AI-Assisted Arbitrators – A Swiss Perspective

SIMON GABRIEL, HANNES LATZEL*

Summary

This article analyzes the admissibility of AI use in arbitral decision-making under Swiss law. For this purpose, it distinguishes between AI-Assisted Drafting, Supervised AI Arbitrators and Autonomous AI Arbitrators. To reconcile the admissibility concerns surrounding Autonomous AI Arbitrators, the authors propose a hybrid model featuring a human appellate arbitrator.

A. Introduction

With the increasing use of AI in arbitration,¹ the question arises as to whether AI may be involved in the arbitral decision-making process.

For non-authoritative dispute resolution methods such as mediation,² the legal position appears straightforward: As long as the parties consent to AI’s involvement in the decision-making process, its admissibility can hardly be called into question.³

Arbitration is authoritative in nature. The arbitrator is empowered to adjudicate the dispute⁴ and his or her decision-making process culminates in a binding and widely enforceable arbitral award.⁵ While this power derives from the arbitration agreement, its exercise is subject to the minimum requirements acceptable to the state as set out in the lex arbitri.⁶ Hence, the question whether AI may be involved in arbitral decision-making is governed by party autonomy, but within the limits of the applicable lex arbitri.

This article assesses whether the involvement of AI in arbitral decision making is permissible under the Swiss lex arbitri where the parties agree to such AI use. In the authors’ view, the answer to this issue is determined by the extent to which AI is applied in the decision-making process, i.e. AI’s cognitive control relative to that of the arbitrator.⁷ Accordingly, the question is analyzed by reference to three scenarios reflecting increasing degrees of AI’s control:⁸

• First, scenarios where arbitrators use AI to perform ancillary tasks in support of the arbitrator’s drafting and to generate specific sections of the award – excluding the substantive decision-making part (“AI Assisted Drafting”).

• Second, cases in which the entire award is, in a first step, generated by AI and, in a second step, scrutinized and adopted by a human arbitrator (hereinafter referred to as “Supervised AI Arbitrator”).
• Third, the scenario where an award is written and issued solely by AI, without any human involvement. We hereinafter refer to this AI use as “Autonomous AI Arbitrator”.

In Sections B–D, the authors will examine whether these party-agreed AI applications conform to the Swiss lex arbitri and – where helpful for the argument – whether they are in compliance with the New York Convention (“NYC”).

B. AI-Assisted Drafting

I. Practical Application

Surveys show that arbitration stakeholders currently use AI mostly for document review, data analysis and research.⁹ Such tasks are usually also central to the work of arbitrators. Arbitrators can therefore leverage AI to fulfill their responsibilities more efficiently,¹⁰ particularly with regard to the drafting of the award.¹¹ AI-Assisted Drafting can for example take the form of AI summarizing the procedural history or the parties’ key arguments,¹² conducting sanity checks of the arbitrator’s legal reasoning or assisting the arbitrator with mastering voluminous files or legal research.¹³

II. State of Discussion

To the best of the authors’ knowledge, there are no scholarly opinions according to which AI-Assisted Drafting within the scope described above would be deemed inadmissible. It is merely discussed whether there is a duty for arbitrators to disclose such AI utilization to the parties.¹⁴ Therefore, it can be inferred that scholars holding this view must consider AI-Assisted Drafting admissible in any case – as in the scenario discussed in this article – assuming the agreement by the parties.

III. Analysis

In the authors’ view, it is justified to accept the admissibility of party agreed AI-Assisted Drafting.

With AI-Assisted Drafting, AI either contributes no text to the award at all (e.g. when conducting research) or its input is restricted to specific administrative or procedural parts. Even in the latter case, the arbitrator remains in charge of the substantive decision-making process as the arbitrator authors the merits and operative section of the award without any AI input. AI’s contribution in such scenarios is therefore negligible in its impact on the arbitrator’s final decision. Hence, there are in our view no compelling reasons as to why AI-Assisted Drafting should not be permissible where the parties have agreed thereto.

As in AI-Assisted Drafting, AI is but a tool without independent judgment function, the authors submit that arbitrators may in principle use it even without the parties’ consent. This requires, however, that the use of AI does not compromise the required degree of confidentiality.

