Courts finally ruled. The USPTO rewrote the playbook. Anthropic settled for $1.5 billion. Here’s what the legal fog looks like now — and why the “just use AI” strategy is shakier than anyone’s admitting.
- No court has ever granted copyright to a work made solely by AI — the D.C. Circuit confirmed this March 2025
- The USPTO (November 2025) treats AI as a lab instrument, not a co-inventor; “conception” must happen in a human mind
- Training-data liability is still an open legal question — the Copyright Office says commercial copying from pirated sources is probably not fair use
- Anthropic settled a training-data class action for up to $1.5 billion in 2025; no AI company has won a definitive fair-use ruling yet
- Documenting human creative input isn’t optional anymore — it’s the only thing standing between you and an unprotectable output
My client called me at 11 p.m. last October. Their marketing team had spent four months building a campaign with AI-generated visuals — great stuff, genuinely — and they’d just found out a competitor was using nearly identical imagery. I had to tell them: they probably couldn’t sue. Because they hadn’t documented a thing. No prompts saved. No revision logs. No record of which human made which creative call. Four months of work, legally defenseless.
I’ve seen this movie three times in the past year. The law has moved faster than most companies’ workflows. So here are the seven rules that actually matter right now — not theoretical constructs, but principles that courts and regulators codified in 2025.
- The Human Authorship Requirement — still the bedrock
- Work-for-Hire in an AI Context — employment agreements need updating
- Platform Licensing — you don’t own what you think you own
- Patent Inventorship — the USPTO rewrote the test in November 2025
- Training Data Liability — the $1.5B question
- Joint Ownership — when two teams build one thing
- International Divergence — China recognized AI copyright; the US didn’t
On March 18, 2025, the D.C. Circuit affirmed what the Copyright Office had been saying for years: human authorship is a bedrock requirement for copyright registration. The case was Thaler v. Perlmutter. Stephen Thaler — the computer scientist who’s been running this legal experiment for years — created a work entirely through his AI system, the “Creativity Machine,” then applied for copyright protection.
The court said no. Again. Thaler petitioned the Supreme Court in October 2025 — we don’t yet know if they’ll take it.
But here’s what the ruling actually doesn’t resolve: how much human input is enough. The Copyright Office’s January 2025 report on AI and copyrightability is Tier 2 — credible institutional guidance, not binding law careful to say it’s case-by-case. A human who writes 50 prompts, reviews iterations, selects elements, and edits the output is in much better shape than someone who types two words and accepts whatever the model spits out.
The dangerous part isn’t not knowing the rule. It’s that AI-generated outputs look complete and professional — they don’t signal their own legal vulnerability. A human designer’s rough sketch comes with obvious signals of authorship. A polished Midjourney image doesn’t. You can publish, market, and license that image for months before anyone challenges it. By then, you’ve made business decisions downstream of a legally shaky foundation.
The practical fix: maintain a creative process log. What prompt did you use? What did you reject? What did you edit? That file is your evidence of authorship.
“Where AI merely assists an author in the creative process, its use does not change the copyrightability of the output. At the other extreme, if content is entirely generated by AI, it cannot be protected by copyright.”
U.S. Copyright Office, Copyright and Artificial Intelligence Part 2: Copyrightability (January 2025)Rule 02Work-for-Hire Still Works — But Your Contracts Don’t Know That Yet
Under standard work-for-hire doctrine, employers own what employees create within the scope of their jobs. This extends to AI-assisted work — but your employment agreements were probably written before your team started using AI tools daily.
Three specific gaps I see constantly:
Gap one: “work product” isn’t defined to include AI-assisted output. A contract that says “all creative work produced by the employee” might face a challenge if the employee argues they only operated the AI, and the AI produced the work. Ambiguous contracts invite disputes.
Gap two: contractor agreements are worse. Freelancers retain rights to their work unless written agreements say otherwise. If your contractors are using AI tools and the agreement doesn’t address it — specifically — you have a problem. The contractor might argue they own the output because they operated the tool.
Gap three: which AI platforms are authorized matters. If an employee uses a personal Midjourney account rather than a company-licensed platform, the ownership question gets murkier. Some platform terms vary by account type.
