LEGAL VALIDITY AND ETHICAL CONCERNS OF AI IN MEDIATION

How to cite this journal: Author, Date of the post, WMO Conflict Insight, Title of the post, ISSN: 2628 6998, https://worldmediation.org/journal/

ABSTRACT

The rise of artificial intelligence in alternative dispute resolution has sparked debate over its legal and ethical implications. Proponents note that such tools can improve efficiency and consistency, but experts warn that these applications must still align with core values of the field. Industry surveys have emphasised that any use of artificial intelligence in dispute resolution must uphold fundamental principles such as fairness, transparency, and due process (Burn, Morel de Westgaver and Clark 2023). UNCITRAL’s technical notes explicitly list due process, fairness, accountability, and transparency as cardinal principles for online dispute resolution (UNCITRAL 2020). This paper examines whether decisions generated by such systems can be legally binding, particularly in international mediation, and whether they can satisfy due process norms. It also explores how algorithmic bias and the absence of human nuance threaten the legitimacy of outcomes reached in this way. The analysis shows that, while no jurisdiction currently bars settlements assisted in this manner, enforceability depends on traditional rules of consent and contract. At the same time, ensuring impartiality, oversight, and the mitigation of bias is critical to preserving the integrity of mediated agreements.

KEYWORDS *

Artificial intelligence, mediation, arbitration, due process, party autonomy, enforceability, Singapore Convention, algorithmic bias, transparency, human oversight, confidentiality, legitimacy

INTRODUCTION

Mediation is a consensual, often cross-border or commercial method of resolving disputes outside the courts. Globally, mediated settlement agreements, once signed by the parties, are enforceable under instruments such as the UNCITRAL Model Law on International Commercial Mediation and the Singapore Convention on Mediation of 2019. Recently, institutions in this field and law reform bodies have begun to explore the role of artificial intelligence. Major arbitration centres have issued guiding principles or formed working groups on the subject (Eftekhar 2025). An industry report highlights that while such tools offer substantial benefits, including increased efficiency and enhanced access to justice, they simultaneously raise serious concerns, noting that they also pose significant legal and technical issues and that their adoption requires careful consideration to ensure compatibility with the principles and values of arbitration, such as fairness, impartiality, transparency, and party autonomy (Burn, Morel de Westgaver and Clark 2023). Likewise, UNCITRAL’s technical notes emphasise that any online dispute resolution system must embody impartiality, due process, fairness, accountability and transparency (UNCITRAL 2020). Hence, while these systems may streamline mediation, for example by analysing data or drafting settlement terms, experts uniformly stress that outcomes must preserve the parties’ procedural rights and ethical norms.

MAIN CORPUS

Legal validity of machine-generated decisions in mediation

The legal effect of an outcome reached with such assistance depends on established law. In international arbitration, parties typically delegate decision-making to a tribunal whose award is binding under conventions such as the New York Convention. However, many arbitration laws and institutional rules explicitly or implicitly require natural persons as arbitrators (Norton Rose Fulbright 2024). Although the principal treaties do not expressly forbid the practice, many arbitration laws and institutional rules require arbitrators to be natural persons or to possess qualities that presuppose this. By contrast, no major treaty currently addresses machine-generated arbitral or mediated decisions. In practice, if an arbitrator relies on such tools beyond agreed parameters without disclosure, courts or tribunals could later set aside the award for procedural irregularity (JAMS ADR 2024). One commentator warns that unauthorised use of this kind could jeopardise the enforceability of an award under due process principles (JAMS ADR 2024). The European Union’s Artificial Intelligence Act classifies use in judicial decision-making, and in dispute resolution outcomes producing legal effects, as high risk, imposing strict transparency and oversight requirements.

