AI IP Litigation in 2025: Where the Lines Actually Moved ⚖️🤖
Copyright, trademarks, patents, and contracts after a year of real decisions
By the end of 2025, “AI IP litigation” stopped being a thought experiment and became a crowded docket:
- Judges in the Northern District of California issued first-wave fair-use rulings on training datasets. (Copyright Alliance)
- Anthropic agreed to a $1.5 billion copyright settlement with authors, including destruction of a pirated-books corpus. (Reuters)
- The New York Times launched a fresh copyright + trademark + hallucination suit against Perplexity AI. (The Guardian)
- A consolidated class of authors survived a motion to dismiss against OpenAI on direct output-infringement theory. (Banner Witcoff)
- Trademark and publicity cases over hallucinated citations, brand misuse, and voice clones started to get real holdings. (JD Supra)
- The USPTO rescinded its 2024 AI inventorship guidance and replaced it with a new, stricter human-inventor standard. (Federal Register)
Layered over that, there are now around 47+ U.S. copyright suits targeting AI training and outputs, with class-wide relief emerging as the “sleeping bear.” (Debevoise)
This piece doesn’t try to catalogue every complaint. Instead, it pulls out the doctrinal shifts that actually matter for product and deal work in 2026.
Training data: fair use didn’t “win,” but it got a foothold 📚⚖️
Courts finally started writing on the central question: is training on copyrighted text fair use? The early answers are mixed—but not a blanket win for either side.
Two Northern District of California decisions—often described together in practitioner summaries as Bartz v. Anthropic and Kadrey v. Meta—took a nuanced approach: (Copyright Alliance)
- Training on lawfully acquired books to improve a general-purpose model can be fair use in principle, particularly where:
- the works were obtained through authorized channels; and
- the use is highly intermediate and non-substitutive.
- Building a “central library” of pirated ebooks to power your training pipeline is not fair use; it’s just large-scale infringement.
You then have a very different signal from settlement:
- Anthropic agreed to a $1.5B settlement with authors, including promises to destroy pirated training data and certify it was not used in commercial products. (Reuters)
Taken together:
- Courts are open to fair use for training on clean, lawfully licensed data.
- Courts and plaintiffs are not forgiving about pirated or grey-market corpora.
- Money on the table (Anthropic) is pushing everyone toward structured licensing rather than “we’ll litigate fair use later.”
For drafting, that’s enough to assume:
If your training data room looks like a random torrent dump, you’re likely outside the fair-use envelope the early cases are sketching.
Licenses vs. lawsuits: 2025 turned into a business model fork 🧾⚔️
Two trends crystallized at the same time:
- Big-ticket settlements and licensing deals
- Media and dataset owners moving to structured AI licenses
- A rising number of publishers, data vendors, and SaaS companies are quietly signing AI training and syndication deals with model providers instead of litigating. (Debevoise)
The upshot for product and corporate counsel:
- If you own valuable content or a curated data asset, AI training is now a separate revenue line you can price and contract for.
- If you build models, you should expect counterparties to show up with Anthropic-style term sheets: licensing fees, usage reporting, certs on where models are allowed to use (and not use) specific corpora.
The 2023–24 question—“Is this even copyrightable?”—has largely been replaced by “Do we litigate or license?”
Outputs: from “hallucinations” to trademark and output infringement 🧠💥
2025 also saw meaningful movement on output liability:
Authors’ consolidated case vs. OpenAI
In the SDNY, the consolidated class action brought by authors against OpenAI survived in part:
- The court denied dismissal of a core direct infringement claim based on allegedly infringing outputs, allowing discovery into when and how models “regurgitate” protected text. (Banner Witcoff)
The same court continues to wrestle with the NYT v. OpenAI/Microsoft case you’ve already written about—where the discovery fight over user logs is as important as the core copyright claims. (U.S District Court)
NYT v. Perplexity: copyright + marks + hallucinations
In December 2025, The New York Times filed a new suit against Perplexity AI that combines: (The Guardian)
- Copyright infringement (training and output regurgitation of Times articles, including paywalled content);
- Trademark and false designation under the Lanham Act (Perplexity allegedly displaying NYT marks next to fabricated “articles” that look like Times reporting); and
- A narrative that hallucinations become a brand and reputational problem when your model attributes invented content to someone else’s trademark.