Fix: update your agreements. Add explicit definitions of “work product” that cover AI-assisted outputs. Specify which platforms employees and contractors are authorized to use. Require documentation of AI usage. This isn’t paranoia — it’s hygiene. (Damn, I feel like a broken record on this, but four of my last eight consultations were exactly this problem.)
Rule 03You Don’t Own What You Think You Own — Platform Licensing
Every major AI platform has a terms-of-service document you agreed to without reading. Most grant you rights to your outputs. But the details matter.
OpenAI’s terms grant users rights to outputs while retaining a license to use those outputs for service improvement. Google’s Gemini terms are similar. Midjourney — and this catches people — has a free tier where you do not get commercial rights. Paid subscriptions do. If your company has one paid account that multiple employees use, check whether the terms cover that use case.
The real risk isn’t platform rights, though. It’s indemnification. Does the platform indemnify you against claims that their AI output infringes someone else’s copyright? Some enterprise tiers do. The consumer-facing products generally don’t. If you’re putting AI-generated content in a product that competes with, say, Getty Images’ content — and it turns out the model trained on Getty’s images — that’s your legal exposure, not the platform’s.
| Platform | User Owns Output? | Commercial Use? | Enterprise Indemnification? | ⚠ Adversarial note |
|---|---|---|---|---|
| OpenAI GPT-4o | Yes (per ToS) | Yes | ChatGPT Enterprise tier only | Terms can change; enterprise indemnification has scope limits and carve-outs. Not legal insurance. |
| Adobe Firefly | Yes (paid) | Full commercial rights | Yes — enterprise tier | Trained on licensed/owned content, but “commercially safe” claims are marketing, not legal guarantees. |
| Midjourney | Yes (paid plan) | Yes (paid) | No | Free tier outputs are licensed to Midjourney; commercial use restricted. Easy to mix up plans in team settings. |
| Stable Diffusion (self-hosted) | Yes | Yes | N/A | Training data origins are disputed; Andersen v. Stability AI is still proceeding through courts. Self-hosting doesn’t insulate you from derivative-work claims if those ever land. |
| Google Gemini | Yes (per ToS) | Yes | Workspace Business/Enterprise | Google retains broad license for service improvement; enterprise tier limits this somewhat but not entirely. |
Rule 04Patents: The USPTO Rewrote the Test in November 2025
This is the one that surprised most of my R&D clients.
On November 28, 2025, the USPTO published revised inventorship guidance that scrapped the Biden-era 2024 approach entirely. The old framework had applied the Pannu factors — the joint-inventorship test for multiple human inventors — to human-AI collaborative work. It was creating confusion about how much human contribution was “significant enough.”
The new guidance is cleaner: AI is a tool. Like a spectrometer or a database. Conception — the moment when the inventor has a definite and permanent idea of the complete, operative invention — must occur in a human mind. That’s the test. Same as it’s always been. AI might surface the candidate compound, but the inventor must understand what it is and why it matters.
The practical consequence: companies that have been loose about inventorship records for AI-assisted R&D need to fix this now. As Brownstein noted in their analysis, incorrect inventorship can render patents unenforceable or invalid. If a future competitor challenges your patent and you can’t demonstrate that a natural person actually conceived the claimed invention — with documented evidence — you’re exposed.
The patent guidance and the copyright framework are pulling in the same direction but with a meaningful asymmetry. Copyright requires human creativity in the output — what was selected, arranged, edited. Patent law requires human conception in the process — what was understood and intended in the inventor’s mind. A company using AI for drug discovery can satisfy patent requirements by documenting that their scientists understood and selected among AI-generated candidates. But that same company might still face copyright challenges on AI-generated documentation, diagrams, or reports if the human creative input to those outputs wasn’t captured. The two frameworks demand documentation of different things at different stages. Most companies are tracking neither.
Editorial synthesis — sources: USPTO Federal Register Nov. 2025; U.S. Copyright Office Jan. 2025 Report; Mayer Brown analysis Dec. 2025Rule 05Training Data Liability — The $1.5 Billion Question
This is the most unsettled area in all of AI IP law. And the stakes are enormous.