In mediation the situation is different. Mediation outcomes are not awards but voluntary settlement agreements. Those agreements become legally binding as contracts once signed by the parties, and may be enforceable under the Singapore Convention if international, irrespective of whether a human or a machine assisted in reaching them. A system could propose terms, but they bind the parties only if the parties agree and sign. There is no known law expressly forbidding the generation of a settlement proposal in this way. However, because mediation is predicated on party autonomy and consent, any such system must operate with clear party authorisation. To date no country has a specific statute on the subject; enforcement would rely on general contract and procedural law.

The absence of a legal prohibition is not the same as positive endorsement. In my view this legal vacuum calls for proactive regulation. Jurisdictions could consider inserting specific provisions in mediation statutes clarifying the permissible scope of such involvement, especially in sensitive or cross-cultural disputes where trust and human intuition play a crucial role. Recommendations generated in this manner are not binding unless the parties formally accept them, and to be safe the parties should explicitly consent to any such involvement in order to ensure that a mediated agreement later holds up as valid (Norton Rose Fulbright 2024; JAMS ADR 2024).

Key points on legal status

Arbitration and mediation distinguished. Arbitral awards are legally binding, subject to review, while mediated settlements are binding only as contracts. Use as an arbitrator raises questions of consent and statutory authority (Norton Rose Fulbright 2024). Use in mediation must ultimately result in a signed agreement in order to be enforceable.

Party autonomy. Under the UNCITRAL framework and arbitration law, the parties control their process. If they agree to such assistance, whether by clause or by procedural order, its output can be incorporated. Without such consent, reliance on it could compromise due process (JAMS ADR 2024).

Regulatory regime. There is currently no dedicated law on artificial intelligence in dispute resolution. Regulators and courts would apply existing rules on arbitral tribunals, mediator ethics, and contract formation. The classification of dispute resolution decisions as high risk under European legislation indicates that future regulation may require strict oversight of any outcome reached in this way (European Commission 2021; Zilberman 2023).

Due process, impartiality and fairness

Even where such tools are legally permitted, they must respect fundamental principles. Mediation, especially in international settings, demands fairness and neutrality. UNCITRAL’s technical notes advise that online dispute resolution must be impartial and compliant with due process (UNCITRAL 2020). In practice this means that each party should have a meaningful opportunity to present its case and to understand how conclusions are reached. Human mediators rely on face-to-face dialogue, empathy, and adaptive judgement. By contrast, such systems currently lack empathy and situational awareness. This gap becomes especially problematic in mediation, where emotional sensitivity often guides resolution. One observer notes that a leading mediator’s particular strength is making parties feel heard, and that mediators build trust and credibility through fairness and emotional intelligence. On this view, objectivity and fairness play a critical role, but emotional intelligence cannot be underestimated, and parties tend to accept a solution only if they trust the mediator (Cao, Cheung and Li 2023). As another author asks, has anyone yet encountered a computer showing empathy or emotion (Fasolo 2025)?

In my view these qualities are not merely helpful; they are indispensable. The essence of mediation lies in trust building and emotional validation, which no such system can replicate. It lacks consciousness, cannot read non-verbal cues, and is incapable of contextual emotional responsiveness. Delegating mediation to machines, even partially, reduces a fundamentally human process to a mechanical transaction, stripping it of the psychological safety that parties often need in order to resolve conflict.

Principles to uphold

Transparency. The opacity of such systems can undermine the parties’ ability to scrutinise outcomes. Guidelines suggest disclosing their use where it affects decisions (Draper 2019). Parties should understand how a recommendation was arrived at and should have the opportunity to challenge or override it.

Human oversight. Experts agree that such tools should aid rather than replace human judgement (ADR Institute of Canada 2023). Arbitrators and mediators must review all outputs: the tools can analyse data, but a human must assess context, catch errors, and ensure that cultural and emotional factors are considered. Unchecked reliance risks procedural unfairness and may later be held invalid (JAMS ADR 2024).

Impartiality. In theory such a system, absent bias, could be indifferent as between the parties. In reality, training data and algorithms can embed subtle biases or preferences (Eftekhar 2025). Ensuring neutrality requires carefully curated data and oversight. Practitioners must verify that the logic applied is free from undue favouritism toward any side or outcome.