If arbitrators use a Local AI System¹⁵ for AI-Assisted Drafting, they are in our view not required to specifically notify the parties, as such AI systems offer the highest possible level of confidentiality by design. For arbitrators that intend to utilize an Enterprise AI System,¹⁶ we submit that they should first hear the parties on the confidentiality aspect. If one of the parties objects to such AI use, this objection must, in the authors’ view, be observed. The reason is that the use of Enterprise AI Systems always involves an AI system of a third party for processing user data. The parties should remain the masters of their data when services of third parties are involved. Consumer AI Systems¹⁷ should only be used if the parties expressly agree as there is a significant risk that the AI system’s provider utilizes confidential data for training purposes.

IV. Conclusion

Under the Swiss lex arbitri, the parties are free to agree that the arbitrator may apply AI-Assisted Drafting. If arbitrators use a Local AI System, they may apply AI-Assisted Drafting even without the parties’ consent in the sense of an IT-tool which does not pose any risks for breach of confidentiality.

C. Supervised AI Arbitrators

I. Supervised AI Arbitrators Practical Application

AI arbitrators are already used in practice, albeit merely sporadically. The American Arbitration Association’s “AI Arbitrator”, for instance, is, in our view, a Supervised AI Arbitrator as it involves an AI-generated decision reviewed and finally adopted by a human arbitrator.¹⁸

Given the arbitration institutions’ growing focus on AI and AI’s expanding capabilities,¹⁹ more cases will likely be decided by Supervised AI Arbitrators in the future.

II. State of Discussion

Legal scholars are divided on the admissibility of Supervised Arbitrators. On the one hand, there is consensus that this type of AI use is admissible as long as the arbitrator fulfills his review function and provided that he does not transfer any final decision-making power to AI.²⁰ Disputed, however, are the precise circumstances under which AI’s drafting of the award would constitute an inadmissible delegation of decision-making authority.

For some scholars, the answer to this question depends on which sections of the award are generated by AI. According to this view, AI may draft the award’s procedural section. The substantive section, however, must be personally drafted by the human arbitrator as it forms part of the arbitrator’s decision-making authority.²¹ According to another view, human arbitrators retain decision-making authority as long as all decisions in the award result from the arbitrators’ own assessment – irrespective of which sections are drafted by AI.²² 

The issue at hand closely resembles the question to what extent the tribunal secretary (“Secretary”) may be tasked with drafting the award. Various authors therefore suggest that the considerations pertaining to Secretaries be applied to Supervised AI Arbitrators by analogy.²³

III. Analysis

Given the similarities between the two issues, the authors concur that the considerations applicable to Secretaries should in principle extend to Supervised AI Arbitrators. Same as a human Secretary, the Supervised AI Arbitrator is able to perform cognitive operations and thus, in principle, replace the work of the human arbitrator. In both instances, the human arbitrator may be tempted to rely on these types of cognitive assistance and may be influenced in his or her judgment by a first draft. It is therefore appropriate to briefly outline the state of Swiss doctrine on Secretaries and then apply the same to the present subject:

Under the Swiss lex arbitri, arbitrators must not delegate their decision making authority to Secretaries.²⁴ Within certain limits, the Secretary may assist in the award-drafting without thereby assuming decision-making authority. According to a minority view, the Secretary may prepare only certain parts of the award.²⁵ Yet, the majority view is that the Secretary may draft the entire award provided that the arbitral tribunal closely supervises the drafting process and issues clear prior instructions.²⁶ It goes without saying that certain institutional arbitration rules provide for their own (different) standards in this regard.²⁷ 

If the latter view is applied in the context of AI-generated awards, Supervised AI Arbitrators are in principle admissible provided that they are sufficiently instructed and supervised. Yet, it remains unclear under what conditions the arbitrator’s duty to properly instruct and supervise AI is satisfied.

Against this background, the authors propose a two-pronged test:

With respect to the criterion of instruction, a distinction must be drawn between procedural and substantive parts of the award.

For the procedural part, the requirement of proper instruction is already satisfied, if AI is prompted to generate the respective section in a certain manner. For the substantive part, however, proper instruction requires that arbitrators not only prompt AI to create a draft but also specify how and on what factual and legal basis the key issues are to be decided. Similarly, for the dispositive section of the award, the instruction requirement is only satisfied if the arbitrators’ prompt determines which prayers for relief are to be granted and which are not.

