Human Authorship in 2025: Who Really Owns It?
Legal Analysis · April 2025 · ~2,300 words
Copyright law has a clear answer right now. It’s probably not the one you’re hoping for. Here’s what the actual rulings say, what remains genuinely unsettled, and what creators can do today to protect their work.
The short version
- US copyright law currently does not protect purely AI-generated work. At all.
- The Zarya of the Dawn case (2023) established that human-selected, human-arranged elements can be protected — even when AI generated the raw images.
- “Substantial human contribution” is the operative phrase. No court has quantified it. That ambiguity is your main risk.
- The Getty v. Stability AI case, still pending in Delaware, will likely be the most consequential decision for training data rights in the next two years.
- Three things you should do right now: document your creative process, read the TOS of every AI tool you’re using, and don’t assume “edited” equals “owned.”
Copyright law in 2025 is having a bad time. The rules were written for a world where humans picked up pens, or cameras, or keyboards. Now they’re being applied to a world where an author spends six hours prompting, selecting, arranging, and editing content that a model generated in four seconds. The law hasn’t caught up. Rulings are coming out yearly. The practical situation for working creators is murkier than anyone advertising AI tools will tell you.
So let’s get into what’s actually known.
Start with the baseline. The US Copyright Act protects “original works of authorship.” The Copyright Office and federal courts have consistently interpreted “authorship” to require a human. Not a company. Not a machine. A human.
This was tested directly in Thaler v. Vidal — a case that ran through the Federal Circuit in 2022 — where Stephen Thaler argued his AI system DABUS should be listed as an inventor on a patent. The court said no. Same principle, patent law, but the logic applies to copyright: authorship and inventorship require a human being. The Copyright Office’s March 2023 guidance document makes this explicit for copyright specifically: works produced entirely by AI with no human creative control are not eligible for protection.
The Zarya of the Dawn case — what it actually established
Kristina Kashtanova created a graphic novel called Zarya of the Dawn using Midjourney-generated images. The Copyright Office initially granted, then partially rescinded, her registration. The final ruling (February 2023) protected her written text and the selection and arrangement of images — but not the individual AI-generated images themselves. Source: US Copyright Office correspondence, February 21, 2023 — publicly available
This distinction matters a lot. It means the curation of AI content can be protected. The raw AI output cannot. If you’re selecting, sequencing, combining, and making genuinely creative choices about AI-generated material — that selection may be protectable. The individual outputs probably aren’t.
Courts outside the US are landing in different places. The UK explicitly denies copyright to AI-generated works unless a human author can be identified. China’s Beijing Internet Court ruled in late 2023 that AI-generated images can receive copyright protection if the user exercised “creative selection and arrangement” — which is actually closer to the US Zarya logic than the US approach is sometimes characterized. Tier 2 — Beijing Internet Court ruling, November 2023, reported by CNBC and Reuters
The patchwork here is real. If your work has international commercial potential, this matters.
“Copyright law protects the selection and arrangement. It does not protect the AI output itself. These are not the same thing.”
Editorial synthesis — sources: US Copyright Office guidance (March 2023), Zarya of the Dawn correspondence (February 2023)
Getty Images sued Stability AI in January 2023 in both the UK and Delaware federal court, alleging Stability AI scraped over 12 million Getty-licensed images to train Stable Diffusion without compensation or authorization. Tier 2 — court filings, Delaware District Court, Case 1:23-cv-00135; confirmed by Getty press release
The Delaware case is still pending as of early 2025. It’s arguably the most consequential active copyright case in the AI space because it tests whether training on licensed images without permission constitutes infringement — a question that will affect every generative image model in existence.
Getty’s complaint includes something specific that’s worth flagging: some Stable Diffusion outputs apparently reproduce Getty’s watermark. That’s not a metaphor for resemblance — it’s a literal rendering of their branded watermark on generated images, which suggests the model memorized specific training images closely enough to reproduce identifying marks. If that’s substantiated, it shifts the case from “is training fair use?” toward something more like direct copying claims.
