California sits at the intersection of cutting-edge AI, a dense concentration of creative industries, and platform-driven electronic commerce. In 2025, those forces produced a noticeable shift in copyright and trade mark law: courts began moving from early procedural disputes to substantive decisions that now shape product design, contracting and compliance choices.

In the copyright arena, many of the disputes in 2025 concerning generative AI training focused less on whether a model can occasionally regurgitate protected content and more on whether the training process involved unlawful copying at scale. The legal analysis, in turn, tends to centre on fair use, data provenance and market harm.

Fundamentally, copyright generally protects expression, not ideas or facts:  novels, screenplays, and song lyrics sit at the core of what copyright protects. That baseline matters because it frames why AI ingestion or training disputes are not simply about abstract learning but about copying protected expression, and raise questions about copying protected expression at scale and whether permission or fair use applies.

In 2025, California was also on the leading edge of novel trade mark disputes involving digital-native products, creator branding and platform-scale naming collisions. These cases did not replace traditional infringement and counterfeiting disputes, but they layered in new fact patterns where confusion and reputation harm can travel faster than litigation timelines. For clients, the practical issue is risk management in naming, advertising claims and the internal enforcement mechanisms of major platforms.

Trade mark law protects brand identifiers that signal source, sponsorship or affiliation. In digital commerce, those identifiers can be words, logos or even the presentation of an online product experience. The year’s cases from California show that courts continue to apply traditional confusion analysis, but they are increasingly asked to do so in contexts where the product is virtual or the marketing is platform-native, especially where “dupe” or comparison marketing implies equivalence.

Key California 2025 Copyright and Trade Mark Cases at a Glance

• Bartz v Anthropic (N.D. Cal., June 2025) – training held fair use on the record presented, but the court treated piracy and “central library” retention as a separate problem that could still create liability.
• Kadrey v Meta (N.D. Cal., June 2025) – Meta won summary judgment on fair use, but the court’s reasoning emphasised market substitution and warned that the next cases will turn on whether plaintiffs develop meaningful evidence of market harm and market dilution.
• Concord Music Group v Anthropic (N.D. Cal., March 2025) – court declined a broad preliminary injunction against alleged lyric use, highlighting overbreadth of the requested relief and insufficient proof of irreparable harm.
• Lyon v Adobe (N.D. Cal., filed 16 December 2025) – proposed class action against Adobe alleging unlawful book copying for training SlimLM models, illustrating the next wave of data supply chain allegations.
• Yuga Labs v Ripps (9th Cir., July 2025) – the Ninth Circuit held non-fungible tokens (NFTs) can qualify as goods under the Lanham Act, but required a traditional confusion analysis.
• Cameo v OpenAI (N.D. Cal., November 2025) – Temporary Restraining Order disallowing OpenAI’s use of “Cameo” branding for a feature, highlighting naming collisions and reverse-confusion risk.
• Williams-Sonoma v Quince (N.D. Cal., November 2025) – false advertising and unfair competition claims targeting “dupe” marketing and implied equivalence.

Copyright in California in 2025 – Emerging Themes in AI Training Cases

The leading decisions tend to converge on one point: AI training can be characterised as transformative in purpose. The divergence is not whether AI can be transformative, but whether the defendant’s data acquisition and retention practices can be justified and whether the plaintiff can prove meaningful market harm. For clients, these disputes are increasingly less about slogans (“training is fair use” versus “training is theft”) and more about facts: provenance, retention, proof of market harm and the scope of requested relief.

In practical terms, courts are pushing parties toward evidence. Plaintiffs are being asked to show more than hypothetical harms, and defendants are being asked to show more than general innovation narratives. Courts are also increasingly attentive to the mechanics of copying: where the data came from, how it was stored, and whether a defendant’s internal practices resemble a permanent library rather than a narrow use tied to training.

The emerging themes can be summarised as follows.