That hybrid copyright/trademark theory echoes an earlier 2025 decision in Advance Local v. Cohere, where a court suggested that misattributed AI hallucinations can create Lanham Act exposure when they plausibly confuse users about source or sponsorship. (JD Supra)
Voice clones and publicity: trademark isn’t the only hook
A 2025 decision involving AI voice cloning (plaintiffs alleging unauthorized use of their voices in advertising) knocked out federal trademark and copyright claims, but allowed state right-of-publicity theories to proceed. (Fredrikson & Byron)
Takeaway: when models simulate a recognizable voice or persona:
- Trademark may be a stretch unless the voice itself is a source identifier.
- Publicity and unfair-competition theories are often stronger.
Contract and trade-secret fights: licenses are becoming weapons 🧩🕵️
The most interesting “non-headline” case for in-house lawyers this year might be Fastcase (Clio) v. Alexi:
- Fastcase alleges that rival legal-tech firm Alexi used its licensed legal database to train a competing AI legal-research tool, violating a data-licensing agreement that expressly barred such uses, and misappropriating trade secrets. (Reuters)
Regardless of outcome, the complaint reads like a model for how content licensors will plead AI-misuse:
- explicit license field restrictions;
- “no training” clauses;
- post-merger competitive dynamics (Fastcase + vLex + Clio vs. Alexi);
- alleged scraping or over-use beyond the API’s intended scope.
Practically, that means:
- If you license in data, your counterparty will start insisting on Alexi-style language that fences off training and model building.
- If you license out data, you should assume your agreement will be Exhibit A the next time someone trains a model in ways you didn’t anticipate.
The IP fight is as much contract interpretation and trade secrets as it is copyright.
Patents: USPTO resets the AI inventorship rules 🧬📑
On November 28, 2025, the USPTO rescinded its February 2024 AI inventorship guidance and published Revised Inventorship Guidance for AI-Assisted Inventions in the Federal Register. (Federal Register)
Key points:
- AI systems still cannot be named as inventors—consistent with the Thaler line of cases.
- But the Office clarified how to treat AI-assisted inventions:
- There must be a natural person who makes a significant contribution to each claimed invention under traditional conception standards.
- Merely “pushing a button” on an AI tool is not enough; there must be intellectual contribution about the problem, solution, or the way the model’s output is selected and applied.
- Applicants must clearly document human inventorship—who did what, and how that goes beyond routine use of AI. (Morgan Lewis)
For patent strategy, that means:
- You need internal records showing which humans conceived which aspects of the AI-assisted invention.
- Claims should be drafted to reflect human-conceived technical contributions, not just the fact that “the model did something clever.”
In short: AI-assisted inventions remain patent-eligible, but only with careful human-inventor framing.
2025 AI IP themes in one table 📊
| Theme | What actually moved in 2025 | Practical result for 2026 deals |
|---|---|---|
| Training data & fair use | ND Cal decisions suggest some training on lawfully acquired works may be fair use; pirated corpora are not. (Copyright Alliance) | Clean provenance matters. Expect more dataset reps, audits, and indemnities tied to acquisition chain. |
| Class actions & settlements | ~47 copyright suits against AI companies; Anthropic’s $1.5B settlement + data destruction. (Debevoise) | Plaintiffs are using class-wide relief and structural terms (data destruction, guardrails) as leverage. |
| Outputs, brands & hallucinations | Authors’ output claims vs OpenAI survive MTD; NYT v Perplexity ties copyright, trademarks, and hallucinations; Cohere decision recognizes Lanham exposure for misattributed outputs. (Banner Witcoff) | Output-side risk now clearly includes copyright + trademark + publicity. Expect tighter output filters, logging, and indemnity caps. |
| Contracts & trade secrets | Fastcase v Alexi frames training on licensed databases as breach + trade-secret misappropriation. (Reuters) | “No AI training” and field-of-use clauses are becoming standard—and litigated. Your license boilerplate is now critical IP infrastructure. |
| Patents & inventorship | USPTO 2025 guidance rescinds prior AI memo; reemphasizes significant human contribution as inventorship touchstone. (Federal Register) | Patent programs must trace human contributions in AI-assisted R&D and reflect them in claim strategy and disclosure. |
Practical takeaways for 2026 contract and product work 🧰
For AI developers and platforms
- Clean your training stack
- Map which portions of your corpus are licensed, scraped, or third-party; assume pirated / shady segments are future settlement fuel.