Second-order mechanism
The reason training data rights are hard to resolve isn’t just legal complexity — it’s that AI companies and copyright holders are measuring different things. Companies argue training is “learning” not “copying”; the model compresses and transforms rather than storing. Rights holders point to the watermark evidence and argue the distinction is thinner than claimed. The honest answer is that courts haven’t resolved this, and won’t until Getty v. Stability AI or a comparable case reaches a final ruling.
Three other active cases are worth tracking: Andersen v. Stability AI (class action by visual artists), Authors Guild v. OpenAI (filed 2023, alleging training on copyrighted books without authorization), and NYT v. OpenAI (filed December 2023, New York Times alleging its articles were reproduced and paraphrased in ChatGPT outputs). None have reached final judgments. The legal landscape will look different in 18 months.
This phrase appears everywhere in Copyright Office guidance. It’s never been quantified. No court has said “40% editorial input is enough” or “three hours of revision creates ownership.” That ambiguity is not an accident — courts don’t want to set a numeric threshold that could be gamed — but it leaves creators exposed.
What the Copyright Office’s guidance does say: the relevant question is whether the human made “creative choices” about the expression, not just whether they provided input. Typing “a watercolor portrait of a sad dog” is input. Selecting from 40 AI-generated portraits, cropping, adjusting color balance, adding a background, and integrating the result into a larger composition is closer to what courts are looking for.
The distinction that’s emerging from registrations the Copyright Office has accepted vs. rejected: prompting alone is probably not enough. Selecting, arranging, transforming, and combining — especially when those choices require taste and expertise — has a better chance. The more your creative decision-making is legible in the final output, the stronger your position.
| Scenario | Copyright Likely? | Legal Basis | ⚠ Caveat |
|---|---|---|---|
| You type a prompt; AI generates an image; you use it as-is | No | Copyright Office guidance, March 2023: AI-only output not protectable | Applies in US. EU and UK have similar logic. China is more permissive under some interpretations. |
| You prompt, select from many outputs, arrange into a larger work with significant editorial choices | Possibly — for selection/arrangement | Zarya of the Dawn precedent; selection and arrangement protectable | The underlying AI images remain unprotected. Your arrangement is protected. Distinction matters for enforcement. |
| You write a novel using AI for drafts, then substantially rewrite and restructure | Likely yes — for human-authored elements | Copyright protects original expression; significant rewriting creates protectable authorship | “Substantial” has not been defined numerically. Evidence of your revision process matters. Keep drafts. |
| You use an AI tool whose ToS grants the company broad licensing rights to outputs | Complicated | Contract law, not copyright; ToS may limit your rights even if you hold copyright | OpenAI’s current ToS (updated January 2024) assigns output ownership to users but retains broad rights for service improvement. Read it. |
Here’s something most “protect your AI work” articles skip. Several publishers, academic journals, and platforms now require disclosure of AI tool use. Etsy banned AI-generated art from its “handmade” category in 2023. Amazon Kindle requires disclosure of AI-generated content. Many academic journals now require explicit AI authorship statements.
So if you disclose AI involvement, you may lose commercial access on some platforms. If you don’t disclose, you may violate platform ToS, breach academic integrity standards, or create legal exposure around misrepresentation.
That tension doesn’t have a clean resolution right now. The practical answer is: know each platform’s rules before publishing, not after. Some distinguish between AI-generated (primary content produced by AI) and AI-assisted (human-authored work that used AI tools in the process). That distinction is doing a lot of work, and different platforms draw the line in different places.
Cross-source synthesis — not present in any single cited source
Reading the Copyright Office guidance, the Zarya correspondence, and the platform policy changes together produces a finding none states individually: the legal framework and the commercial infrastructure are moving in opposite directions. Copyright law is trending toward “human selection and arrangement can be protected even with AI tools.” But commercial platforms are trending toward disclosure requirements and categorical restrictions that may make that protection commercially worthless if the work can’t be distributed on the platforms where revenue is generated. A creator might own copyright to an AI-assisted work and still be unable to sell it where their audience is.
You might own the copyright to an AI-assisted work and still be unable to sell it where your audience lives.
