This week, a federal court in Tennessee transferred to California a lawsuit brought by several large music publishers against a California-based AI company, Anthropic PBC. Plaintiffs in Concord Music Group et al. v. Anthropic PBC[1] allege that Anthropic infringed the music publishers’ copyrights by improperly using copyrighted song lyrics to train Claude, its generative AI model.  The music publishers asserted not only direct copyright infringement based on this training, but also contributory and vicarious infringement based on user-prompted outputs and violation of Section 1202(b) of the Digital Millennium Copyright Act for allegedly removing plaintiffs’ copyright management information from copies of the lyrics.  On November 16, 2023, the music publishers also filed a motion for a preliminary injunction that would require Anthropic to implement effective “guardrails” in its Claude AI models to prevent outputs that infringe plaintiffs’ copyrighted lyrics and preclude Anthropic from creating or using unauthorized copies of those lyrics to train future AI models. 

Last week, in Vidal v. Elster, the Supreme Court upheld the Lanham Act’s prohibition against registering a trademark that includes a living person’s name without their consent.[1]  This case is the latest in a trilogy of challenges to the constitutionality of trademark registration bars in the Lanham Act.  The Court previously struck down as unconstitutional the clauses in Section 2(c) prohibiting registration of marks constituting “disparagement” and “immoral or scandalous matter.”[2]  In a departure from those decisions, the Court upheld the U.S. Patent and Trademark Office’s refusal to register a trademark for “Trump Too Small”—a piece of political commentary that the applicant sought to use on apparel to criticize a government official.  The Court reasoned that, unlike the other provisions, the “names” prohibition is viewpoint-neutral, and thus does not violate any First Amendment right. 

In a recent en banc decision concerning the standard for assessing obviousness challenges to design patents, the United States Court of Appeals for the Federal Circuit discarded its long-standing standard, known as the Rosen-Durling test and regarded by many as overly-rigid, and held that the standard for design patents should be the same as for utility patents.  The decision in LKQ Corporation v. GM Global Technology Operations LLC[1] will have significant implications for design patent applicants and owners going forward.

Yesterday, the Supreme Court denied certiorari in Hearst Newspapers, LLC v. Martinelli, declining to determine whether the “discovery rule” applies in Copyright Act infringement cases and under what circumstances.  As a result, most circuits will continue to apply the rule to determine when an infringement claim accrues for purposes of applying the Copyright Act’s three-year statute of limitations.

Last week, a divided Supreme Court held in Warner Chappell Music, Inc. et al. v. Nealy et al. that a copyright plaintiff who timely files an infringement lawsuit based on the “discovery rule” may recover damages for infringements that occurred outside the Copyright Act’s three-year statute of limitations period.[1]  A claim generally accrues when an infringing act occurs, but many circuits apply a “discovery rule,” pursuant to which a claim accrues when a plaintiff has (or with reasonable diligence should have) discovered the infringement, which could be many years later.  Courts applying this rule have recently disagreed on how far back damages are available, with the Second Circuit holding that a copyright claimant may recover only three years’ of damages, even if the suit was otherwise timely under the discovery rule.  The Supreme Court rejected that conclusion, holding that “no such limit on damages exists” in the Copyright Act, which “entitles a copyright owner to recover damages for any timely claim” no matter when the infringement occurred.  

Last week the Fourth Circuit reversed a $1 billion copyright verdict against an internet service provider and ordered a new trial on damages allegedly arising from illegal music downloads by its subscribers.  In Sony Music Entertainment et al. v. Cox Communications Inc. et al.,[1] a group of music producers belonging to the Recording Industry Association of America brought suit against Cox for contributory and vicarious copyright infringement based on allegations that Cox induced and encouraged rampant infringement on its service.  In 2019, a jury found Cox liable on both theories for infringement of 10,017 copyrighted works and awarded $99,830.29 per work, for a total of $1 billion in statutory damages.  On appeal, the Fourth Circuit issued a mixed ruling – upholding the finding of contributory infringement but reversing the vicarious liability verdict and remanding for a new trial on damages. 

This week saw yet another California federal court dismiss copyright and related claims arising out of the training and output of a generative AI model in Tremblay v. OpenAI, Inc.,[1]a putative class action filed on behalf of a group of authors alleging that OpenAI infringed their copyrighted literary works by using them to train ChatGPT.[2]  OpenAI moved to dismiss all claims against it, save the claim for direct copyright infringement, and the court largely sided with OpenAI. 

In Punchbowl, Inc. v. AJ Press, Inc., the Ninth Circuit revived a trademark infringement case previously dismissed on grounds that the First Amendment shields “expressive” trademarks from Lanham Act liability unless plaintiff can show the mark (1) has no artistic relevance to the underlying work, or (2) explicitly misleads as to its source.[1]  This is known as the Rogers test, and effectively operates as a shield to trademark liability where it applies.  Last year, the Supreme Court limited application of the Rogers test in Jack Daniel’s Properties, Inc. v. VIP Products LLC, [2] holding that it does not apply where the challenged use of a trademark is to identify the source of the defendant’s goods or services.  In those instances, a traditional likelihood of confusion or dilution analysis is required. 

This third part of our four-part series on using synthetic data to train AI models explores the interplay between synthetic data training sets, the EU Copyright Directive and the forthcoming EU AI Act.