Only if Supervised AI Arbitrators meet this requirement, can it be ensured that the cognitive process leading from the parties’ submissions to the decision(s) is first carried out by the arbitrator and not by AI. Otherwise, there is a considerable risk that Supervised AI Arbitrators’ initial output unduly influences the human arbitrators’ findings.

Regarding the requirement of supervision, a distinction must be drawn between the party-agreed use of Supervised AI Arbitrators and their application absent the parties’ consent.

The authors propose that in the latter case (no consent), Supervised AI Arbitrators must not be used for the drafting of any substantive reasoning. The parties have a legal right to a human decision maker and author of the award – but they must then bear the time- and cost-consequences.

For party-agreed use of Supervised AI Arbitrators, the supervision will include the specific review of the main pillars of the substantive decision including double-checks on correct quoting of evidence. A double-check of every detail of the award is, in the authors' view, not required (unless differently agreed by the parties). Indeed, the main advantage of Supervised AI Arbitrators is the increased time- and cost-efficiency.²⁸ Hence, requiring arbitrators to fully review the AI-generated award would defeat the underlying purpose of opting for a more efficient dispute resolution.

When this standard is applied to Supervised AI Arbitrators, human arbitrators must at a minimum verify the accuracy of the main pillars of the award’s reasoning and their consistency with the key evidence. This allows for leveraging the efficiency of artificial intelligence while preserving the accuracy of human intelligence.

Finally, one could argue that it is sufficient for a human arbitrator to do a sanity check after receipt of an AI-generated draft award (i.e. without giving prior instructions on the outcome). The authors considered this approach as well and opted against it: First, in practice, the temptation to accept the draft award of the AI is hard to resist and may thus unduly influence the sanity check by the human. Second, even if the human resists this temptation, the work required for a proper sanity check will be comparable to a proper instruction before the AI drafts the award. Hence, the sanity check after the draft award of the AI has been prepared involves the risk of undue influence, while hardly adding to efficiency.

IV. Conclusion

Under the Swiss lex arbitri, Supervised AI Arbitrators are in principle admissible provided that they are agreed upon by the parties as well as sufficiently instructed and supervised by a human arbitrator.

Regarding the award’s substantive and operative sections, the requirement of instruction is satisfied if the arbitrator’s prompts specify how the key issues are to be decided and which prayers for relief are to be granted.

For the procedural part, the mere prompt for AI to generate this section in a certain manner satisfies the requirement of instruction.

The supervision requirement is satisfied if the arbitrator verifies the accuracy of the main pillars of the award’s reasoning and their consistency with the key evidence and the prayers for relief.

D. Autonomous AI Arbitrators 

I. Practical Application

Use cases of Autonomous AI Arbitrators have yet to emerge in legal practice. Their admissibility remains highly contested, as will be outlined below.

II. State of Discussion

Insofar as Swiss scholars address the issue, they agree that AI must not make decisions independently. Hence, they are likely to consider Autonomous AI Arbitrators inadmissible. It has, however, not yet been addressed in Swiss legal doctrine whether Autonomous AI Arbitrators are inadmissible even where the parties specifically agree to AI adjudication.²⁹

In international literature, the admissibility of Autonomous AI Arbitrators has received virtually no attention in relation to specific leges arbitri but is instead mainly discussed in view of the NYC:

The majority view is that awards rendered by Autonomous AI Arbitrators are per se unenforceable under the NYC.30 The main argument against the admissibility of Autonomous AI Arbitrators is that the NYC confines arbitral functions to human arbitrators. This view is based both on historical and grammatical interpretations of the NYC’s provisions.

Proponents of a minority view argue that priority should be given to a teleological interpretation, according to which Autonomous AI Arbitrators are consistent with the NYC’s object and purpose.³¹

III. Analysis

The majority view is persuasive. The idea of AI deciding a case autonomously deviates so significantly from the common understanding of authoritative decision-making that Autonomous AI Arbitrators cannot be considered admissible under the current wording of the NYC.

The same holds true for the Swiss lex arbitri. Provisions such as Art. 189 PILA, which require the original award to be signed, indicate that the legislator only envisaged human beings to serve as arbitrators.³² Moreover, it is clear that under the lex arbitri, legal entities cannot be arbitrators.³³ If non human entities with legal personality cannot act as arbitrators, non-human entities without legal personality, such as AI, are a fortiori excluded from this function. Hence, Autonomous AI Arbitrators are not admissible under Swiss lex arbitri, even in cases where the parties agreed on this form of adjudication.

