The Britannica v. OpenAI dispute could signal a shift of AI liability from training inputs to branded outputs, and the trademark theory may be harder to solve than copyright.

From training to outputs: the next frontier?

Last year, Britannica and Merriam-Webster sued Perplexity for AI-generated hallucinations dressed up as authoritative citations (see Citation frustration: When AI makes stuff up (and gets sued for It). Six months later, the same plaintiffs have filed suit against OpenAI, the maker of ChatGPT featuring a Lanham Act claim alleging that ChatGPT’s hallucinations and undisclosed omissions, displayed alongside Britannica’s and Merriam-Webster’s famous marks, constitute false designation of origin and trademark dilution under 15 U.S.C. § 1125(a) and (c). 

Not your usual training-data fight

The AI copyright docket is crowded with cases sharing a common anatomy: Authors, publishers, and visual artists have challenged the legality of ingesting protected works to build large language models. They contest the input side (was copying for training fair use?). The Britannica complaint pleads copyright counts too (training and RAG copying, vicarious and contributory infringement). But the Lanham Act claim targets the output and presentation side: how generated content is labeled, attributed, and displayed relative to third-party brands. 

Why this matters

This case does not depend on proving copying of protectable expression and could reach conduct even where training was licensed or fair. It implicates UI and attribution design and product decisions, not just model architecture, and exposes developers, and potentially the deployers and enterprises that surface branded, attributed AI outputs to end users. If the claim remains viable, it opens an output-focused liability channel independent of the copyright fair-use battles.

The theory: hallucinations as false designation

The complaint alleges two distinct trademark harms:

First, fabricated content falsely attributed to Plaintiffs. ChatGPT “sometimes generates hallucinations in its outputs and wrongly attributes that text to Plaintiffs using Plaintiffs’ trademarks.” (¶68.) The claim is that when a user asks a question and ChatGPT responds with invented information tagged to “Britannica” or “Merriam-Webster,” users are misled into believing the content is “factually correct, complete, and authoritative because [it is] sourced from, associated with, sponsored by, or approved by Plaintiffs.” (¶111.)

Second, undisclosed omissions and alterations. ChatGPT omits or modifies portions of Plaintiffs’ content while purporting to reproduce it, without disclosing the changes, falsely implying completeness and accuracy. (¶69.)

Together, Plaintiffs say these practices amount to (a) false designation of origin and likelihood of confusion under § 1125(a)(1), and (b) dilution by blurring and/or tarnishment under § 1125(c). They allege willfulness: OpenAI has “actual knowledge” of the marks and awareness that ChatGPT falsely attributes content. (¶¶107, 112.)

What to do now

For AI developers and model builders: 

• Guardrails: Implement guardrails that suppress or flag third-party brand names in generated outputs unless paired with high-confidence, verifiable source links.

• Attribution logging: Maintain audit trails of when and how brand names are surfaced alongside generated text.

• Disclaimers and provenance: Where brand names appear in outputs, consider disclosure language distinguishing AI-generated summaries from publisher-endorsed content.

For product, data, and engineering leaders:

• Audit UI: Review how third-party marks, logos, or publisher names are displayed alongside generated text. Source panels, citation badges, and branded summaries could raise liability for trademark claims.

• Gate branded citations behind verification. Don’t display attributions unless retrieval confirms the output matches the cited source.

• Suppression and corrective mechanisms: Enable takedown or correction when hallucinated or outdated brand attributions are identified: proactive monitoring beats reactive litigation.

For enterprise users:

• Treat AI citations as leads, not proof: Worth repeating prior guidance here - verify every branded attribution before relying on it externally. 

• Assess downstream exposure. An organization that surfaces AI outputs to customers or the public, could share liability for false attributions embedded in those outputs.

For legal, compliance, and procurement:

• Update brand-protection programs. Add AI-output monitoring to your trademark watch: run periodic test prompts to see how your marks are being used (and misused) across major AI tools. 

• Tighten vendor contracts. Require AI vendors to indemnify against trademark claims arising from hallucinated attributions; negotiate attribution-control and takedown rights.

• Revisit terms of use. Prohibit unauthorized use of marks in AI-generated outputs; specify remedies and monitoring rights.

Bottom line

The best time to audit your citation and attribution architecture was before this complaint was filed. The second-best time is now – stay tuned for Part III, which will lay out a phased, practical playbook for getting there.