Summary: On March 13, 2026, Encyclopedia Britannica and Merriam-Webster sued OpenAI in Manhattan federal court over nearly 100,000 articles and entries allegedly copied without permission to train GPT models. The complaint centers on memorization, showing model outputs set against source text and claiming GPT-4 can produce near-verbatim copies. The publishers also allege AI answers substitute for their pages and cannibalize web traffic.
Manhattan federal court gained a fresh copyright fight on Friday, March 13, 2026, when Encyclopedia Britannica and its subsidiary Merriam-Webster sued OpenAI; Reuters broke the news on March 16. According to the complaint, OpenAI copied nearly 100,000 Britannica articles and Merriam-Webster entries without permission and used them to train the GPT family of large language models; the publishers press both copyright and trademark infringement claims.
What earns this case a place on the watchlist is its evidence strategy. The publishers lead with “memorization” — the assertion that GPT-4 will return Britannica text essentially word for word when asked — rather than confining the dispute to training data. Any team building models on web content now faces both legal fronts in court at once: whether the training infringed, and whether the outputs do.
The Claims, in Three Layers
Read the Reuters and The Verge coverage together and the case stacks three claims: nearly 100,000 articles copied at scale to serve as training corpus; model outputs the suit labels “substantially similar” to the publishers’ material; and injury to Merriam-Webster’s trademark interests on top of the copyright counts.
The strongest exhibits are laid out side by side — OpenAI model responses set against Britannica source text, with passages matching nearly word for word. As the complaint puts it, “GPT-4 itself has ‘memorized’ much of Britannica’s copyrighted content and will output near-verbatim copies” when asked.
Memorization Becomes the Main Battleground
AI companies have long answered copyright suits with the same standing line: “Models learn statistical patterns and style, they don’t store copies.” Britannica’s filing runs straight at it. When a prompt alone makes a model hand back source text, that output is itself a substitute copy of the original, and the fair-use defense loses much of its footing.
The New York Times trod this same evidentiary path when it sued OpenAI in December 2023. Britannica now carries the fight onto structured reference works — encyclopedias and dictionaries — and so puts a wider question to the test: do the selection, arrangement, and prose of entries and definitions merit copyright protection at all?
Engineering teams should read the exhibits as a practical warning: memorization can be tested. Prompt a model with the opening words of a proprietary document and see whether it continues verbatim — a cheap, automatable check, and the same idea behind canary tests for training-data leakage. If plaintiffs’ counsel can surface word-for-word matches in federal court, nothing stops your own evaluations from surfacing them too.
Cannibalized Traffic and a Wave of Publisher Suits
On business harm the complaint does not hedge: AI answers “substitute, or directly compete” with Britannica’s content instead of pointing users back to the publisher’s own pages, as a conventional search result would — “cannibalizing” their web traffic, in the plaintiffs’ words. For a knowledge publisher that lives on subscriptions and visits, this is framed as a threat to survival, not a marginal loss.
Industry-wide, the choice is narrowing to license or litigate. The Times case has dragged past the two-year mark with no resolution. In September 2025, Anthropic paid $1.5 billion to settle the authors’ class action over books used in training — “license before you train” pressure converted into a concrete dollar figure. By filing when it did, Britannica carries that license-or-litigate choice out of news publishing and across knowledge publishing as a whole.
Takeaways for Developers and Product Teams
Three concrete effects. RAG and citation design climb the priority list — retrieve-then-cite carries far less legal risk than closed-book generation. Training-data governance turns into a compliance function: write down your data sources, licensing status, and dedup strategy, and fold output-side memorization checks into the evaluation pipeline. The content-licensing market may widen as well — if Britannica wins outright or settles big, licensing prices for encyclopedias, dictionaries, and database content all get recalculated. The contractual side matters too: teams building on third-party model APIs should pull output-side copyright risk into vendor review, weighing what a provider discloses about training data and indemnification at least as heavily as a benchmark delta.
For OpenAI, the timing stings — GPT-5.4 mini and nano launched that very week. The more capable the model and the smoother its output, the easier it becomes to elicit “reproduce the source on demand” evidence, and output-side risk climbs right along with it.