While accepting this conclusion from a doctrinal viewpoint, the authors deem it worthwhile to examine whether the position is also supported by functional and normative considerations.

First, not permitting Autonomous AI Arbitrators could be justified by their lack of accountability. Yet, this concern is mitigated where liability is governed by the contract between the parties and the provider of the AI system.

Second, the question arises whether arbitral authority is legitimate only if exercised by humans. If this were true, authority wielded by party-agreed Autonomous AI Arbitrators would be considered illegitimate. Yet, it is difficult to conceive how the parties’ agreement to submit their dispute to a non-human arbitrator could be at odds with notions of justice.³⁴

One could argue that human arbitrators possess attributes that render them better suited for decision-making than AI systems. However, even if this were the case, this would be irrelevant where the parties, by agreeing to an Autonomous AI Arbitrator, had accepted any of its potential limitations.35 The assumption that only humans can be vested with legitimate decision-making authority therefore appears doubtful.

If this position were nevertheless followed, the guarantee of access to a judicial authority would be limited to human adjudication. This conclusion is difficult to justify as party-agreed Autonomous AI Arbitrators in the authors’ view broaden access to justice. This applies particularly to low-value claims for which initiating arbitration would ordinarily not be worth the effort.

IV. Proposal: Autonomous AI Arbitrator with Human Appeal Option

To reconcile the requirement of human decision-making with the advantages of (future) Autonomous AI Arbitrators, the authors propose the following: When parties agree on Autonomous AI Arbitrators, they should at the same time foresee a human appellate arbitrator. Analogous to the mechanism in Dispute Adjudication Boards, the award of the Autonomous AI Arbitrator would only become final if no party objected thereto.

In case a party is dissatisfied with the award, it would be entitled to refer the case to the human appellate arbitrator. If no party appeals the award of the Autonomous AI Arbitrator it appears to be acceptable to the parties and becomes final.

Finally, it is proposed that the party that wishes to seize the appellate arbitrator must advance the full costs for the appeal (subject to final allocation by the human arbitrator). That way, the losing party must be sufficiently convinced that the Autonomous AI Arbitrator made a mistake, and appeal cases are reduced to the extent necessary.

This approach could give the parties the benefit of highest efficiency as long as the quality is right and the chance for human correction, if it is not.

E. Summary Chart for AI-Assisted Awards

In the sense of a summary, the following chart reflects the different types of AI-systems used and the different use-cases as analyzed in the present contribution for ease of reference:

Footnotes

* Dr. Simon Gabriel LL.M., Attorney at Law, Partner at GABRIEL Arbitration AG, s.gabriel@gabriel-arbitration.ch; Dr. Hannes Latzel, Associate at GABRIEL Arbitration AG, h.latzel@gabriel-arbitration.ch. 

1. See WHITE & CASE/QUEEN MARY UNIVERSITY, 2025 International Arbitration Journey, The path forward: Realities and opportunities in arbitration, 2025, p. 3, (https://www.whitecase.com/insight-our-thinking/2025-international-arbitration-survey), last accessed 30 April 2026. 

2. BERNHARD BERGER, International and Domestic Arbitration in Switzerland, Bern 2025, para. 154. Mediators may solely make non-binding proposals, which become binding only if the parties agree, see DANIEL GIRSBERGER/NATHALIE VOSER, International Arbitration, Zurich 2024, para. 36 et seq.

3. It is therefore unsurprising that autonomous AI mediator services are already being offered (ABILASHA VIJ, Arbitrator-Robot: Is A(I)DR the Future?, in: ASA Bull. 1/2021, pp. 123-146, p. 131-132 with reference to “Smartsettle One”).

4. See BERGER (Fn. 2), para. 9. 

5. GARY B. BORN, International Commercial Arbitration, Alphen aan den Rijn 2021, 2.02[B] [1][b][iv]. 

6. See MICHAEL E. SCHNEIDER/MATTHIAS SCHERER, Art. 182 IPRG, in: Karl Spühler/Luca Tenchio/Dominik Infanger (eds.), Basler Kommentar Internationales Privatrecht, 4th ed., Basel 2021 (cited as: BSK IPRG-AUTHOR), Art. 182 N 1.

