A chatbot’s account of its own conduct helped a defamation claim against Google survive a motion to dismiss. In Starbuck v. Google LLC, the Delaware Superior Court allowed allegations about a generated audience figure and deliberate design to support the claim at pleading. Its July 24, 2026 opinion left the truth of those accounts for later examination. The question for an AI governance team is what evidence could test them.

The model’s account is only one part of that inquiry. The court also considered earlier warnings about Bard, later notices to Google’s legal department, and allegedly invented sources. A record could undermine the model’s story about its audience or design while preserving evidence for another part of the claim. Its significance depends on which proposition it tests.

What survived dismissal

Robert Starbuck alleged that Google’s AI products generated false statements associating him with serious misconduct, sometimes supported by invented sources. His original complaint also alleged that other people received false statements and that Google had received notice before the outputs for which he sought relief. These were allegations the court assessed under Delaware’s notice-pleading standard. The opinion expressly declined to endorse their truth.

Google assumed for its motion that an AI output could be defamatory and challenged publication, actual malice, and damages. The court found the allegations sufficient on each issue. It applied reasonable conceivability, a standard the opinion distinguished from federal plausibility pleading. It found the damages allegations sufficient under Tennessee law and noted that Google made no argument that they failed under Delaware law, so no choice between them was needed. The ruling did not establish liability or resolve the general legal treatment of AI-generated defamation.

The audience figure was an allegation too

The publication dispute concerned whether other people received the actionable statements. Starbuck alleged that his children and colleagues received false outputs and described two encounters with people who had seen information about him from Google AI. He also alleged that Gemini reported distributing false information about him to 2,843,917 unique users. That precise figure was a generated statement recounted in the complaint, not an independently verified audience measurement.

The court gave the figure a particular role. It linked the broader allegations about recipients to an inference that some people had received the specific outputs challenged in the complaint. Google argued that a system accused of unreliability should not be believed about its own audience. The court declined to draw that inference for Google at dismissal, while recognizing that proof could be difficult later.

The resulting evidence question is narrower than the apparent scale of the number. What would establish that someone other than Starbuck received an actionable statement, and in what circumstances? A prompt and response record might establish what appeared in one interaction. At pleading, the court rejected a demand for detailed recipient-by-recipient circumstances, drawing an analogy to mass publication. Later proof would still need to establish publication, but the ruling did not require a named witness for every alleged recipient. The generated audience figure does not settle how that proof could be made.

Notice has to reach people

The same distinction matters more sharply for actual malice. The parties treated Starbuck as a public figure for the motion. The court explained that he would ultimately have to prove knowledge of falsity or reckless disregard for truth by clear and convincing evidence. For an institutional defendant, the relevant state of mind must be connected to the people responsible for publication. A model’s apparent hostility is not, by itself, a finding about those people’s knowledge.

The complaint alleged that Gemini described a deliberate design to damage Starbuck’s reputation. The court would not rule out, at pleading, the possibility that the system could accurately describe its programming. It allowed the allegation to support discovery without finding that the model had that capability or that anyone had engineered the claimed bias. Establishing a generated statement’s technical basis therefore remains distinct from establishing its attribution to responsible people.

Earlier contacts supplied a further basis for the court’s reasoning. Starbuck alleged X posts in 2023 and 2024 directed to Google and its employees, followed by correspondence with a Google employee about Bard’s false statements. The court declined to disregard those allegations, emphasizing the pleaded combination of Bard with Gemini and the possibility of a shared defamatory design.

The opinion also identified a route through human notice. Starbuck alleged written notices to Google’s legal department on July 31 and August 12, 2025, before the 22 outputs identified as actionable in the original complaint. The court did not have those notices and did not know their exact contents. It nevertheless considered it possible that they identified the relevant falsehoods, reached people responsible for the outputs, and were followed by continued publication despite an ability to correct the problem. Each step remained something to establish.

This reasoning gives complaint handling evidentiary significance. The question is not settled by showing that a message entered a legal inbox. The content of the notice, its route within the organization, the responsibilities of its recipients, their knowledge, and their response bear on the possible sequence the court described. The opinion did not decide that everyone in a legal department is responsible for publication, or that receipt of a complaint alone establishes actual malice.

The court identified another allegation supporting actual malice: Google AI’s alleged fabrication of sources. It noted that the parties had not explicitly addressed that ground in their arguments on the issue. The opinion treated fabrication as potentially relevant evidence of actual malice, without finding that fabrication occurred or that responsible people had the required state of mind.

A real objection survives the ruling

Google’s reliability objection exposes a serious difficulty. If a plaintiff can rely on a system’s generated accusation against its maker while alleging that the same system invents facts, an unsupported story could help impose the burden of discovery. The concern becomes especially acute when the story describes internal design that the person prompting the system cannot otherwise verify. Accepting that concern does not establish that the prompts here were manipulative; the court refused Google’s invitation to assume they were.