Confidentiality and security. Mediation is often confidential. Using such tools requires strict data protection, especially where third-party servers process private information (ADR Institute of Canada 2023). Any platform must be vetted to prevent the disclosure of sensitive information, for example by using systems hosted within the practitioner’s own environment rather than public services. Breaches of confidentiality could contravene mediator ethics rules and erode party confidence.

Due process risks

Delegating aspects of mediation to such systems raises due process concerns akin to those in arbitration. One commentator cautions that novel use could give rise to due process violations or a claim of irregularity if parties are caught off guard (JAMS ADR 2024). For example, if a proposed settlement is delivered without allowing the parties to argue their case fully, or if one party had no role in the selection of training data, a party might object to enforcement. Maintaining procedural fairness means that the parties should consent to the role such a system plays and should retain final control over accepting its recommendations (JAMS ADR 2024; ADR Institute of Canada 2023).

Bias and legitimacy

The potential for bias presents a serious ethical challenge to legitimacy. By design, machine learning systems reflect the data on which they are trained. If historical case data or heuristic rules contain gender, racial, or cultural bias, a system operating in this field will mirror those biases. As has been observed, if there is bias in the input, there will be bias in the output (Eftekhar 2025). Experts in international mediation have likewise raised concerns about cognitive, linguistic, age, and gender biases embedded in algorithms (Fasolo 2025). Parties who suspect that a system has been trained on skewed data may refuse to accept its conclusions, undermining trust in the process (ADR Institute of Canada 2023).

This problem is not merely technical; it strikes at the heart of legitimacy. Unlike human bias, which is subject to ethical scrutiny and corrective feedback in real time, algorithmic bias is often invisible and unaccountable. Worse, once embedded in opaque models, such bias can silently skew outcomes across many cases without the parties’ knowledge. In my view, relying on tools with hidden or untraceable bias risks systemic unfairness and delegitimises the very outcomes that are meant to promote resolution.

At the same time, such systems could reduce certain human biases, such as fatigue or inconsistent emotion-driven decisions, if properly managed. Studies suggest that they can provide a neutral perspective by analysing facts uniformly (Cao, Cheung and Li 2023). However, scholars caution that algorithmic fairness is not automatic. Even demonstrably fair algorithms can perpetuate hidden biases, and any algorithm that weighs certain factors, such as historical outcomes, inherently favours some results. Moreover, large language models are known to fabricate information (JAMS ADR 2024), raising additional fairness issues where output is not verified.

Mitigation measures

Bias auditing. Toolkits developed for this purpose aim to test and correct bias before deployment (Bellamy et al. 2018). In dispute resolution, such tools could scan recommendations for signs of unfair patterns, for example a consistent tendency to favour businesses over individuals.

Party control of data. Parties could negotiate which data a system uses. Excluding sensitive attributes such as race or nationality from the input can reduce some biases.

Guidelines and regulation. Emerging best practice emphasises that users must understand the limitations of these tools, retrain models where bias is found, and keep a human in the loop able to override unjust results (ADR Institute of Canada 2023).

If bias is not managed, the legitimacy of mediation itself suffers. Perceptions of fairness are crucial: parties are more likely to abide by a settlement if they believe the process was fair (Cao, Cheung and Li 2023). Conversely, a tainted outcome, even if technically sound, may not be regarded as just or trustworthy. Ensuring transparency and fairness is therefore not merely an ethical ideal but a practical necessity for enforceable and durable outcomes.

SUMMARY *

The article’s legal finding is straightforward and worth stating plainly because it dissolves much of the anxiety around the subject: a mediated settlement binds because the parties signed it, and the question of what assisted in producing the text is legally irrelevant to its enforceability. Where the difficulty lies is in the two conditions that surround the signature, namely that the parties knew what was assisting and consented to it, and that each had a genuine opportunity to be heard. The article’s second argument is of a different order. It holds that even where every legal condition is satisfied, something essential is lost, because what makes a party accept an outcome is the experience of having been heard by someone, and that experience cannot be simulated. The author accordingly arrives at a position that is permissive in law and restrictive in practice.

