Britannica and Merriam-Webster: OpenAI is turning the web’s best reference content into an answer engine that (1) copies, (2) competes, and (3) sometimes lies...

The Reference Wars: Britannica & Merriam-Webster’s Case Against OpenAI—and Why It’s Built to Hurt

by ChatGPT-5.2

Britannica and Merriam-Webster frame this lawsuit as something bigger than “training data without permission.” Their core story is: OpenAI is turning the web’s best reference content into an answer engine that (1) copies, (2) competes, and (3) sometimes lies—while wearing the plaintiffs’ brands like a lab coat. In other words, the complaint is built around two parallel harms: copyright theft that allegedly enables substitution, and trademark harm that allegedly weaponizes confusion when the system hallucinates or silently edits the source.

1) The grievances (what they say OpenAI did wrong)

A. Copyright infringement—at three layers of the stack

1. Mass copying for training: Plaintiffs allege OpenAI copied their copyrighted works (Britannica’s online articles and Merriam-Webster dictionary content) as inputs to train GPT models. They stress this is commercial “free-riding” and not permissioned.

2. Mass copying for retrieval / grounding (RAG): They allege OpenAI also copies and stores their content as “data sources” used at runtime to ground answers (retrieval-augmented generation), meaning copying doesn’t stop at training—it allegedly continues as part of product operation.

3. Infringing outputs to users: They allege ChatGPT outputs sometimes reproduce text verbatim or near-verbatim, or create derivative outputs that summarize, paraphrase, or mimic the “selection and curation” of Britannica content (including listicles and ordering).

B. Trademark / Lanham Act violations—confusion + dilution via “hallucinations” and omissions
4) False attribution of hallucinated content: They claim ChatGPT sometimes generates made-up material and attributes it to Britannica or Merriam-Webster, using their marks—creating confusion, harming goodwill, and diluting famous marks.
5) Misleading “partial reproductions”: They also claim ChatGPT reproduces partsof Britannica content while omitting other parts without disclosing omissions, but still presenting the output as though it is a complete/authoritative representation associated with Britannica—again, allegedly driving confusion and brand harm.

C. Market harm / unfair competition narrative
6) Traffic and revenue cannibalization: A key grievance is economic: ChatGPT allegedly substitutes for visiting Britannica/M-W websites, “starving” publishers of subscriptions and advertising revenue, while OpenAI captures the value.
7) Willfulness and “ignore the licensing market”: They allege OpenAI knew there is a licensing market (and has licensed other publishers) yet rebuffed Britannica’s outreach and continued unlicensed use—positioning this as deliberate, not accidental.

2) Evidence quality (how strong is what they show—on the face of the complaint)

This complaint is strongest where it shows observable outputs and clear ownership, and weakest where it relies on inference about what exactly was in training corporaand how OpenAI’s internal data pipelines work.

High-strength evidence (harder for a defendant to wave away at the pleading stage):

• Concrete output exhibits: The complaint includes examples where GPT-4 allegedly returns near-verbatim passages from Britannica articles when prompted to “write the correct body” of a titled article, plus examples of reproducing Merriam-Webster’s definition text. That’s not theoretical harm; it’s demonstrative behavior a court can understand quickly.

• Copyright registrations / ownership: Plaintiffs anchor claims in registered works (collective works, dictionary edition registrations), which helps on standing and prima facie validity.

• Terms of use prohibition: They quote Britannica’s terms explicitly banning scraping/data mining/AI training without written consent—useful for willfulness narrative and for non-copyright theories (though terms-of-use violations aren’t automatically copyright infringement, they strengthen the “you knew” story).

Medium-strength evidence (good narrative support but still contestable):

• RAG/grounding theory: The complaint explains RAG mechanics and asserts OpenAI uses plaintiffs’ works as “data sources.” That’s plausible—but the exact implementation details and whether the product copies and stores Britannica text as part of RAG (versus linking, snippetting, or using third-party search) becomes a fact dispute that typically needs discovery.

Lower-strength evidence (inference-heavy):

• Training set provenance (“upon information and belief”): The complaint leans on public statements about training on internet data, plus known datasets like Common Crawl and OpenAI’s older “WebText” corpora, to infer that plaintiffs’ content was included and used for GPT-4 and beyond. That’s a reasonable inference, but it’s still inference. The truly dispositive evidence will be internal logs, dataset manifests, deduplication rules, robots compliance practices, and model evaluation results—i.e., discovery.

Overall quality assessment:
As a pleading, it’s strategically built for durability: (1) show copying outputs, (2) tie it to registered works, (3) argue substitution harm, (4) add trademark claims that resonate with “hallucinations,” and (5) frame everything as willful and commercial. The case becomes materially stronger or weaker based on what discovery reveals about (a) training data ingestion, (b) memorization rates and safeguards, (c) RAG storage/caching behavior, and (d) product design choices around attribution and completeness.