The short version: dozens of lawsuits are pending against AI companies for using copyrighted material without permission to train their models. The Copyright Office released a 108-page report in May 2025 saying the answer depends on the facts — but commercial copying from pirated sources to generate content that competes with the originals is probably not fair use. That’s the Copyright Office’s view, which is advisory not binding, and was further complicated by the Trump administration dismissing the Register of Copyrights the day after the report dropped. Tier 3 — this institutional guidance is contested and may not represent final policy
What actually happened in courts in 2025:
Anthropic settled Bartz v. Anthropic — a class action over training-data use — for up to $1.5 billion. The case involved allegations that Anthropic had downloaded roughly 7 million books, some from pirated sources, to train its Claude models. According to reporting by Reuters, an adverse ruling could have exposed Anthropic to billions more. As of October 2025, no court had issued a definitive fair-use ruling on AI training data — the next expected decisions are summer 2026 at the earliest.
In the Northern District of California, a court ruled that Meta’s use of books for model training was fair use — but that ruling is specific to those facts, and different judges are reaching different conclusions.
Here’s what the AI companies’ fair-use defense gets right, and why this isn’t as clear-cut as rights-holders want it to be: In Bartz v. Anthropic, Judge Alsup described LLM training as “transformative — spectacularly so,” comparing it to human reading and learning. There’s a genuine argument that a model trained on books doesn’t reproduce those books — it learns patterns. Courts haven’t closed this door. The problem for AI developers is that some models do appear to memorize and reproduce training content verbatim under certain prompting conditions. That’s where “transformative” breaks down and infringement becomes more plausible. The line between learning and copying isn’t settled. Probably won’t be until 2026 or later.
For companies using third-party AI tools: your exposure is indirect but real. If the model you’re relying on is found to have been trained on infringing data, and the outputs you’re using are substantially similar to the training material, you could face claims. Enterprise-tier indemnification (see Rule 3) is the practical hedge. Check if you have it.
“Making commercial use of vast troves of copyrighted works to produce expressive content that competes with the original works in existing markets… may not be a fair use.”
U.S. Copyright Office, Generative AI Training Report (May 2025, pre-publication) — note: advisory only; institutional position contested following Register of Copyrights dismissalRule 06Joint Ownership — When Two Teams Build One Thing
Joint IP ownership is a mess even without AI. Add AI into a multi-party development arrangement and it gets genuinely complicated.
The default legal rule in most U.S. jurisdictions: joint owners each have an equal right to exploit the IP — including licensing it to third parties — without the other owner’s consent. They may owe an accounting of revenue. But they don’t need permission. If you co-develop an AI system with a partner company and don’t have a written agreement governing this, either party could license your jointly-developed model to your direct competitor. Legally.
The AI-specific wrinkle: when companies build together, contribution tracking is hard. Who contributed the training data? Who did the architecture work? Who refined the prompts? In traditional software development, code commits create a natural record. In AI development, contributions are often informal, iterative, and undocumented. That makes “who contributed what” almost impossible to establish after the fact.
Stop me if this sounds familiar: two companies partner on an AI pilot, both contribute resources and personnel, the project produces something valuable, and then they argue about who owns it because no one thought to write a joint ownership agreement before they started.
Fix it before you start. Not after. A joint IP agreement should specify: contribution tracking requirements, independent licensing rights (or restrictions), revenue sharing, what happens when a party wants out, and who controls defensive patent filings.
Rule 07International Divergence — and Why It’s Your Problem Now
The U.S., EU, and China have taken meaningfully different approaches, and if your company operates across borders — or just licenses content internationally — this matters more than most people realize.
The U.S. position is clear after Thaler v. Perlmutter: no human authorship, no copyright. Full stop.
China took the opposite approach in 2023. A Beijing court recognized copyright protection for an AI-generated image, on the grounds that it reflected a human’s intellectual effort — the prompting, selection, and curation. Tier 2 — one district-level ruling; not binding national precedent That’s not established national precedent, but it signals how Chinese courts are thinking about this.
The EU, under the AI Act and existing copyright directives, maintains human authorship requirements but has text-and-data-mining exceptions for commercial AI training that are narrower than what U.S. companies have been operating under. European rightsholders have stronger tools to opt out of AI training than their American counterparts.