• “Transformative training” is not a blanket immunity for all data practices.
• Courts may ask for a clean provenance story – what was sourced, how, when and why.
• Internal retention and reuse policies matter, especially if datasets become shared resources across teams.
• Market impact remains the most contested factor and will likely be the focal point of the next wave of merits decisions.

Bartz v Anthropic: training can be fair use, but piracy does not disappear

In Bartz v Anthropic (23 June 2025), Judge Alsup issued a detailed order that is now a central reference point for AI training disputes. The court did not treat ingestion as one act. Instead, it separated different categories of copying and evaluated them under the fair use factors, creating a practical roadmap for how companies and plaintiffs can map their practices to the categories the court analysed.

On the record presented, the court held that using books to train Anthropic’s models was fair use. The court described training as “exceedingly transformative” and observed that “[t]he technology at issue was among the most transformative many of us will see in our lifetimes”, emphasising that the purpose was to build a tool rather than to republish the books. The court also held that converting lawfully purchased print books into digital copies for a searchable internal library could be fair use, treating internal format-shifting differently from mass acquisition from unauthorised sources.

The opinion, however, treated piracy and long-term retention as a distinct risk. The court emphasised that downloading pirated copies from unauthorised sources and retaining them as part of a “central library” was not automatically excused by the later claim of transformative training. That distinction is important for clients because it means that even if training itself can qualify as fair use, acquisition and retention practices can still generate separate liability exposure. Data provenance and retention are not side issues; they are central risk variables.

The lessons of Bartz:

• “transformative training” is not a blanket immunity for all data practices;
• courts may ask for a clean provenance story – what was sourced, how, when, and why; and
• internal retention and reuse policies matter, especially if datasets become shared resources across teams.

Kadrey v Meta: a defence win that reads like a warning

In Kadrey v Meta (25 June 2025), Judge Chhabria granted Meta summary judgment on fair use on the record presented. The court reasoned that model training is aimed at creating a functional tool, not distributing the books as books, and for that reason it treated the use as transformative. The decision, however, is better understood as an evidentiary ruling than as a broad endorsement of training without permission.

The court emphasised that the fourth factor – market effect – is “undoubtedly the single most important element of fair use”, and aligned its analysis with the Supreme Court’s emphasis in Warhol on market substitution in commercial settings. The court identified several market-harm theories that commonly appear in AI training cases, including “regurgitation”, harms to licensing markets and indirect substitution (sometimes described as market dilution). The court viewed indirect substitution and market dilution as potentially significant but found that plaintiffs in this case had presented “no meaningful evidence” to support that theory.

The opinion also included a cautionary passage that has been widely cited in commentary. Even while granting summary judgment to Meta, the court suggested, in conclusion, that it is difficult to justify copying books without permission to develop a tool expected to generate massive commercial value while enabling a potentially endless stream of competing works. The decision turned on the evidentiary record before the court rather than on a broad rule approving all AI training practices.

For clients, Kadrey’s message is two-sided. Plaintiffs must build market evidence, particularly for indirect substitution and dilution. Defendants must be prepared for that evidence and cannot assume that “transformative” alone will carry the day in a future case with a stronger market record. Kadrey is a preview of the next fight, not the last word.

What Kadrey means:

• market impact is where the next cases will turn, and economic evidence will matter;
• courts may not presume a new training data licensing market exists; plaintiffs must show the market and the harm; and
• a defence win at summary judgment can co-exist with judicial scepticism about the underlying conduct.

Concord Music Group v Anthropic: courts are cautious about broad injunctions

The publishers’ lawsuit against Anthropic over song lyrics illustrates another theme: courts may be hesitant to grant sweeping preliminary injunctions when requested relief is broad and the record is underdeveloped. In March 2025, the Northern District of California declined to grant the publishers’ requested preliminary relief, emphasising the overbreadth of the request and the lack of sufficient proof of irreparable harm on that early record.