7. ANNE-CATHRINE HAHN, Limits to the Use of AI Tools by Swiss Arbitral Tribunals: A Tentative Analysis, in: ASA Bull. 4/2025, pp. 746–760, p. 757, uses this terminology in relation to the disclosure of AI use.

8. See KAVERI G. KURUVATTI, A Study of the Interface Between Artificial Intelligence and Arbitration, in: Indian J.L. & Legal Rsch. 4/2022-2023, pp. 1-13, p. 4.

9 WHITE & CASE/QUEEN MARY UNIVERSITY (Fn. 1), p. 28.

10 DOMINIK STEFER/VICTORIA FRICKE, From algorithms to awards: exploring the technical and legal boundaries of AI's contributions to the work of arbitrators, in: William W. Park (ed.). Arbitration International 1/2025, pp. 49–70, p. 50.

11 See STEFER/FRICKE (Fn. 10), p. 64 et seq.

12 STEFER/FRICKE (Fn. 10), p. 65.

13 STEFER/FRICKE (Fn. 10), p. 64–65.

14 See HAHN (Fn. 7), p. 754; JONATHAN SILBERSTEIN-LOEB, Arbitrators, Decision Making, and Generative AI, in: ASA Bull. 4/2023, pp. 831-840, p. 836; see KURUVATTI (Fn. 8), p. 4.

15 Systems deployed and run on individual user devices which are detached from any AI processing in third-party data centers; i.e. confidential by design

16 Systems offered by professional providers which contractually guarantee that user data is not used for training of AI-models or stored after AI-processing; i.e. confidential by contract.

17 Publicly accessible AI-tools that may store user data and/or use it for training purposes.

18 Compared to the Supervised AI Arbitrator (supra A.) this system adds an additional review stage, allowing the parties to validate AI's summary of their submission before the award is generated, see: (https://www.adr.org/news-and-insights/ai-arbitrator-faq/), last accessed 30 April 2026.

19. MARIKE PAULSSON/SUPRITHA SURESH, AI: The Modern Tribunal Assistant – Impact on Enforceability of Arbitral Awards under the New York Convention, in: Jus Mundi Arbitration Review, 1/2024, pp. 101-118, p. 115. 

20. SILBERSTEIN-LOEB (Fn. 14), p. 833 et seq.; see also YAN LU/FATT HEE TIE, Utilizing Artificial Intelligence in Investor State Dispute Settlement: Present Issues and Future Challenges, in: JMCL 1/2024, pp. 31-62, p. 59 et seq; see also PAULSSON/SURESH (Fn. 19), p. 110. 

21 ABIR AL MAHDOURI, Public Policy and the Enforcement of AI-Generated Arbitral Awards, in: J. Int’l Arb. 6/2025, pp. 759-790, p. 765 with further references; see also SCC, SCC Guide, Guide to the use of artificial intelligence in cases administered under the SCC Rules, p. 3.

22 See SILBERSTEIN-LOEB (Fn. 14), p. 833 et seq.; HAHN (Fn. 7), p. 754; LU/TIE (Fn. 20), p. 59 et seq.; PAULSSON/SURESH (Fn. 19), p. 110.

23 HAHN (Fn. 7), p. 754; SILBERSTEIN-LOEB (Fn. 14), p. 831; MAHDOURI (Fn. 21), p. 765; PAULSSON/SURESH (Fn. 14), p. 104, p. 110.

24 CHRISTOPHER BOOG/SONJA STARK-TRABER, Art. 365 ZPO, in: Christopher Boog et al. (eds.), Berner Kommentar Schweizerische Zivilprozessordnung, Art. 353-399 ZPO und Art. 407 ZPO, Band III – Binnenschiedsgerichtsbarkeit, Bern 2014 (cited as: BK CPC-AUTHOR), Art. 365 N 20; STEFAN GRUNDMANN, Art. 365 ZPO, in: Thomas Sutter-Somm et al. (eds.) Kommentar zur Schweizerischen Zivilprozessordnung (ZPO), Art. 1-218 ZPO (cited as: CPC Comm.-AUTHOR), Art. 365 N 10; PHILIPP HABEGGER/MICHAEL FEIT, Art. 365 ZPO, in: Karl Spühler/Luca Tenchio/Dominik Infanger (eds.), Basler Kommentar Schweizerische Zivilprozessordnung, 4th ed., Basel 2024 (cited as: BSK CPC-AUTHOR), Art. 365 N 6a.

