Liability could be proved if records show that a generative AI system—a tool that creates new text from user input—delivered an identifiable, false accusation to another person. Liability could be defeated by missing publication evidence, credible warnings, or a lack of proof that the provider probably knew the output was false. The premise remains unproven in the leading dispute described here. On July 24, 2026, the Delaware Superior Court allowed Robby Starbuck's claim against Google to proceed into discovery, but it did not find that Google defamed him or that any accusation was true.
Table of Contents
- What would prove publication?
- The complete exchange matters
- Public figures face a higher proof burden
- When do warnings help the defense?
- What should a potential claimant preserve?
What would prove publication?
A defamatory statement generally must reach someone besides the person it concerns. A private exchange visible only to the claimant may therefore present a major publication problem. Useful evidence could include the user's exact query, the complete response, its date and time, product and model information, account records, and delivery logs. Recipient testimony could confirm what appeared, whether the recipient understood it as a factual accusation, and whether it referred to the claimant.
Starbuck alleged that Google's system connected him to sexual assault accusations and the January 6 riot. He also alleged that other people received such responses. At the dismissal stage, the court treated those statements as allegations, not proven facts. The Delaware Superior Court said unnamed recipients could suffice at the pleading stage, while warning that finding and deposing them could become a serious discovery problem. A large user count does not establish that any particular person received the disputed statement.
The complete exchange matters
A screenshot of one alarming sentence may not show enough. Investigators would need the full exchange to determine what the user asked, whether the system identified the claimant, and whether surrounding text framed the accusation as fact, uncertainty, fiction, or error.
The strongest record would preserve: The defense could use the same material. A full transcript might reveal that the claimant was not identifiable, the accusation never appeared, or no third party saw it. Metadata could also show that a different product or version generated the disputed text.
- The exact user query and every follow-up.
- The complete, unedited response.
- Any citations, links, warnings, or uncertainty notices.
- The date, time, account, product, and displayed model version.
- Evidence showing who received or viewed the response.
Public figures face a higher proof burden
A public-figure claimant must prove actual malice by clear and convincing evidence. In this context, actual malice concerns knowledge that the challenged statement was false or serious doubt about its truth—not simply hostility, carelessness, or awareness that generative systems sometimes make mistakes. Relevant company records could include prior reports of the same false association, internal testing, employee escalations, correction requests, design decisions, and remediation efforts. Repeated notice followed by continued distribution could support an inference about knowledge.
Records showing prompt investigation, effective correction, or no awareness of probable falsity could weaken that inference. A known general risk is not necessarily knowledge about one specific statement. The National Institute of Standards and Technology defines "confabulation" as confidently generated erroneous content and explains that it can arise from next-token prediction. That technical risk alone does not prove publication or actual malice.
When do warnings help the defense?
Warnings can matter, but their existence cannot simply be assumed. Evidence should show the warning's exact wording, size, placement, timing, and visibility to the person who received the accusation. A warning may help establish that a reasonable recipient should not treat the response as verified fact. Its force may be weaker if it was hidden, appeared after the accusation, or used vague language that did not address factual reliability.
The closest documented defense result is Walters v. OpenAI. As described by the Delaware Superior Court, the Georgia court granted summary judgment after discovery where the recipient had seen disclaimers and the plaintiff lacked evidence that anyone at the provider knew the received output would probably be false. That result does not make every disclaimer decisive. The Starbuck court declined to rely on asserted warnings without evidence showing how they appeared to the particular users involved.
What should a potential claimant preserve?
Preservation should begin before an output disappears, changes, or becomes difficult to authenticate. A claimant should avoid editing the original record or relying entirely on a recreated exchange, which may involve a different system version or produce different wording.
Practical steps include: A lawyer may need to seek provider-held logs before ordinary retention practices erase them. The most useful first package is the untouched exchange, its metadata, and an independent recipient who can testify that the accusation actually reached them.
- Save the complete exchange through screenshots and any available export function.
- Record the date, time, web address, account, product, and displayed version.
- Preserve every warning or disclaimer exactly as presented.
- Ask actual recipients to retain their own copies and account records.
- Record when each recipient saw the accusation and what they understood it to mean.
You Might Also Like
- Severe Turbulence Causing Unrestrained Passenger Injuries: What Evidence Could Prove—or Defeat—Liability?
- An AI Companion Allegedly Encouraging Self-Harm: What Evidence Could Prove—or Defeat—Liability?
- A Wrongful Arrest Caused by Faulty Facial Recognition: What Evidence Could Prove—or Defeat—Liability?