This matters for clients because it signals what kinds of remedies are realistically attainable early. Plaintiffs may find it easier to obtain narrower operational commitments, guardrails or preservation orders than to secure a broad “stop training” injunction. Defendants should expect that courts may be sceptical of blanket relief that effectively regulates an emerging industry before full merits development.

Lyon v Adobe: the next wave focuses on data lineage and “supply chain” provenance

The proposed class action in Lyon v Adobe (filed 16 December 2025) signalled a broadening of targets and theories. The complaint alleges Adobe used copyrighted books, including pirated books, to train small language models referred to as SlimLM. Public reporting described SlimLM as smaller models aimed at document-related tasks and mobile deployment, illustrating that training disputes are not limited to the largest general-purpose models and can also involve smaller models embedded in products.

The case is early, but it highlights a growing focus on training-data lineage. Plaintiffs increasingly allege that companies relied on publicly released datasets marketed as open-source, even though those datasets may incorporate text drawn from pirated book collections. That approach puts pressure on companies to show what diligence they performed on dataset sourcing and how they documented and audited the data they used, rather than relying on the fact that a dataset was publicly available.

For clients, the practical consequence is straightforward. Data provenance is not just a question of whether the company scraped content directly. It also requires being able to show where third-party training datasets came from, what vetting was done before adopting them, and what controls the company maintained over their use and retention.

The reported USD1.5 billion Anthropic settlement

The reported USD1.5 billion settlement in Bartz (reported in October 2025 and now proceeding through a court approval process) changed the negotiation posture around AI training disputes. Whatever the ultimate law on training as fair use, the settlement demonstrates that “pirated library” allegations can create existential exposure. Public reporting also emphasised operational components, including destruction of disputed pirated data and certification tied to whether that data was used in commercial products.

For clients, the lesson is not the headline number alone. The settlement underscores that risk is driven by provenance facts, retention practices and internal controls. Companies that cannot tell a coherent sourcing story may find that even favourable legal arguments become difficult to monetise at scale, because the litigation risk becomes too large to carry.

Trade Mark and Unfair Competition in California in 2025

NFTs and virtual goods

In Yuga Labs v Ripps (9th Cir., July 2025), the Ninth Circuit held that NFTs can qualify as “goods” under the Lanham Act. That holding matters because it provides a direct answer to a threshold question that often complicates virtual goods disputes: whether trade mark law can apply to digital-native assets sold in online marketplaces.

The Ninth Circuit’s decision, however, did not hand brands an automatic win. It reversed summary judgment for Yuga on infringement and cybersquatting because likelihood of confusion had not been established as a matter of law. The court required a traditional confusion analysis and emphasised that even in digital markets, trade mark claims must still be proven with evidence tied to consumer perception and the way the product is marketed and sold.

For clients, Yuga provides both an enforcement opportunity and a caution. Brands can assert rights in NFT markets, but defendants can still force a rigorous confusion inquiry. The “virtual” nature of the good does not eliminate trade mark protection, but it also does not eliminate defences or the need for proof.

Naming collisions and reverse confusion risk: Cameo v OpenAI

Cameo v OpenAI (N.D. Cal., November 2025) highlights a practical issue many clients underestimate: feature naming can become litigation. In November 2025, the Northern District of California issued a Temporary Restraining Order (later extended through February of 2026 after a hearing on a preliminary injunction) forbidding OpenAI from using “Cameo” or confusingly similar marks in connection with a feature, service or functionality. The order reflects a classic reverse-confusion problem: a well-funded entrant with massive reach can overwhelm a smaller incumbent’s market identity.

For clients launching products in fast-moving categories, the lesson is simple. Naming decisions should be treated as clearance decisions, not purely marketing decisions. A naming collision can trigger emergency relief, force a rebrand at a critical launch window, and create reputational harm that persists even after litigation settles.

