The AI Company Is the Evidence Now: What Britannica v. Perplexity Teaches About RAG and UAL Discovery

When Your AI Vendor's Retrieval Index Becomes the Evidence, Your Own Stack Needs to Survive That Same Request

A Southern District of New York court has ordered Perplexity AI to produce snapshots of its own retrieval-augmented-generation (RAG) database and months of its User Activity Log (UAL) data as evidence in a copyright case brought by Britannica and Merriam-Webster. The ruling, covered by ACEDS on September 10, 2026 in a piece republished from the Minerva26 Blog (originally posted September 9, 2026), treats Perplexity's own infrastructure — not a chat transcript a party typed in, not a review tool counsel used — as the discoverable record of alleged infringement. Britannica and Merriam-Webster allege Perplexity's "answer engine" copies their content into its retrieval index at the input stage and reproduces or paraphrases it at output, plus a trademark claim over hallucinated content falsely attributed to their brands.

What did the court actually order Perplexity to hand over?

Snapshots of the RAG retrieval index plus months of UAL records showing what users asked, how content was ranked, and what the model returned.

ACEDS breaks the UAL down into four logged elements per query: the user's question, how the system retrieved and ranked source material, the instructions sent to the underlying language model, and the final answer served. That's the granular, per-query audit trail that turned a "what did the AI say" dispute into a "show us the pipeline" discovery order.

Why should this matter if you're not litigating against Perplexity or OpenAI?

Because any AI tool a party in your case relies on now has a retrieval database and an activity log that opposing counsel can demand.

ACEDS is explicit that this isn't a niche AI-industry problem: "you're going to need this vocabulary" once you're dealing with any party — vendor, client, or opponent — running a RAG-based tool, however small the company. The order is effectively a discovery template: ask what's in the index, ask what's logged, ask where the record of the system's own behavior actually lives.

What should buyers ask their own e-discovery or AI vendor now?

Whether their platform can produce a defensible, source-linked snapshot of its retrieval index and query logs on demand, under their own custody.

The coverage focuses on Perplexity as a defendant being forced to expose infrastructure it presumably didn't build for discovery. Firms deploying RAG tools on privileged material should flip that scenario: if a regulator or opposing party later demanded proof of exactly what your platform retrieved and why for a given answer, could you produce it without also exposing other clients' data or sending anything outside your own environment? Architecture — single-tenant, self-hosted under your own keys, with every answer already tied back to a cited source document — is what makes that a non-event instead of an emergency motion.

How is this different from the OpenAI copyright fight?

ACEDS flags a timing distinction: the dispute turns on when Perplexity accesses the allegedly copyrighted material, not just whether it does.

The full mechanics weren't detailed in what's public so far, but the framing matters for discovery scope: input-stage retrieval and output-stage generation may be logged, stored, and discoverable differently, which is exactly the kind of distinction a requesting party needs to nail down before drafting a request for production.

Frequently asked questions

Is Perplexity's user activity log the same as a chat history?

No. UAL records the system's internal retrieval and ranking decisions and the instructions sent to the model, not just the visible question and answer.

Does this ruling apply outside copyright cases?

The order is specific to this case, but the discovery vocabulary — RAG index, UAL logs — applies anywhere a party's AI tool's own behavior is in dispute.

Sources: ACEDS (republished from the Minerva26 Blog).