On Friday, March 13, 2026, Encyclopedia Britannica and its subsidiary Merriam-Webster filed suit against OpenAI in Manhattan federal court; Reuters broke the story on March 16. The complaint alleges that OpenAI copied nearly 100,000 Britannica articles and Merriam-Webster entries without permission to train the GPT family of large language models, and asserts both copyright and trademark infringement.
The evidence strategy is what makes this case worth watching. Rather than resting on the training-data question alone, the publishers put “memorization” front and center — the claim that GPT-4 will reproduce near-verbatim copies of Britannica text when asked. For every team training models on web content, that puts two separate legal fronts, whether training infringes and whether outputs do, into the courtroom at the same time.
What the Complaint Alleges
Drawing on Reuters and The Verge’s reporting, the claims stack in three layers. First, OpenAI copied nearly 100,000 articles as training corpus at scale. Second, the models produce results the suit calls “substantially similar” to the publishers’ material. Third, Merriam-Webster’s trademark interests were harmed alongside the copyright claims.
The sharpest material in the filing is a set of side-by-side exhibits: OpenAI model responses placed next to Britannica text, with passages that appear to match word for word. The complaint states that “GPT-4 itself has ‘memorized’ much of Britannica’s copyrighted content and will output near-verbatim copies” on demand.
Memorization Evidence Takes Center Stage
“Models learn statistical patterns and style, they don’t store copies” has been the standing defense from AI companies. Britannica’s complaint attacks it directly: once a model can reproduce source text on prompt, the fair-use argument narrows considerably, because the output itself functions as a substituting copy.
The New York Times, which sued OpenAI in December 2023, built its case on the same evidentiary path. Britannica extends the front line to structured reference works — encyclopedias and dictionaries — which tests a broader question: whether the selection, arrangement, and prose of entries and definitions earns copyright protection in the first place.
For engineering teams, the exhibits double as a practical warning: memorization is testable. Feeding a model the opening words of a proprietary document and checking whether it continues verbatim is a cheap, automatable check — the same idea behind canary tests for training-data leakage. If a plaintiff’s lawyers can produce word-for-word matches in federal court, your internal evaluations can produce them too.
Traffic Cannibalization and the Publisher Litigation Wave
The business damage in the complaint is described plainly: AI answers “substitute, or directly compete” with Britannica’s content instead of sending users to the site the way a traditional search engine would — what the plaintiffs call “cannibalizing” their web traffic. For a knowledge publisher that monetizes subscriptions and visits, that is an existential framing, not a rounding error.
The sue-or-license squeeze is playing out across the industry. The Times case has run more than two years without resolution. In September 2025, Anthropic settled a class action from authors over books used in training for $1.5 billion, converting “license before you train” pressure into a real cash figure. By filing now, Britannica pushes the license-or-litigate choice from news publishing into knowledge publishing at large.
What It Means for Developers and Product Teams
Three practical effects. First, RAG and citation design move up the priority list: retrieve-then-cite is materially safer from a legal-risk standpoint than closed-book generation. Second, training-data governance becomes a compliance topic — document sources, licensing status, and dedup strategy, and add output-side memorization checks to your evaluation pipeline. Third, the content-licensing market may widen: a Britannica win or a large settlement would reprice licensing for encyclopedias, dictionaries, and database content across the board. A fourth consideration is contractual: if you build on third-party model APIs, output-side copyright risk now belongs in vendor review — what a provider says about training data and indemnification deserves at least as much weight as a benchmark delta.
The timing is awkward for OpenAI — the same week it launched GPT-5.4 mini and nano. The stronger the model and the more fluent the generation, the easier this kind of “reproduce the source on demand” evidence becomes to produce, and the higher the output-side risk climbs.
Sources
- Encyclopedia Britannica sues OpenAI over AI training — Reuters
- Encyclopedia Britannica is suing OpenAI for allegedly ‘memorizing’ its content — The Verge
- Britannica sues OpenAI for copyright and trademark infringement — Engadget
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.