Editorial synthesis — sources: Copyright Office guidance (2023), platform policies (Amazon KDP, Etsy, academic journal requirements, 2023–2024)
Keep your prompts. Keep your rejected outputs. Keep your revision history. If you’re ever challenged on ownership, the question isn’t “did you use AI” — it’s “what did you contribute?” A folder of drafts showing your selection and editing choices is your evidence. This costs you nothing and matters a lot.
Read the ToS of every AI tool you’re using
This is not optional. OpenAI’s ToS assigns output ownership to users but retains rights to use that content to improve services. Midjourney’s free tier means the company holds a license to your outputs for promotional use. Adobe Firefly is trained on licensed content and explicitly designed for commercial use — different risk profile than Stable Diffusion trained on scraped data. These are not the same tool legally. Treat them differently.
Think about what you’re actually protecting
If your AI-assisted work has commercial value, consider what element of it you’re relying on. The specific images? Probably not protectable. The selection, sequence, and arrangement that makes your book, campaign, or product distinctive? Closer. Your written text, even if AI-drafted and substantially rewritten? Yes, probably. Know which part of your work you’re actually betting on.
For: Independent Creators (Writers, Visual Artists, Designers)
Your copyright situation is fragile, and most advice online is wrong
Look, the “40% human input means you own it” framing floating around creator forums is not from any legal ruling. It’s invented. The actual standard is qualitative, not quantitative: did you make creative expression choices? Your job is to make those choices legible — to yourself first, as documentation, and to any future dispute second.
What you do: Start a process log. A simple Google doc. Prompt you used → options you saw → what you chose and why → what you changed. Not for every project. For anything you’re planning to sell, license, or publish commercially. Six months from now this costs you nothing. In a dispute it could be the difference.
Here’s what’s going to stop you: it feels like bureaucracy for something creative. It is. Do it anyway. The platforms enforcing AI disclosure requirements are going to keep multiplying; the ones without policies today will have them in 18 months. You want your documentation habit in place before you need it.
Stop doing this: claiming you “own” AI-generated work because you paid for the subscription. The ToS gives you a license to use the output. That’s not the same as copyright ownership, and several of those licenses have commercial use restrictions that most people don’t read until they’re in a client dispute.
For: Creative Teams at Companies (Marketing, Content, Brand)
Your legal exposure is different from an individual creator’s — and probably bigger
Individual creators risk losing ownership of work they made. Companies risk something larger: infringing on someone else’s IP at scale, with institutional pockets worth suing. The Getty case wasn’t brought against individual Stable Diffusion users — it was brought against the company that trained the model. But if your team is using a model trained on unlicensed content and your outputs end up resembling source material, you’re in the chain.
What you do: build a vendor assessment into your AI tool procurement. Two questions: What data was this model trained on? Does the vendor provide indemnification for IP claims against outputs? Adobe Firefly answers both favorably. Many tools don’t answer the first question at all. The absence of an answer is itself information.
Here’s what’s going to stop you: procurement cycles are slow and legal review is slower. Start with the tools your team is already using. Run the two questions against each one this quarter. You don’t need a perfect policy; you need to know what you’re using and why.
Stop doing this: treating “AI disclosure” as a PR question rather than a compliance question. Some jurisdictions are moving toward mandatory AI disclosure requirements for commercial content. The FTC has signaled interest in deceptive AI use claims. Getting ahead of this is cheaper than responding to it.
The copyright question around AI isn’t going to be stable for a while. We’re probably 18–24 months from a few major rulings that will set clearer precedent. Until then, the operative strategy is documentation, ToS literacy, and understanding what part of your work is actually valuable and why.
Related: Prompt Engineering for Creators — full guide at BestPrompt.art
The law will catch up. It always does. Just probably not on your timeline.
https://www.bestprompt.art/ai-and-human-creativity/
https://www.bestprompt.art/the-ethical-use-of-ai-prompts/
https://www.bestprompt.art/ai-ip-rights/
https://www.bestprompt.art/ownership-of-ai-outputs/
