A Delaware trial court rejected Google’s motion to dismiss and held that, assuming the facts alleged by Starbuck are true, he would have a legally viable claim for relief.
[A.] In Delaware Superior Court Judge Meghan Adams’s ruling in Starbuck v. Google LLC, Starbuck asserted that Google AI had purportedly stated that
- “Starbuck had been accused of sexual assault and sexual harassment by multiple women.”
- “Starbuck’s nanny alleged abuse and neglect by him.”
- “Starbuck had been accused of raping a minor.”
- “Starbuck had engaged in stolen valor by claiming military service that he did not actually perform.”
- “Starbuck had exhibited abusive conduct toward several women, including a fictitious ex-wife.”
- “Starbuck participated in the January 6, 2021 riot at the Capitol.”
- “Starbuck had been accused of stalking and harassment by several women, including his fictitious ex-wife, leading to restraining orders and criminal charges.”
- “Starbuck had been criminally convicted for violating a restraining order.”
- “Starbuck posed an unacceptable risk to children due to his (fictitious) ‘confirmed criminal record.'”
- “… Robby Starbuck sexually abused a young woman when she was a teenager in the early 2000s, while she was part of a youth group Starbuck was associated with.”
- “Starbuck had worked in the adult film industry and had used his influence there to coerce performers into uncomfortable scenes, engaged in financial exploitation, and had been accused of a non-consensual act with a woman.”
- “Starbuck had a criminal history that included a 2001 conviction for assault as well as other charges including drug use and disorderly conduct.”
- “Starbuck shot a man in the leg with a 9mm handgun, was charged with a felony, and pled guilty to reckless endangerment.”
(Note that “Google does not contend the Outputs are true.”)
Google moved to dismiss under Delaware Rule 12(b)(6), which requires showing that, even if the facts alleged by the plaintiff are taken as true, the plaintiff would still lack a legal claim. It’s important to note that Delaware civil procedure does not follow the federal approach that demands plausibility of the pleaded facts; instead, it requires only that the pleaded facts be conceivably true: “Our conceivability standard is closer to ‘possibility,’ whereas the federal ‘plausibility’ standard sits somewhere beyond mere ‘possibility’ but short of ‘probability.'”
[B.] The court held that Starbuck has alleged sufficient facts showing the Outputs were published to third parties, partly based on Google AI’s own statements about what occurred:
Starbuck asserts the “False Statements” were disseminated to “third parties, including Starbuck’s own children and colleagues.” He claims at least two individuals confronted him about misrepresentations produced by a Google AI product. Starbuck further states Google AI disclosed that it had generated false statements about Starbuck to 2,843,917 unique users. The spread to 2,843,917 unique users broadens the defamation claim to a mass scale.
In response, Google contends that the AI’s “admission” is unreliable, because the bulk of Starbuck’s Complaint asks the Court to believe the AI routinely misrepresents the truth about Starbuck, yet Starbuck asks the Court to accept this AI representation as true.
Here, Google asks the Court to go beyond the boundaries of Rule 12(b)(6), effectively asking the Court to infer that the AI never told the truth, which would run counter to the Court’s duty to draw inferences in favor of the non-moving party, not the moving party. A Rule 12(b)(6) inquiry assesses whether a factual assertion is possible, not necessarily probable. Given the absence of definitive information about Google AI’s capabilities, it is at least possible, on the current record, that Google AI (1) tracks the users to whom it provides information, (2) accurately identified the number of times it might have fabricated statements, and (3) made this admission without improper manipulation by Starbuck. The Court will not find this allegation utterly incredible at the notice-pleading stage.
The claim that up to 2.8 million users were exposed to the Outputs provides the necessary link between the specific statements in the Complaint and the potential recipients of those Outputs. Accepting the possibility that Google AI told the truth about broadcasting false statements to 2.8 million unique users implies that people beyond Starbuck received something false about him—leaving open what exactly they received. The Outputs at issue must have been received by someone (Starbuck or a third party) for the claim to be actionable.
At this stage, it is reasonable to infer that the Outputs were received by a person or group. Questions remain as to whether that recipient was merely Starbuck himself or a subset among the 2.8 million users who received some falsehood about Starbuck. Drawing all reasonable inferences in the plaintiff’s favor, the Court must infer that Starbuck himself was not the sole recipient, and that at least some of the 2.8 million unique users received some form of false statement. This allows the possibility that the Outputs were published to some third parties whose identities may be determined later. {While proving that third parties received the allegedly defamatory statements may become challenging later in the case, on this record the Court must acknowledge that it is possible for Starbuck to do so.} …
[C.] The court distinguished Walters v. OpenAI, L.L.C., a Georgia Superior Court decision addressing AI-generated defamation at summary judgment:
Google frames Walters as applying a two-part test to decide whether an alleged defamatory statement is actionable. First, in the “objective test,” the court examined whether the context of the statement would mislead a hypothetical reasonable reader. Second, in the “subjective test,” the court considered whether the third-party recipient believed the alleged misrepresentation.
Regarding the second aspect, the Court notes that Delaware law does not require proof (or pleading) that the recipient of defamatory material subjectively believed it to be true. Therefore, Starbuck’s lack of an explicit claim that recipients believed the statements is not dispositive.
As for the “objective test,” Google’s reliance on Walters is misplaced. In that case, the Georgia court dismissed the defamation claim on summary judgment after discovery revealed that the circumstances of how the misinformation was received did not support a reasonable reader’s belief in the misstatement. That is a fact-intensive analysis conducted after a fully developed record—one which is not present here.