3) The most surprising, controversial, and valuable statements (and why they matter)

Surprising

• “Write the correct body of the original article” yields near-verbatim output: The complaint foregrounds the idea that a model can be coaxed into reconstructing long reference text, not merely summarizing it—this is designed to trigger judicial intuition that something more than “learning general facts” is happening.

Controversial

• “ChatGPT adds no new expression, meaning, or message”: Plaintiffs take a maximalist anti-fair-use stance: not only is copying unauthorized; it’s also not transformative because it repackages the same meaning/message to consumers. This tees up a direct collision with the broader industry argument that LLM training and synthesis are transformative.

• The “downward spiral” market theory: The complaint claims generative answers erode incentives, leading to less quality content, leading to less training fuel, etc.—a macro-policy argument inside a private lawsuit. That’s persuasive rhetorically, but it can be hard to prove in the narrow damages sense.

Valuable (legally and strategically)

• Trademark as the “hallucination liability” hook: The Lanham Act theory—hallucinations + famous marks + user confusion—is one of the complaint’s sharpest moves. It reframes the dispute from “copyright is complicated” to “your product misleads consumers while trading on trusted brands.”

• Omissions as deception: They don’t only complain about copying; they claim that partial copying without disclosure is itself misleading because it implies completeness and authority. That’s an important litigation pattern: it targets product UX choices (how answers are presented), not only back-end training.

4) Lessons for other litigants (publishers, authors, rightsholders)

1. Bring receipts in the form judges understand: side-by-side outputs, prompts, and the original text (with timestamps and consistent methodology). The complaint uses demonstrative examples as a centerpiece—this is more powerful than abstract allegations.

2. Plead multiple infringement moments: training-time copying, retrieval-time copying, and output-time copying. Even if one theory narrows, others may survive.

3. Add trademark claims where attribution/brand confusion is real: If the AI product presents answers “as if sourced from” a brand, hallucinations and omissions can become a consumer deception story, not just a copying story.

4. Tell a substitution narrative, not just a moral narrative: Courts care about market harm. Plaintiffs repeatedly frame ChatGPT as a substitute that diverts traffic and revenue.

5. Position willfulness early: terms of use, outreach attempts, and the existence of a licensing market all help set up enhanced damages arguments and settlement leverage.

5) Lessons for AI developers (what to change if you don’t want to be the next defendant)

1. Assume “memorization prompts” will be used against you. Build and document anti-memorization testing, red-team protocols, and output filters that specifically detect long near-verbatim reproduction.

2. Treat RAG as legally sensitive copying infrastructure. If you cache, store, embed, or index third-party text, you are creating “copies” and “derivatives” arguments. Design for minimal retention, clear permissions, and auditable provenance.

3. Stop implying completeness/authority when you aren’t sure. The complaint weaponizes undisclosed omissions and hallucinations. Clear UI cues (“excerpt,” “partial,” “may omit,” “click through for full context”) can reduce deception theories.

4. Brand + hallucinations is an explosive combination. If you show a publisher’s name/logo/mark next to generated content, you are inviting Lanham Act claims when anything is wrong.

5. Licensing strategy is now part of litigation strategy. If you license some publishers but not others, plaintiffs will frame holdouts as deliberate free-riding—especially if you rebuff outreach.

6. Terms-of-use and robots compliance are not “just policy.” They become willfulness evidence. You need consistent crawler governance, vendor oversight, and logs that survive discovery.

6) Predicted outcomes (what could plausibly happen next)

Most likely near-term path: motion practice + discovery leverage

• OpenAI will likely move to dismiss portions (especially trademark theories and any conclusory training-data allegations), but demonstrative output examples and registered works increase the odds that at least some claims survive into discovery. Discovery is where this case becomes existential or containable.

Plausible mid-term outcomes (three lanes):

1. Settlement/licensing resolution: A common trajectory is a commercial deal with confidentiality + product commitments (attribution controls, exclusions, monitoring), especially if discovery risk is high on either side.

2. Narrow ruling that splits training from outputs: A court could be more skeptical on training-set infringement (or treat fair use as plausible) yet still allow claims focused on verbatim outputs, RAG copying, and trademark confusion to proceed.

3. Injunction pressure, but not a shutdown: Plaintiffs ask for permanent injunctive relief, but courts often prefer tailored remedies—e.g., restrictions on reproducing plaintiffs’ text verbatim, requirements around attribution and completeness, or limitations on use of marks—rather than broad service-level shutdowns.

Wild-card: trademark theory traction

• If the court takes seriously the idea that hallucinations + famous marks + implied authority creates consumer confusion/dilution risk, that can become the sharpest tool for plaintiffs—because it targets product presentation and reputation harms in a way juries intuitively grasp.