The UK is genuinely weird here. British law has a provision — Section 9(3) of the Copyright, Designs and Patents Act 1988 — that specifically recognizes copyright in “computer-generated works,” attributing it to the person who arranged for the work’s creation. Some U.K. commentators have argued this could cover AI outputs. But the courts haven’t confirmed it in an AI context, and the Getty v. Stability AI case in 2025 — which the U.K. court largely rejected on evidentiary grounds — didn’t resolve the underlying question. The result: the U.K. is operating without a meaningful legal answer on AI training data, which is itself a problem.
Practical takeaway: if you’re registering IP internationally, check the requirements jurisdiction by jurisdiction. U.S. copyright registration requires human authorship. In China, your prompting process might be enough. In the EU, your training data practices might create liability even if you’re fine under U.S. law.
The documentation gap is now a liability gap
The legal risk here isn’t just theoretical — it’s operational. Your company is almost certainly generating AI outputs that lack the documentation needed to defend copyright ownership. The question isn’t whether to document human creative input; the 2025 decisions settled that. The question is when you build that process.
What you do: Audit current AI tool usage across all departments. Identify where AI-generated outputs are being used commercially. Build a minimum documentation standard: prompt logs, iteration records, human editorial decisions. This doesn’t require new software — a structured shared folder works. Make it part of onboarding for any team using AI tools.
Here’s what’s going to stop you: Creative teams find documentation friction annoying and will skip it without enforcement. The fix isn’t persuasion — it’s building it into the workflow tool. If the team is using Figma or Notion, the log lives there. If it’s ChatGPT Enterprise, configure it to export sessions.
Stop doing this: Treating platform indemnification as a substitute for copyright documentation. Enterprise-tier indemnification covers claims arising from the platform’s own training data — not from your organization’s failure to establish human authorship in the output. Those are different things. Many in-house teams conflate them.
Your AI-assisted IP is only as strong as your invention records
If you’re building with AI tools in R&D and you haven’t updated your invention disclosure process, the November 2025 USPTO guidance just created a specific audit obligation. Patents derived from AI-assisted research need documented human conception — not just the AI’s output, but evidence that a natural person understood, selected, and built on that output with a specific, operative invention in mind.
What you do: Talk to patent counsel before your next filing, not after. Specifically: who at your company is the named inventor, what did they actually conceive, and what’s the documented record? For drug discovery and materials science especially — where AI is generating candidate structures and a human is selecting among them — the selection-and-understanding record is your inventorship evidence.
Here’s what’s going to stop you: This is harder in fast-moving R&D environments where discovery and documentation are treated as separate phases. The discipline required is logging what the AI surfaced versus what the human decided to pursue — and why. Teams that don’t build this habit in year one are doing archaeology in year three when the patent gets challenged.
Stop doing this: Treating AI-assisted discovery as inherently patentable by default. The USPTO guidance is explicit: using an AI tool doesn’t invalidate a patent application, but not being able to identify a human inventor who conceived the claimed invention does. “Our AI found it” is not an inventorship record. “Our scientist reviewed the AI-generated candidates, identified compound X as promising based on mechanism Y, and designed the test protocol” is starting to be one.
The law here is genuinely unsettled in some areas and firmly settled in others. Courts have been unambiguous on human authorship for copyright. The USPTO has been clear on human conception for patents. Training-data liability — 52 active lawsuits, no definitive fair-use ruling as of this writing — is still being worked out in real time.
What that means practically: the companies that will be best positioned when this shakes out are the ones that built documentation habits now, when it’s cheap, rather than retroactively, when it costs a lot.
One more thing — and I say this because I’ve watched people skip it: get actual legal counsel with AI specialization. Not a generalist, not a blog. The frameworks are moving fast enough that the right answer in Q1 2025 might not be the right answer in Q4 2025, and the USPTO literally scrapped and replaced its guidance mid-year. That’s not normal.
Legal disclaimer: This article is for informational purposes only and does not constitute legal advice. AI IP law is evolving rapidly; specific situations require qualified legal counsel. Sources cited represent the author’s research as of April 2026; check for subsequent developments before relying on any legal principle described here. Internal links to BestPrompt.art are provided for navigation. External links are provided for source attribution only.