“Dupe” marketing and false advertising: Williams-Sonoma v Quince

Williams-Sonoma v Quince (filed in Northern California in November 2025) illustrates a broader trend in trade mark-adjacent disputes. Rather than alleging classic counterfeiting, the complaint focuses on comparative marketing and alleged implied equivalence. Public reporting highlighted slogans and comparison framing, including “Like Williams-Sonoma, but half the price”, and asserted false advertising and unfair competition theories.

This approach reflects a common enforcement strategy in modern e-commerce. Where product design or trade dress claims might be complicated by functionality or by the expense of consumer surveys, brand owners increasingly focus on the truthfulness of comparative claims. If a seller implies equivalence, the seller may be required to substantiate it. For clients who use comparison marketing, the practical risk is not only trade mark confusion but also false advertising exposure.

E-commerce and Platform Dispute Systems: The Practical First Forum

In 2025, many IP disputes were resolved – or at least defined – inside platform dispute systems. Amazon’s rights-owner tools, TikTok Shop’s intellectual property reporting channels, and Meta’s brand protection systems operate with their own filing requirements, evidentiary expectations and appeal workflows. For many clients, the first meaningful decision about an IP dispute is a platform decision, not a court ruling.

That shift has practical consequences. Platform decisions often occur on significantly shorter timelines than court proceedings, and the economic impact of a takedown, suspension or delisting can be immediate. Platform decisions also shape litigation posture: if a party wins reinstatement, settlement leverage changes; if a party loses access, the pressure to seek emergency relief increases.

Clients should treat platform submissions as a form of mini-litigation. The quality of the record matters, and the submissions must be understandable to non-lawyer reviewers. A clear ownership packet, concise infringement narrative and well-organised exhibits often matter more than the sophistication of legal citations.

Online Contracting: The Continuing Decline of “Browsewrap”

California’s e-commerce environment is also shaped by online contract formation and arbitration enforceability. In 2025, the Ninth Circuit continued to emphasise that users must receive reasonably conspicuous notice of terms and must unambiguously manifest assent. Cases such as Godun v JustAnswer (9th Cir. April 2025) underscore the Ninth Circuit’s continued scrutiny of online sign-up and payment flows where the terms are presented only by hyperlink and the interface does not clearly tell users that clicking the button will bind them to those terms.

For clients, the takeaway is operational. If arbitration and terms are part of a risk-management strategy, the checkout and sign-up flows must be designed for enforceability. Clear language near the action button (“By clicking Pay Now, you agree to the Terms of Service”) and a conspicuous hyperlink remain the most defensible approach.

Practical Takeaways for 2026

For AI developers and deployers:

• audit provenance and keep an evidentiary record that you can produce in discovery;
• separate use cases in policy and practice: training, fine-tuning, evaluation, and general library retention;
• align marketing statements with reality, especially claims about ethical sourcing and commercial safety; and
• prepare for market-harm arguments by understanding where your product may substitute for human-created works.

For creators and rights holders:

• build market evidence early, especially if the theory is indirect substitution or market dilution;
• focus on concrete conduct that courts can remedy: unlawful acquisition mechanics, retention, and distribution; and
• consider targeted relief and operational commitments where broad injunctions are unlikely on an early record.

For brands and e-commerce operators:

• treat product and feature naming as a clearance issue, not a branding exercise;
• review comparative advertising for implied equivalence and substantiation; and
• develop platform-ready enforcement and defence packets, because platform decisions often set the litigation posture.

Conclusion

California’s 2025 IP landscape reflects a shift from speculative debates to operational scrutiny. Courts and platforms increasingly expect proof about data sourcing, market effects, and the realities of how products reach consumers. For clients, that means legal risk is now tied to governance, documentation, and process – often long before any complaint is filed.

The unifying theme is operational readiness. The strongest positions in 2026 may well belong to parties who can demonstrate a coherent provenance story, develop credible market evidence, and manage platform disputes with the same discipline applied to court filings. Those who treat these disputes as abstract policy debates may find that the real decisions are being made elsewhere, faster, and with higher immediate commercial stakes.