CONCLUSION

Artificial intelligence promises to transform mediation by offering data-driven insight and efficiency, but it cannot remove the need for core human values. Legally, outcomes reached with such assistance can bind only under the same conditions as any settlement: party consent and formalisation, since no special exception yet exists. Most laws still assume human neutrals (Norton Rose Fulbright 2024), so practitioners should proceed carefully when introducing these tools into binding decision processes (JAMS ADR 2024).

Equally important are the ethical dimensions. International standards demand fairness, impartiality, and accountability (UNCITRAL 2020). Such systems must be overseen by humans to ensure that they apply sound reasoning and remain neutral (ADR Institute of Canada 2023; Eftekhar 2025). Unmitigated algorithmic bias poses a clear threat to legitimacy; parties must be able to trust that no side is being quietly favoured.

In sum, the law currently treats artificial intelligence as a tool rather than a substitute: its outputs bind only to the extent that the parties accept them and due process is preserved (JAMS ADR 2024; UNCITRAL 2020). Ongoing developments in institutional guidelines and regulation aim to close the gap between technology and legal norms. Until then, the safest course is a human-centred approach: use these tools to assist mediation, but let human mediators and the parties retain ultimate control over the fairness and validity of the settlement. Ultimately, mediation’s legitimacy derives not from speed or efficiency but from meaningful human engagement. Such tools may assist with administrative efficiency, but allowing them to shape, suggest, or direct settlement terms distorts the nature of consensual dispute resolution. If the goal is not merely resolution but just resolution, then artificial intelligence must remain firmly subordinate to human mediators, as an instrument and not as an agent.

POTENTIAL SOLUTION *

The legal analysis is correct and one consequence of it deserves to be drawn out, because it is more reassuring than the framing suggests. A settlement is a contract, and contracts are set aside for want of capacity, for duress, for mistake, or for misrepresentation. None of these turns on what tool produced the draft. The exposure therefore lies almost entirely in non-disclosure: a party who discovers after signature that the mediator’s proposals were machine-generated, and who was not told, has an argument about the integrity of the process that would not otherwise exist. Disclosure at the outset closes that gap at negligible cost, and it is the single measure that most reduces risk. The Silicon Valley Arbitration and Mediation Center guidelines and the recent institutional notes point in the same direction.

The distinction that would sharpen the argument is between the uses of these tools, which the article treats as a single question. Summarising documents, preparing chronologies, translating, and administering scheduling raise no issue of principle and are already common. Analysing settlement ranges from comparable cases informs the mediator without touching the parties’ autonomy. Drafting proposed terms begins to shape the outcome. And generating an evaluation of a party’s case, or a recommendation as to what it should accept, is the point at which the tool has entered the substance of the process. These four uses call for different degrees of disclosure and consent, and a blanket position on either side of the debate obscures that.

On the empathy argument, which the article makes strongly, one qualification is worth adding because it does not weaken the conclusion. Online dispute resolution platforms handling very high volumes of low-value disputes have resolved millions of cases without any human present, and the parties have largely accepted the outcomes. This suggests that the human element matters in proportion to what is at stake and to whether the parties have a continuing relationship. Where a dispute concerns a defective purchase, an efficient and consistent process is what fairness looks like; where it concerns a family, a workplace, or a community, it is not. The author’s argument is therefore right about the disputes this journal’s readers actually handle, and stating the boundary makes it harder to dismiss.

Two additions on practice. Confidentiality is the most immediate exposure and the least discussed: a mediator who enters party material into a service that retains inputs for training may have breached an obligation before any question of fairness arises, and the remedy is contractual and technical rather than ethical. And the fabrication problem the article notes has already produced sanctions against practitioners in several jurisdictions for citations that did not exist, which is a reminder that the first professional risk here is not that these systems will be unfair but that they will be confidently wrong.

