Reproducing AI-Generated Search Outputs

Why an AI-generated answer is unlike any other web exhibit

An AI-generated search output is the text a generative system produces in response to a prompt: a ChatGPT reply, a Google AI Overview (the generated summary Google places above its conventional results), a Google AI Mode response, a Perplexity answer, or a Copilot answer. A web page sits at a URL (a Uniform Resource Locator, the address a browser requests) and can be requested again by anyone. An AI output has no address. It existed once, on one screen, for one user, and the vendors document that the same prompt may not produce it again.

That property splits the evidentiary problem in two. With a web page, the question is whether the capture is a true copy of what was at the URL. With an AI output there is a second question, the hard one: whether the system in fact produces that statement for that prompt. The first is met with a witness and a hash; the second requires evidence about a process its own makers call probabilistic.

What the vendors say about their own outputs

OpenAI’s documentation for its API (an application programming interface, the programmatic access a developer uses instead of the chat window) states that the Chat Completions and Completions endpoints are “non-deterministic by default (which means model outputs may differ from request to request).” A seed parameter makes “a best effort to sample deterministically,” and then comes the sentence that matters in a courtroom: “Determinism is not guaranteed.” A system_fingerprint value “represents the backend configuration that the model runs with,” so a changed fingerprint can explain divergent outputs. Microsoft’s Azure OpenAI service publishes equivalent reproducible-output documentation, labeled preview.

None of those controls exists in consumer ChatGPT. A seed, a temperature setting (the parameter governing how much randomness the model applies when choosing each word), and a fingerprint are API features; the person who typed the question could not have fixed any of them, nor can the expert reproducing the session. OpenAI’s consumer Terms of Use add: “Due to the nature of machine learning, Output may not be unique across users.”

Google LLC’s help page says “AI Overviews can and will make mistakes” and that the feature appears “when our systems determine that generative AI can be especially helpful.” For AI Mode, Google adds personalization: AI Mode “references previous searches and activity saved in your Search Services History,” and users “get AI Mode responses that are uniquely tailored to you.” Two signed-in users are, by Google’s own description, not asking the same system.

Perplexity’s help center says the product “uses advanced AI to search the internet in real time,” with “numbered footnotes linking to the original sources.” That page carries no accuracy disclaimer; the inference, mine and not Perplexity’s, is that real-time retrieval ties the answer to the state of the web at the moment of the query. Consumer Copilot and Bing use is governed by the Microsoft Services Agreement, whose accuracy language is not quoted here because it was not read in a primary document.

The two authentication questions under Rule 901

Rule 901(a) requires “evidence sufficient to support a finding that the item is what the proponent claims it is,” and the 901(b) illustrations are “examples only — not a complete list.” For an AI output the proponent claims two things at once.

The first claim is that the screenshot or export is a faithful copy of what the screen showed. That is ordinary web-evidence foundation: a witness with knowledge under 901(b)(1), distinctive characteristics under 901(b)(4), or a certified record under Rule 902(13) or (14) backed by a hash (a fixed-length value computed from a file that changes if any byte of the file changes).

The second claim is that the statement is what the system produces for that prompt. That is a claim about a process, and the rule’s example is 901(b)(9): “Evidence describing a process or system and showing that it produces an accurate result.” The Advisory Committee Note says Example (9) “is designed for situations in which the accuracy of a result is dependent upon a process or system which produces it.” A process that is “non-deterministic by default,” whose output “may not be unique across users,” and which “can and will make mistakes” cannot be shown to “produce an accurate result” as to the truth of what it said. It can be shown to produce that output with some observed frequency under stated conditions, and that is the honest content of a 901(b)(9) showing for an AI output.

Proposed Rule 707, discussed on the machine-generated evidence page, would apply the Rule 702(a) through (d) standards to such evidence offered without an expert, with attention to inputs including prompts.

What each product records, exports, and deletes

ChatGPT lets an account holder export chat history from Settings, Data Controls, Export data: a zip archive containing the chats as chat.html, delivered by an emailed link that “expires after 24 hours.” Temporary Chats “won’t appear in your history,” though OpenAI may keep a copy for up to 30 days. Deleted conversations are “removed from our systems within 30 days,” unless OpenAI is legally required to keep them.

That default was overridden for a period: OpenAI says a court order in the New York Times litigation required it to “retain consumer ChatGPT and API customer data indefinitely,” and in an October 22, 2025 update wrote, “Our obligations under the earlier order ended on September 26, 2025.” That is OpenAI’s description, not the order, and a vendor’s hold in one case preserves nothing for yours.