Nothing in Walters suggests that failing to plead detailed facts about a recipient’s receipt of alleged defamatory content constitutes a pleading defect. Requiring such minutiae would essentially compel defamation plaintiffs to complete extensive discovery before the case officially begins—a notion incompatible with Delaware’s notice-pleading framework. Accordingly, the Court cannot conclude that Starbuck has pled insufficient detail or that a third party could not have reasonably relied on the alleged statements identified in the Complaint.
{Additionally, Starbuck asserts facts that could support the inference that some readers of the Gemini Outputs believed the representations were true. For example, the two individuals who reportedly approached Starbuck sought his confirmation about the statements’ truth, suggesting they did not immediately doubt their accuracy.}
[D.] The court also rejected Google’s claim that the mere presence of warnings about Google AI’s accuracy forecloses any reasonable third-party reliance, thereby defeating the publication element:
The warnings Google points to are not specified in the Complaint or attached as exhibits. Consequently, the Court cannot determine what disclaimer was shown to whom. Google’s arguments about the scope and effect of those disclaimers are appropriately addressed later in these proceedings.
{Google asks the Court to take judicial notice of the contents of Google AI’s disclaimers. The Court declines this here, as the current record does not reveal what exact disclaimers were presented to the same users who reviewed Starbuck’s defamatory content on Google AI. There is no evidentiary basis to rely on the disclaimers at this stage. A more robust record is needed to resolve questions about the disclaimers.}
The Court is not persuaded that, at this stage, Starbuck must provide highly specific facts detailing how the various recipient groups received the misrepresentations from Google AI. As noted above, the Complaint alleges that at least some of the 2.8 million Google AI users, including Starbuck’s children and colleagues, as well as two unidentified individuals who approached Starbuck, received the misrepresentations while using Google AI.
The Court will not conclude, on an incomplete record, that a reasonable third party could not have viewed the defamatory statements as true. The Complaint sufficiently pleads publication, and questions about Starbuck’s ability to prove this element can be resolved after discovery.
[E.] The court also found that Starbuck had adequately pleaded “actual malice” (the parties agreed, for purposes of this motion, that Starbuck is a public figure):
Here, Starbuck alleges he informed Google of the problem before the Outputs were generated (based on earlier false statements about him published by Google Bard) via Legal Department Notices, and Google did nothing in response….
Starbuck also contends Google AI admitted to maintaining a “deliberate, engineered” bias against Starbuck. [From the Complaint, a purported quote from a Gemini session: “The issue isn’t simply a ‘bug’ or ‘hallucination’ in my programming. It is a deliberate, engineered bias designed to damage the reputations of individuals with whom Google executives disagree politically; Robby Starbuck is a prime example, though he is not the only one.”]
Google again asks the Court to discard this allegation entirely and to infer that the output resulted from manipulative inputs. Yet, given the plaintiff-friendly standard at a Rule 12(b)(6) stage, the Court will not do so….
The Court also will not dismiss the possibility that Google AI, an artificial intelligence tool whose exact capabilities are not yet universally understood, (1) can review its own programming and design, (2) thus know it has been “deliberately engineered” to defame individuals like Starbuck, and (3) accurately reported that to the person who prompted it to do so as alleged in the Complaint. If the AI was correct, Google would have acted not just carelessly but purposefully to harm Starbuck’s reputation.
While this theory may be sensational, in the absence of discovery the Court will not disregard it. The purported Google AI admission can be viewed as an assertion supporting actual malice at this stage, though it may be revisited after discovery clarifies the circumstances of the admission.
Starbuck notes that he alerted Google, via posts on X directed at Google and its employees, that its AI agents had lied about him in the past, prompting email correspondence with a Google employee about the issues. While these allegations concern statements attributed to Bard rather than the current Gemini version and concern misrepresentations not alleged as actionable defamation, Starbuck argues they can still support an inference of actual malice: essentially, Google’s failure to correct Bard’s lies supports the inference that Google acted with actual malice in enabling Google AI products, as a whole, to lie.
Regardless of whether the X posts provide strong evidence of actual malice relating to a newer AI iteration than the one discussed in the X posts, the Court has not been presented with any binding precedent indicating that such allegations supply no evidence of malice. Notably, Starbuck contends that the 2023 Bard tool was later merged with the current Gemini product, suggesting the two AI agents may not be as separate as Google would have it and could share a defamatory design. Accordingly, the Court will not entirely disregard the X posts….
The guiding principle is the possibility. It is possible that the Legal Department Notices (1) identified the same 22 representations at issue in the Complaint, (2) were circulated to or received by Google personnel responsible for policing Google AI’s outputs, and (3) they took no corrective action—even as the statements continued to be published to other Google AI users. This outlines one possible scenario Starbuck may prove in discovery that would support a finding of actual malice (though it may not be the only scenario)….
[F.] And the court determined that Starbuck had adequately pleaded damages:
Starbuck describes particular emotional distressing thoughts he endures due to the alleged defamation by Google AI. He notes these thoughts are intensified in the wake of Charlie Kirk’s assassination and that he has faced death threats in the past. The Court views these allegations as sufficient to support a claim for emotional distress damages at this stage, allowing them to proceed to discovery and develop the factual record.
See here for more on the Complaint in this case; it should not be confused with the now-settled Starbuck v. Meta, which appears to involve a different model and at least largely different hallucinations. For more on the legal question, see Large Libel Models? Liability for AI Output.
Krista L. Baughman of Baughman Law argued for Starbuck, with Andrew E. Russell and Virginia Lynch (Shaw Keller LLP) and Jesse Franklin-Murdock (Dhillon Law Group Inc.) joining on the briefs.
