Iffy Claim That Arbitrator “Outsourc[ed] His Adjudicative Role to Artificial Intelligence” Dismissed for Procedural Reasons

From the motion to vacate the arbitrator’s award in LaPaglia v. Valve Corp. (S.D. Cal.):

Arbitrator Saydah’s opinion for Mr. LaPaglia’s case has telltale signs of AI generation. The facts section cites facts that are both untrue and not presented at trial or present in the record. For example, Arbitrator Saydah’s decision states, without source attribution, that “Other platforms such as Roblox innovate in other ways with more mature content like horror elements paying off.” But Roblox is a children’s game with no horror elements. No testimony or document in the record, or anything, suggested otherwise. This sort of hallucinating or mixing up of facts is frequent when using AI tools to write content. Arbitrator Saydah’s seemingly random, uncited reference to Roblox’s marketing strategy that is only tangentially related to the parties’ dispute betrays the use of artificial intelligence to find “facts.”

Similarly, the decision states that “Just last year Sony and Microsoft partnered together to explore cloud gaming and streaming solutions using Microsoft Cloud Azure” and that “There is also major competition from China with their own developers and platforms, and also competition from companies in the United States, in the race to capture the Chinese market for PC Games.” Neither of these statements were in the record or otherwise evidenced or even argued, and neither fact findings bear any citations, again demonstrating Arbitrator Saydah relied on generative AI to determine the facts of the case and make decisions on market power and competition for him.

Prompted by these observations, Mr. LaPaglia’s counsel’s law clerk asked ChatGPT whether it believed the Roblox paragraph was written by a human or AI. ChatGPT stated that the paragraph’s awkward phrasing, redundancy, incoherence, and overgeneralizations “suggest that the passage was generated by AI rather than written by a human.”

The decision must be vacated because, by relying on AI to reach his ruling, Arbitrator Saydah exceeded his authority bound by the scope of the parties’ arbitration agreement. The parties’ arbitration agreement empowers “a neutral arbitrator” to resolve disputes between them. Dkt. No. 1-5 (Exhibit 4). The agreement also binds the arbitrator, who is responsible for supplying “a written decision” and a “statement of reasons” for their holding. Id. An arbitrator’s reliance on generative AI to replace their own role, and the parties’ submissions, in the litigation process betrays the parties’ expectations of a well-reasoned decision rendered by a human arbitrator.

Just as courts have vacated awards when the decision-making is outsourced to a person other than the arbitrator appointed, so too must a court vacate when that decision making is outsourced to an AI. See Move, Inc. v. Citigroup Global Mkts., 840 F.3d 1152, 1159 (9th Cir. 2016) (vacating where “there is simply no way to determine whether” an unqualified “imposter” on the arbitration panel “influenced other members of the panel or that the outcome of the arbitration was affected by his participation”), citing Stivers v. Pierce, 71 F.3d 732, 747 (9th Cir. 1995) (“Particularly on a small board, … it is difficult if not impossible to measure the impact that one member’s views have on the process of collective deliberation. Each member contributes not only his vote but also his voice to the deliberative process.”); See also Bassett’s Adm’r v. Cunningham’s Adm’r, 50 Va. 684, 692 (Va. 1853) (“The authority of an arbitrator cannot be delegated to an agent… In this case the arbitrators appointed other persons to examine the accounts and vouchers in the case, and adopted their conclusions without examining the vouchers themselves. This was held sufficient ground for setting aside the award.”)

From Valve’s memorandum in support of the motion to dismiss:

Petitioner contends that Arbitrator Saydah used artificial intelligence (“AI”) to draft the award. His only purported evidence in support of this contention is that a law clerk working for his counsel claims he asked ChatGPT whether AI could have been used in writing one paragraph from Arbitrator Saydah’s 27-page award. Remarkably, Petitioner’s counsel [in the arbitration -EV] was recently sanctioned by an arbitrator for submitting a letter brief containing numerous citations on core issues that had been fabricated through the use of AI, including non-existent cases and cases that did not contain the quotations ascribed to them. (ECF No. 5-2 ¶ 8.) Notably, Petitioner here cites a case—Bassett’s Adm’r v. Cunningham’s Admir, 50 Va. 684 (1853)—that does not contain the quote he attributes to it or stand for the proposition he claims.

The quote given in the petition as being from Bassett’s Adm’r appears to instead be from an annotation is actually from an old print volume reporting the case and then summarizing other caselaw on the subject. That print material is in turn reproduced in this document on the site run by Strongsuit (which calls itself “The Most Advanced Legal AI Platform”).

Because the lawyer’s true superpower is to turn every question into a question about procedure—and every procedural question into a question about procedure related to procedure—Judge Ruth Bermudez Montenegro dismissed the case today on the grounds that it didn’t involve the over $75K in controversy required for federal diversity jurisdiction. No word from the court as to the merits of the underlying dispute about who used AI and when and whether that’s OK.