The July opinion resolved the pleading question in Starbuck’s favor. It did not solve the later problem of proving that the generated account was accurate. Evidence that the system lacked access to audience records or design information could weaken its account of those matters. Compare that with a preserved response citing a source that an independent check establishes never existed: the response could be an unreliable account of the world and reliable evidence of what the system produced.

That distinction matters to the fabrication ground the court identified. Disproving the model’s self-description would not erase an invented source recorded in its output, though proving fabrication would still leave the institutional state-of-mind question. The earlier Bard contacts raise a different question about the connection between products and what responsible people knew across that change. The later legal notices require their own connection to contents, recipients, and subsequent publications. A technical rebuttal to one generated account therefore need not answer the other grounds the court considered.

A warning needs its context

Google also invoked accuracy disclaimers against publication and actual malice. The court lacked a record showing the wording, prominence, timing, and presentation of the relevant warnings to the relevant users. It declined to decide their effect on that basis. The opinion’s discussion of an earlier Georgia defamation decision identified both a different procedural posture, after discovery, and a different theory: general awareness that a system could produce false information differed from allegations of targeted design and specific prior notice.

For a product team, the practical inference is to connect a warning to the experience in which it appeared. A current screenshot of a generic disclaimer would not establish what an earlier recipient saw. A versioned interface record, linked where feasible to the relevant interaction, could help answer that question. It would not establish that a warning defeats every theory of liability. The July ruling left the warning’s effect open rather than holding that disclaimers are useless.

The record may cut both ways

A blue thread runs from an envelope through three blank cards to an open folder.

AI-generated editorial illustration · Following the notice.

For the legal and product teams handling a report of false person-level outputs, a useful record would connect the reported statement and interaction to its cited sources, checks on those sources, the notice, the people assigned to investigate, the findings, the action taken, and any follow-up check. For a report spanning products or versions, identify which system was involved and what connects the earlier report to the later output. Record uncertainty as well: a report may omit the prompt, misidentify a product, or describe an output that investigators cannot reproduce. Those are investigative possibilities, not findings about Starbuck’s reports. The purpose is to make the basis of the response examinable.

Such a record could rebut the model’s story while revealing facts relevant to the notice theory. An unsuccessful correction might document a reasoned attempt to address a report; it might also identify the point at which responsible people learned that a problem persisted. Neither conclusion follows from the existence of a ticket alone. What was known, by whom, and before which publication matters to the actual-malice inquiry the court preserved.

Confidence is high in the limited reading of the July ruling: it allowed these allegations to proceed without validating the model’s account or finding actual malice. Confidence about the merits would require evidence this analysis does not have. A later ruling rejecting the pleading inferences, or evidence about recipients, notice, capabilities, and responsible people, could change the application of the argument. The proposed records are a way to test those questions, not a promise of protection from suit or a duty to retain every conversation indefinitely.

Practical implications

  1. Separate the propositions. Legal teams should distinguish what an output said, whether its cited sources exist, whether another person received it, and what responsible people knew. A generated audience figure or design explanation should not collapse those questions.

  2. Connect notice to a response. Product and legal teams should record the reported falsehood, relevant dates, assigned responsibility, investigation, response, and remaining uncertainty. An inbox receipt alone does not establish what happened next.

  3. Preserve the warning’s context. Product teams should retain the applicable wording and interface version and identify what can be established about presentation in the relevant interaction. Describe evidentiary limits where those links cannot be reconstructed.

  4. Set a proportionate evidence scope. Counsel and system owners should identify the records needed to investigate and test the particular claim, then address access, retention, and applicable preservation obligations. The opinion does not supply a universal logging mandate or make documentation a defense in itself.

Scope and sources

This analysis examines the Delaware Superior Court’s July 24, 2026 ruling in Starbuck v. Google LLC, drawing on the opinion and original complaint. It focuses on the court’s reasoning at the motion-to-dismiss stage and describes the parties’ positions through the opinion. Subsequent pleadings and proceedings fall outside its scope.

Case citations

Starbuck v. Google LLC, C.A. No. N25C-10-211 MAA (Del. Super. Ct. July 24, 2026), especially pages 14–35. Opinion. Original complaint, filed October 22, 2025, especially paragraphs 24, 36–43, 77–86, 91–100, 106, and 110–111.

Disclaimer: AI Lex Intelligence is published for informational purposes only. It does not constitute legal advice, and no attorney-client relationship is formed by reading or receiving this publication. Readers should consult qualified legal counsel about specific legal matters.

AI Lex Intelligence with Zola Valashiya. Independent legal analysis of artificial intelligence, law, and the institutions that govern both.