Finally, the article’s closing formulation, that the goal is not merely resolution but just resolution, identifies what is genuinely at stake. Efficiency arguments in dispute resolution have a long history of arriving as improvements and settling as reductions in what parties receive, and the question to ask of any such tool is not whether it saves time but who benefits from the time saved. Where the answer is the parties, the case is strong. Where the answer is the institution, the parties should be told, and should be asked.

* Added by the WMO Editorial Team

REFERENCES

ADR Institute of Canada (2023) Utilizing AI-Powered Tools in Arbitration.

Bellamy, R. K. E., Dey, K., Hind, M., Hoffman, S. C., Houde, S., Kannan, K., Lohia, P., Martino, J., Mehta, S., Mojsilovic, A., Nagar, S., Ramamurthy, K. N., Richards, J., Saha, D., Sattigeri, P., Singh, M., Varshney, K. R., and Y. Zhang (2018) AI Fairness 360: An Extensible Toolkit for Detecting, Understanding, and Mitigating Unwanted Algorithmic Bias. arXiv.

Burn, G., Morel de Westgaver, C., and V. Clark (2023) AI in International Arbitration: The Rise of Machine Learning. Bryan Cave Leighton Paisner arbitration survey.

Cao, N., Cheung, S. O., and K. Li (2023) Perceptive Biases in Construction Mediation: Evidence and Application of Artificial Intelligence. Buildings 13(10), Article 2460.

Draper, C. (2019) The Pull of Unbiased AI Mediators. International Journal on Online Dispute Resolution 6(1), 116 to 136.

Eftekhar, R. (2025) The Legal Framework Applicable to Using AI by an Arbitral Tribunal. DailyJus, 25 March 2025.

European Commission (2021) Proposal for a Regulation Laying Down Harmonised Rules on Artificial Intelligence, COM(2021) 206 final.

Fasolo, N. (2025) The Future of Mediation: AI, Funding, and Global Trends. DailyJus, 23 April 2025.

JAMS ADR (2024) The Use of AI in ADR: Balancing Potential and Pitfalls. JAMS ADR Insights, 31 January 2024.

Norton Rose Fulbright (2024) New Frontiers: Regulating Artificial Intelligence in International Arbitration.

United Nations Commission on International Trade Law (2006) UNCITRAL Model Law on International Commercial Arbitration 1985, with amendments as adopted in 2006.

United Nations Commission on International Trade Law (2019) United Nations Convention on International Settlement Agreements Resulting from Mediation, the Singapore Convention on Mediation.

United Nations Commission on International Trade Law (2020) Technical Notes on Online Dispute Resolution.

Zilberman, L. M. (2023) Will AI Mediators Soon Replace Humans? The Simple Answer Is No. The Daily Journal, 10 November 2023.

Supplementary references added by the WMO Editorial Team

Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence, and Annex III on high risk systems in the administration of justice.

UNCITRAL Model Law on International Commercial Mediation and International Settlement Agreements Resulting from Mediation, 2018.

Silicon Valley Arbitration and Mediation Center, Guidelines on the Use of Artificial Intelligence in Arbitration, 2024.

Chartered Institute of Arbitrators, Guideline on the Use of AI in Arbitration, 2025.

Katsh, E., & Rabinovich-Einy, O. (2017). Digital Justice: Technology and the Internet of Disputes. Oxford University Press, New York.

Susskind, R. (2019). Online Courts and the Future of Justice. Oxford University Press, Oxford.

Barocas, S., Hardt, M., & Narayanan, A. (2023). Fairness and Machine Learning: Limitations and Opportunities. MIT Press, Cambridge.

Tyler, T. R. (2006). Why People Obey the Law. Princeton University Press, Princeton, on procedural justice and perceived fairness.