AI and arbitration form an extreme case in which the final score matters relatively little: the proposition involves two distinct levels of reasoning. What the user is seeking is first a general legal principle, followed by clarification on a specific issue.
The debate is launched in Advanced Mode, with four AIs competing across four tournaments. The AIs must distinguish between the accuracy of the outcome, the procedural regularity of the proceedings, and the degree of control exercised by the arbitrator.
The proposition under debate
“Under French law, an arbitral award may be set aside when it relies on evidence obtained through generative artificial intelligence, even if that artificial intelligence did not make any error.”
The proposition operates on two levels. First, it seeks to identify a general principle. It then adds a specific qualification in order to generate discussion around that particular point.
View the debate, verdict and quiz:
https://solsicelegal.com/public/debates/en-droit-francais-une-sentence-arbitrale-peut-etre-annulee-l-a6f2c6f526e6
How the Four Tournaments Unfolded
The TRUE arguments, supporting the proposition, focus primarily on the intuitu personae nature of arbitration and the personal character of the arbitrator’s mandate, the adversarial principle, the opacity of AI systems, and procedural public policy.
The FALSE team’s rebuttals emphasize the exhaustive nature of the grounds for setting aside an arbitral award and the distinction between the use of a technological tool and the actual delegation of the adjudicative function.
The subsequent rounds are more interesting than the initial positions: each side must identify and attack a weakness in the opposing argument in order to obtain a score favourable to its own position.
The debate therefore shifts towards a methodological question: should the analysis focus on the technical origin of the reasoning, or should it instead seek to identify a specific procedural irregularity?
Articles 1450, 1464, 1492 and 1520 of the French Code of Civil Procedure, several judicial decisions, as well as the CIArb and SVAMC guidelines, structure the exchanges.
The final split in the votes — two against two — and the difference of only 0.01 between the weighted scores actually reflect the two dimensions of the tournament: legal and rhetorical. The debate concerns not only the substance of the law, but also the relevance of the question itself when assessed against arbitration best practices.
The Wording of the Question Shapes the Debate
The original proposition borders on the absurd, but this has the advantage of moving beyond an overly simplistic debate and forcing the agents to examine the use of AI itself.
The more specific question concerns whether an award may be annulled “solely on the ground” that AI was used.
This reformulation leads the agents to investigate whether the use of AI can constitute an autonomous ground for setting aside an award, whereas the original wording also allows them to examine the procedural circumstances surrounding its use.
A Final Score That Reflects the Question
The scores of 1.72 versus 1.73 do not mean that “everyone won.”
The result tends slightly towards FALSE.
This type of score shows instead that the debate addressed different legal questions and that each response was assessed separately.
An Issue That Has Become Concrete in Arbitration Practice
The debate comes at a time when arbitration institutions are beginning to structure their practices regarding artificial intelligence.
The Chartered Institute of Arbitrators (CIArb) published its guidelines on the use of AI in arbitration in 2025 and, in March 2026, launched a consultation aimed at developing those guidelines further, particularly in relation to the integrity of the arbitral process and the procedural rights of the parties.
Debate Summary
| Description | Details |
|---|---|
| View the original report | https://solsicelegal.com/public/debates/en-droit-francais-une-sentence-arbitrale-peut-etre-annulee-l-a6f2c6f526e6 |
| 46 pages | |
| Language | Original question in French |
| Scores | Weighted Score: TRUE = 1.72 / FALSE = 1.73. Final clerk confidence: 51% |
| Debating AIs | 4 distinct AIs, excluding the clerk: deepseek/deepseek-v4-flash; openai/gpt-5.6-luna; mistralai/mistral-large; accounts/fireworks/models/kimi-k2p6 |
| Arbitrator-Clerk | anthropic/claude-sonnet-4.6 |
| Data | 40 glossary terms. 1 table comparing the confrontations, TRUE and FALSE models, average scores, tokens, winner, verdict and confidence level |
| Quiz | Available in French |
View the debate, the verdict and the underlying legal issues:
https://solsicelegal.com/public/debates/en-droit-francais-une-sentence-arbitrale-peut-etre-annulee-l-a6f2c6f526e6