25 BERGER (Fn. 2), para. 1010.

26 BSK CPC-HABEGGER/FEIT (Fn. 24), Art. 365 N 6a, referring to Decision of the Swiss Supreme Court (“DSC”) 4A_709/2014 of 21 May 2015, para. 3.2.2; BK CPC-BOOG/STARK TRABER (Fn. 24), Art. 365 N 21; MICHAEL FEIT/CHLOÉ TERRAPON CHASSOT, The Swiss Federal Supreme Court Provides Guidance on the Proper Use of Arbitral Secretaries and Arbitrator Consultants under the Swiss lex arbitri: Case Note on DFC 4A_709/2014 dated 21 May 2025, in: ASA Bull. 4/2015, pp. 897–917, p. 908; CPC Comm.-GRUNDMANN (Fn. 24), Art. 365 N 10; GABRIELLE KAUFMANN-KOHLER/ANTONIO RIGOZZI, International Arbitration: Law and Practice in Switzerland, Corby Northants 2015, para. 7.156 and TARKAN GÖKSU, Schiedsgerichtsbarkeit, Zurich 2014, para. 879.

27 See for instance SWISS ARBITRATION CENTRE, Swiss Rules of International Arbitration, Practice Note, 2025, para. 130 et seq.; ICC, Note to Parties and Arbitral Tribunals on the Conduct of The Arbitration, 2021, para. 216 et seq.; LCIA, LCIA Guidance Note for Parties and Arbitrators, 2020, para. 198 et seq.; SIAC, Practice Note on the Appointment of Tribunal Secretaries, 1 January 2025).

28 See MAHDOURI (Fn. 21), p. 765; SAID GULYAMOV, Digitalization of international arbitration and Dispute Resolution by Artificial Intelligence, in: WBML 9/2022, pp. 79-85, p. 83; KURUVATTI (Fn. 8) p. 4.

29 For instance in situations where a regular arbitration would be too costly for small claims.

30 See RHEA SPECOGNA, Grenzen und Möglichkeiten privater Streiterledigung, in: PIRR – Perspectives on Innovation, Risk and Responsibility, 1/2023, pp. 109-125, p. 111–112; MOHAMED S. ABDEL WAHAB, Welcome to the Future: The AI Arbitrator – An Unwelcomed Reality or Science Fiction?, in: Stavros Brekoulakis et al. (eds.), Achieving the Arbitration Dream: Liber Amicorum for Professor Julian D.M. Lew KC XY 2023, pp. 305-316, p. 315; HORST EIDENMÜLLER/FAIDON VARESIS, What is an Arbitration?, Artificial Intelligence and the Vanishing Human Arbitrator, p. 35; ANNABELLE O. ONYEFULU, Artificial Intelligence in International Arbitration: A Step Too Far?, in: The International Journal of Arbitration, Mediation and Dispute Management, 1/2023, pp. 56-77, p. 60; see VIJ (Fn. 3), p. 141.

31 EIDENMÜLLER/VARESIS (Fn. 30), p. 5, p. 34-37; HUANG YING (IRENE) NG/VALERIA BENEDETTI DEL RIO, When the Tribunal is an Algorithm: Complexities of Enforcing Orders Determined by a Software under the New York Convention, in: Katia Fach Gomez/Ana M. Lopez-Rodriguez (eds.), 60 Years of the New York Convention: Key Issues and Future Challenges, pp. 121-134; concurring CHRISTINE SIM, Will Artificial Intelligence Take over Arbitration?, in: Asian Int'l Arb. J. 2/2018, para. 2.1.

32 Regarding the NYC: VIJ (Fn. 3), p. 136.

33 GÖKSU (Fn. 26), para. 873.

34 See MICHAEL J. BROYDE/YIYANG MEI, Don’t Kill the Baby! The Case for AI in Arbitration, in: Journal of Law & Business, 1/2024, pp. 119-173, p. 146-147.

35 The same holds true for any issues relating to confidentiality or impartiality.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.