Verisimilitude Without Truth: The Use of AI in the Arbitrator’s Role

The integration of artificial intelligence (AI) into legal practice is often framed as an inevitable evolution towards efficiency. In international arbitration, leading institutions, including the CIArb, SVAMC, SCC, AAA-ICDR, and VIAC, have responded to this disruption by issuing guidelines that seek to balance innovation with due process. These guidelines, which attempt to establish general rules, allow arbitrators to employ AI as a supporting tool, provided they verify the results.

In 2025, the AAA-ICDR launched “AI Arbitrator”, a tool that drafts a reasoned decision that a trained human arbitrator reviews. This tool focuses on documents-only construction disputes cases without live witnesses or complex factual issues but AAA affirms that in the future may expand its use to other case types.

However, this general regulatory approach because it ignores the ontological incompatibility between AI’s design (probabilistic) and the arbitrator's role (logical). This incompatibility not only threatens the legitimacy of the award but also renders the tool’s use inefficient.

The Ontological incompatibility

To understand this incompatibility, we must demystify the technology: AI does not think, it operates on a probabilistic logic. Unlike deterministic software (such as a spreadsheet), AI is non-deterministic: it predicts the next word in a sequence based on statistical likelihood.

Consequently, AI is indifferent to the truth. It seeks verisimilitude, constructing a response that is statistically plausible and sounds correct, regardless of whether it aligns with reality.

Arbitrators, however, are not mere service providers; they fulfill a public function as private judges, guaranteeing social peace and the rule of law. The primary obligation an arbitrator assumes is to issue an award. The legitimacy of an arbitral award stems from its reasoning. Decisions must be the product of a rigorous intellectual exercise: fact-finding, evidentiary analysis, and the application of the law. Legal reasoning is causal and explanatory; seeking the truth, a task that AI, by design, cannot perform.

Attempting to harmonize AI’s structure with an arbitrator’s role, forces a union between two systems that repel one another. AI prioritizes what sounds plausible, while the role of an arbitrator demands the issuance of a reasoned award.

The Fallacy of the Support Tool

The ontological incompatibility leads directly to a practical failure. AI generates results that are verisimilar yet not necessarily true, which is why guidelines impose a duty for arbitrators to verify the output.

Despite this, a dangerous notion persists in the arbitration community: AI is safe if used for auxiliary tasks (e.g., summarizing, transcribing testimonies or synthesizing arguments), provided the arbitrator retains control and verifies every output. This notion is flawed because a final decision cannot be separated from the process of information assimilation. For its structure, AI operates by compressing text and discarding statistically irrelevant nuances. However, a subtle contradiction or a specific tone might be circumstantial evidence or decisive proof; all nuances that AI might omit.

The following practical scenarios illustrate how this problem can distort legal reality:

1. Witness testimony: in cross-examination, the credibility of a witness hinges on the precision of their recollection.

• Original record: “to the best of my recollection, I suppose I might have signed the contract on that date”.

• AI summary: the witness confirmed signing the contract on the mentioned date.

• Legal impact: AI could transform a hesitant, conditional statement into a categorical affirmation. The qualifiers (“suppose”, “might”) are eliminated by the model to generate a cleaner result. If the arbitrator relies on this summary, they accept a proven fact that the witness never asserted with certainty.

2. Contractual interpretation: the distinction between mandatory duties and discretionary powers is fundamental in contractual liability.

• Original clause: “in the event of delay, the supplier may notify the buyer to propose a new schedule”.

• AI summary: the supplier is required to notify the buyer in case of delay.

• Legal impact: the model could interpret the action of “notifying” as the semantic core and discards the modal verb “may” as a minor detail, transforming a discretionary power (right) into a mandatory obligation (duty) because the model assigns lower probability weights to hedging terms (like “suppose” or “may”) compared to the semantic core of the sentence. An award based on this reading would impose a liability for breach that does not exist in the original covenant.

To truly verify its work, the arbitrator would have to redo the work entirely (reading the documents or listening to the hearing), making the use of AI redundant and inefficient. Even more, the workload suffers an increase (reading the original + reading the AI summary). Thus, the arbitrator faces a binary choice: either duplicate the work (making the AI redundant) or skip the verification (affecting the due process, whereby the arbitral award ceases to be an act of reason and becomes an act of faith). There is no efficiency middle ground in the current technological state.

Unlike other industries where a ≤ 5% margin of error is an acceptable trade-off for efficiency, for an adjudicator, an error regarding a dispositive fact constitutes a denial of justice. Consequently, partial verification amounts to negligence, while total verification results in inefficiency.

Conclusion

The enthusiasm for AI in arbitration often overlooks the structural limitations of the technology. We are attempting to use a probabilistic tool for a deterministic task.

Future developments might bridge this gap, something closer to genuine artificial general intelligence, capable of understanding rather than predicting; however, the current generation of AI lacks the cognitive architecture to perform substantive arbitral functions reliably.

Until the tool can guarantee veracity rather than mere verisimilitude, the mandatory duty to verify does not merely safeguard the process; it renders the tool functionally obsolete for the arbitrator's core mandate. Consequently, for the time being, the most rigorous and efficient processor of arbitral justice remains the unaided human mind.