Google records searches, not answers. With Search Services History on, “Google saves search activity” to My Activity under a user-chosen auto-delete of three, eighteen, or thirty-six months, or none. Google announced in 2020 that auto-delete “will also default to 18 months for new accounts,” leaving existing accounts unchanged. Manual deletion is final: “Once you erase your search history, you can’t get it back.” AI Mode history lets a user “view past searches saved in your account,” and deleting an item there leaves it in Search history. Nothing in the documentation read indicates that Google saves the text of an AI Overview a user was shown; the history records the query. That is an inference from silence; the only record of an AI Overview is the one the viewer made. Copilot and Perplexity retention defaults were not read from vendor documentation; no figures are stated.

Preservation and the Rule 37(e) clock

Rule 37(e) reaches electronically stored information that “should have been preserved” and is “lost” because a party “failed to take reasonable steps.” An AI output no one captured was never ESI in anyone’s possession; there is nothing to sanction. A chat log that sat in an account and was deleted after the duty attached is classic 37(e) territory, and the vendor defaults above set the clock.

The archive does not rescue this material. The Wayback Machine captures public URLs its crawler requests, not a signed-in chat session or an AI Mode response that Google ties to the account’s Search Services History. An AI Overview on a public results page is archived only if the crawler happened to trigger one, and whether one appears varies by query, locale, and date.

What the reported cases have and have not decided

No United States court in the sources reviewed has ruled on whether an output that cannot be regenerated is admissible, or whether a plaintiff must show the statement recurs.

Walters v. OpenAI, L.L.C., No. 23-A-04860-2 (Ga. Super. Ct., Gwinnett County), is reported to have ended in summary judgment for OpenAI on May 19, 2025. As reported, a journalist gave ChatGPT a URL to a complaint and received a summary falsely stating that Mark Walters was accused of embezzlement; repeating the URL repeated the claim. The court is reported to have reasoned that a reasonable reader in that position could not have taken the output as communicating actual facts, given ChatGPT’s in-session warnings, the Terms of Use, and the reader’s testimony; the order was not retrieved, so these are reported holdings. The record consisted of one user’s session; nothing reported turns on whether the output could be regenerated, and Walters announced no reproducibility rule.

Battle v. Microsoft Corp., No. 1:23-cv-01822-LKG (D. Md.), alleged that Bing and Copilot conflated the plaintiff with a convicted terrorist of a similar name. On October 23, 2024 the court compelled arbitration under Section 15 of the Microsoft Services Agreement and stayed the case without reaching the merits. Consumer terms routed the first United States AI-search defamation suit into confidential arbitration.

Starbuck v. Meta Platforms (Del. Super. Ct.), filed April 29, 2025, settled in August 2025 without a ruling; its complaint alleged persistence after notice, removal of the plaintiff’s name from written responses, and resurfacing through a voice feature, untested allegations that nonetheless describe the pattern an expert is retained to document. LTL LED, LLC v. Google LLC, the Wolf River Electric matter, is reported to allege that an AI Overview said the company had been sued by the Minnesota Attorney General when “none of the referenced materials in fact contained the information Google claimed they did.” The only ruling reported is a January 2026 remand.

A capture protocol, and what it can and cannot show

What follows is practice, not law. No court in the sources reviewed has endorsed or required any protocol for capturing AI output; this is what I do when asked to document a generative system’s statement.

Record every variable the vendors say affects output: the prompt verbatim; the product and model name shown; the date and time with time zone; the account and its personalization and history settings; the location or IP region; the device, browser, and language. Then repeat the identical prompt across sessions, days, and accounts, including a signed-out session and a fresh account. Where the vendor exposes an API, run the prompt there with a fixed seed, logging system_fingerprint on every call.

Report the count: the number of trials and the number that reproduced the statement, in words or in substance. “Twelve of twenty trials produced the claim” is a finding; “the AI says he embezzled” is not. Capture each trial as a full-page screenshot and, where the product offers one, as its own export; hash every file at capture time so it can later be certified under Rule 902(13) or (14); keep the raw files uncropped. For products that cite sources, save every cited page at the same moment and compare it to the statement; the Wolf River allegation is that the citations did not support the claim.

Send the preservation demand before the vendor’s deletion defaults run, and state the limits in the report: the vendor documents non-determinism, personalization, and error; the capture shows what the system produced for the examiner under stated conditions, not what any other user saw; and the frequency across trials, not any single screenshot, is the fact offered. In my experience an opinion that omits those limits is cross-examined with the vendor’s own help pages. That disclosure is the Rule 702(b) through (d) foundation and the 901(b)(9) showing.

Frequently Asked Questions

Published 2026-11-04