examples of their work. The risk here is if AI becomes embedded too early in the writing style of someone who has recently joined the profession, they may struggle to develop their own natural professional voice. Their writing may also begin to sound generic or worse still AI influenced if the examples they used for the training are themselves based on AI contaminated material.

Open Session at the Mediation

It is common, and usually encouraged by mediators, to have a meeting attended by everyone early in the mediation (the “Open Session”). I have heard third hand of mediators running AI software during the Open Session that prompts questions or interventions the mediator may make during that discussion. I cannot see the desirability of this. Essentially the parties are wanting a discussion facilitated by a human being, and the interposition of AI is likely to detract from the communication skills, insight and empathy that a mediator will bring to the table. Further, consulting a laptop at a time when the mediator should be giving full concentration and good eye contact to the parties and their lawyers is likely to result in less effective communication. Ethically, if a mediator were to use such AI, it would be desirable to disclose this to the parties, although there is no requirement. It is also interesting to note that while AI generated empathic responses that make people feel heard, when people know the response came from a machine, research shows the effect drops significantly.

Private caucuses at the mediation

Frequently at mediations a trainee solicitor or paralegal will take notes of what is said during the private meetings. Will we reach a situation when this role will be redundant and note taking software will be used? The IPOS Mediation Agreement provides that no audio or video recording or transcript will be made of any part of the Mediation. 

The practice of note taking software being used for business meetings in general is becoming increasingly popular. The consent of everyone in the room would be needed for this to happen, not least for privacy reasons. If the voices are recorded and a transcript subsequently made, the consent to the use of an individual’s voice would be needed. This is unlikely to be forthcoming given concerns of possible misuse of that person’s voice should the audio file unwittingly end up in the wrong hands.

Microsoft Teams has built in a Copilot functionality such that when transcription mode is enabled, participants can ask questions during or after the meeting based on the live transcript. While this can be useful for actions, summaries and clarifications, there can be risks of inaccuracies, particularly with names, professional details and legal terminology. Any such transcripts would need to be very carefully checked, especially if to be read by clients.

Negotiations

I have seen negotiations very effectively supported by legal teams who have created spreadsheets that “number crunch” a variety of settlement outcomes, and how different parameters would alter them. Shareholder disputes where values of share are in issue, valuations of land and chattels in probate disputes, and calculations of termination payments for agency contracts are just some of the examples that come to mind. AI can be usefully employed in the creation of such material, and I would argue that its use need not be disclosed, as usually these documents remain for the private use of one party in deciding what offers to make or accept.

Parties and their advisers may go a stage further and use tools that predict the outcome of a dispute and refer to them during negotiations at a mediation. Indeed, we may reach a time when large law firms collate their own specialised data on certain types of disputes and their outcomes at mediation and/ or trial. The CMC refers in their response (referred to above) to BlueJ as one such predictive tool that has been adopted in the UK. They further state (albeit the CJC Consultation does not specifically address the point) that for mediators and parties to be able to “interrogate” these arguments properly, the fact that AI has been used, as well as potentially which AI tool, should be disclosed. I would suggest if the point were to be discussed in depth, the prompts used should also be disclosed, since they will impact the results. It is quite likely, in any case, that the risk tolerance of the party themselves, the view of counsel (if engaged) and the solicitors are not completely aligned with the AI answer, the latter merely being a further piece of the “jigsaw puzzle” for the party to bear in mind when deciding what a satisfactory solution may look like.  

Impasse

What if an impasse is reached at mediation? I am largely a facilitative mediator, although I am evaluative to the extent of the issues I choose to probe and the questions I ask. At one mediation the parties reached an impasse, and before totally giving up asked me where I thought a solution satisfactory to all might lie. At this point I became more evaluative and agreed I would share my view with them all together, but would not give reasons or answer any questions, since to do so might risk breaching confidentiality. I gave my view, the parties returned to their private sessions, and each told me that yes, they could “live with” my suggestion. The mediation settled on that basis. 

A related, novel use of AI applied as a potential tool to break impasse is reported in the US in a mediation concerning a payment on termination of a lease. The experienced mediator Myer Sankary asked ChatGPT what amount the landlord should request the guarantor to pay. The mediator thought the amount was more than the guarantor would be willing to pay. Nevertheless, he asked both parties if they would agree to the amount stated by ChatGPT if there was an impasse. They agreed to that process. Eventually the parties reached an agreement without resorting to using the ChatGPT intervention, settling at a figure very close to the ChatGPT sum. They subsequently asked what the ChatGPT figure was and remained satisfied with their deal. It is thought that the “push” of the case ultimately being determined by AI if the parties failed to do so motivated the parties to continue. It seemed to turn the process into something approaching a mediation with the possibility to morph into an Early Neutral Evaluation. Cautions about this approach are what AI prompts did the mediator use to generate the settlement outcome? The computer driven answer will depend on the input. 

Fundamentally, if parties are coming to “just a mediation”, wanting an outcome at whatever cost, this is a possible approach. But in my experience, most parties are wanting, at least to some extent from their perspective, a “just mediation” and a large part of that is their involvement in the outcome.

The Settlement Agreement

Should lawyers use AI to assist them to draft a settlement agreement? I usually encourage lawyers to draft wording relating to certain issues as they are agreed during a mediation, to reduce how much drafting needs to be done at the end. AI could be used to check that no important clauses from their client’s perspective have been missed. Certainly, great caution should be exercised before reaching for an AI generated agreement. I have, however, heard of unrepresented parties at a mediation using this as an approach in very straightforward cases. A practical approach, but not without its risks. Custom “skills” or agents can be configured to apply a firm’s branding, fonts, formatting and house style without changing the underlying content. Possibly lawyers could be tempted to cut corners here and reach for an AI template if it is late at night when a mediated settlement agreement needs to be produced. Up to date training in drafting for junior lawyers should cover the risks of this practice and emphasise the need for a lawyer to very carefully check and specifically tailor any AI generated template.

Conclusion

The exponential use and development of AI mean its use by lawyers and mediators needs constantly addressing and updating. Some top tier law firms are already positioning themselves as “AI-first” and placing significant trust in AI tools. It is likely that this tendency will cascade through the wider legal sector.

Lawyers representing clients at a mediation may find it helpful to use AI for data management and creating spreadsheets for possible variations of settlement outcomes. 

As to mediators using AI for their own professional practice, I would caution strongly against this without ensuring confidentiality is preserved, and carefully checking and exercising judgment about any results it produces. 

All mediators earn their crust partly by the quality of the questions they ask. I now find myself occasionally asking: “is it you or AI who thinks that?” and “if it was AI, which AI agent and prompts did you use?” The answer to this may be lengthy, as several chatbots may have been used, with hundreds of iterations. Depending on the context and how time at the mediation can best be spent, it may be enough simply to know that AI has been used. If the topic needs to be discussed further, users might resort to an AI summary of the prompts (but thereby casting a further layer of AI on the issue!).

The success of many mediations lies in stepping away from the detail of the dispute, legal arguments, and evidence and facilitating the parties to gain a pragmatic and human perspective on where a settlement may lie. Using AI is potentially a hindrance to that. How AI has impacted on the belief systems, submissions, preparation and negotiation of parties and their advisers is a relatively new strand to be considered by mediators, likely to grow in importance as AI is increasingly adopted.

No AI agent was used to generate this article. Google search was used to check the surveys.

Rebecca Attree is a Commercial and Workplace Solicitor with IPOS Mediation.