- Build the ability to quarantine and retire specific corpora (Anthropic-style data destruction) without bricking your product.
- Treat outputs as product claims, not purely technical artifacts
- Logging, prompt/output retention (subject to privacy law), and an audit trail are now discovery issues in output litigation.
- Consider contractual carve-outs for certain high-risk uses (news, health, legal, financial advice) where hallucinations + marks are most dangerous.
- Align patents with your product reality
- For AI-assisted inventions, maintain lab notebooks / internal memos that show human conception and selection, not just “the model said so.”
For rights holders (publishers, SaaS providers, data vendors)
- Decide your posture: license, litigate, or both
- If you have a valuable corpus, assume someone is already training on it. The realistic options are:
- license it;
- fence it off technically and contractually;
- or treat it as a future complaint with class-wide relief and structural terms, not just damages.
- If you have a valuable corpus, assume someone is already training on it. The realistic options are:
- Refit your outbound licenses
- Add express AI-use provisions: permitted/forbidden uses, model types, logging duties, and re-training restrictions.
- Bake in audit rights, “no circumvention” language (no scraping around the license), and data-destruction obligations if the deal ends badly.
For everyone else using third-party AI
- Revisit indemnities and caps
- Many standard SaaS MSAs were not written with class-wide AI copyright and trademark exposure in mind.
- At minimum, carve out IP infringement from low caps, or negotiate a separate, AI-specific indemnity tied to training and outputs.
- Clarify who owns what in AI-assisted work
- Internal policies should explain when model-assisted work is:
- proprietary to your company;
- jointly owned;
- or subject to license conditions from model vendors.
- Internal policies should explain when model-assisted work is:
Frequently asked questions 💬
If courts are finding some AI training is fair use, can we stop worrying about licenses?
No.
The early fair-use decisions are:
- fact-specific (lawfully acquired books, particular technical implementations); and
- limited to training conduct, not output regurgitation or market-substitution claims. (Copyright Alliance)
At the same time:
- Anthropic just agreed to a record settlement plus data destruction, despite fair-use arguments being available. (Reuters)
- Plaintiffs are explicitly seeking injunctive and structural relief, not just damages.
From a counseling perspective, fair use is now one defense among several, not a basis to treat unlicensed scraping or grey-market datasets as “safe.” Licenses, provenance, and technical controls still matter.
Do we need to renegotiate old content licenses that never mention AI or “training”?
In many cases, yes or soon—especially if:
- the license is broad (“for any and all purposes”) but predates generative AI; and
- one side now wants to treat AI training as either within or outside that scope.
Old agreements are already being weaponized in disputes like Fastcase v. Alexi, where the licensor says “you used our data to train a rival AI in breach of the license,” and the defendant insists it stayed within its contractual rights. (Reuters)
To avoid litigating what “use in connection with your services” meant in 2018:
- Inventory legacy licenses that involve bulk access to text, images, audio, or logs.
- Decide your preferred policy:
- Expressly allow AI training (perhaps for a fee);
- Expressly prohibit it; or
- Allow it only under tight conditions (model type, security, non-competition).
- Use renewals, amendments, and new SOWs as opportunities to drop in AI-specific clauses, rather than leaving it to later courts to decide what the old language implies.
The 2025 litigation wave shows that AI IP issues are no longer just for the OpenAIs and Anthropics of the world; they’re baked into every content and data contract you draft going forward.