The U.S. Copyright Office issued a landmark ruling in the case of Zarya of the Dawn — a graphic novel written by Kristina Kashtanova and illustrated using Midjourney. The office granted copyright to the text and the arrangement of images Kashtanova chose, but cancelled protection for every individual AI-generated image. The images, it concluded, were not the product of human authorship and therefore could not be owned.
American copyright law, codified in Title 17 of the U.S. Code, has required human authorship since at least 1884, when the Supreme Court decided Burrow-Giles Lithographic Co. v. Sarony. That case established that a photograph of Oscar Wilde could be copyrighted because the photographer made creative choices — lighting, pose, expression. The law has never extended protection to works produced by nature, animals, or machines acting autonomously.
The Copyright Office's 1973 COMPENDIUM (updated continuously since) states flatly: "The Office will not register works produced by nature, animals, or plants." The 2023 guidance added AI systems to that list when the AI's contribution is the creative content itself.
This creates an immediate puzzle for creators using tools like Midjourney, DALL-E, Stable Diffusion, or ChatGPT: the more expressive the AI's output, the less likely you can own it.
Thaler v. Perlmutter (D.D.C. 2023): Computer scientist Stephen Thaler applied to register a painting called "A Recent Entrance to Paradise," listing his AI system DABUS as the sole author. Federal District Court Judge Howell upheld the Copyright Office's refusal, writing that "human authorship is a bedrock requirement of copyright." Thaler's appeal is ongoing as of mid-2025.
Copyright protects original works of authorship fixed in a tangible medium. "Original" means it originated from a human author and shows at least a minimal degree of creativity — the bar is low, but it is not zero and it requires a human.
What copyright does NOT require: that you wrote every word or drew every stroke. A film director doesn't hold a camera or compose music, yet controls the copyright in the film. An architect's copyright covers the building design even if contractors did all the physical work. The creative decisions — what to include, how to arrange, which elements to emphasize — constitute authorship.
This distinction matters enormously for AI users. If you make the expressive choices and use AI as a tool (like using Photoshop's filters), you likely have a copyright claim. If you type a five-word prompt and press Enter, the Copyright Office's current position is that you probably don't.
Selecting and arranging AI-generated pieces into a collage · Writing surrounding text · Editing and transforming raw AI outputs · Choosing which of 200 generations to include and how to sequence them · Adding original artistic elements on top of AI output
The raw image from a short text prompt · The verbatim story ChatGPT wrote when you asked for one · A song generated entirely by Suno from a one-line description · Any output where the AI made all the creative choices autonomously
If your AI-generated work lacks copyright protection, it enters the public domain immediately. Anyone can copy, sell, or remix it without your permission and without paying you. This has significant consequences for creators trying to monetize AI-assisted work — illustrators, musicians, game designers, and content creators building businesses on AI output.
The law is still in flux. Several cases are winding through U.S. courts, and Congress has held hearings specifically on AI and copyright. The European Union's AI Act and its copyright directive approach the problem somewhat differently. We'll examine those differences in Lesson 3.
The more human creative decision-making you can document and demonstrate — prompts, iterations, edits, selections, arrangements — the stronger your potential copyright claim on AI-assisted work. Document your process.
You're going to present creative scenarios to the AI lab assistant and ask it to evaluate whether the human creator likely has a copyright claim — and why. Push back, ask follow-up questions, and test the edges of the rule.
In 2022, when Stability AI launched Stable Diffusion as open-source software, users were surprised to discover it had been trained on billions of images scraped from the internet — including works by living artists who had never consented. Artists Sarah Andersen, Kelly McKernan, and Karla Ortiz filed a class-action lawsuit in January 2023 against Stability AI, Midjourney, and DeviantArt, alleging copyright infringement in the training data itself. The case, Andersen v. Stability AI, is still moving through Northern District of California courts.
Meanwhile, the platforms these tools run on were quietly writing contracts that shaped who owned everything created on them. Most users clicked "I agree" without reading a word.
Every major AI creative platform has terms of service (ToS) governing what happens to your inputs (prompts, uploaded files, reference images) and outputs (the generated content). These terms vary significantly — and they change frequently.
Getty Images filed suit in both the U.S. and UK, alleging Stability AI scraped and trained on over 12 million Getty images without a license, including images with Getty watermarks visible in some AI outputs. The UK case is proceeding; in the U.S., Judge Stephanos Bibas allowed the copyright infringement claims to proceed in August 2023. This is the first major case to directly test whether training on copyrighted images constitutes infringement.
Many AI platforms treat your prompts and uploaded content as training data for future model versions. In April 2023, Samsung engineers accidentally leaked proprietary chip design data through ChatGPT when they used it to debug code — the conversations were stored and potentially visible to OpenAI staff. Samsung subsequently banned ChatGPT for internal use.
OpenAI and other providers typically offer opt-out mechanisms for enterprise users, but consumer accounts often default to opt-in for training data use. Practically: do not put trade secrets, client data, or personal information into free AI tools.
When you use AI-generated output commercially and a third party sues claiming the output infringes their copyright, who covers your legal costs? Most platforms offer no indemnification to free users. Adobe and Microsoft (Copilot) have made commercial indemnification explicit selling points for enterprise clients.
In February 2024, Microsoft launched its "Copilot Copyright Commitment" — promising to defend customers sued for copyright infringement from Copilot outputs, provided the customers used the built-in guardrails. Adobe made a similar commitment for Firefly Enterprise. These commitments do not apply to consumer-tier products.
Before using any AI tool commercially: (1) Read the current ToS — they change. (2) Check whether commercial use requires a paid tier. (3) Understand what license you grant to the platform. (4) For high-stakes commercial work, consider platforms offering explicit IP indemnification (Adobe Firefly Enterprise, Microsoft Copilot).
You're a freelance designer considering using AI tools for client work. Ask the assistant to help you evaluate which tools are safer for commercial use, what questions to ask platforms, and how to protect yourself contractually when delivering AI-assisted work to clients.
Three visual artists — Sarah Andersen, Kelly McKernan, and Karla Ortiz — filed suit against Stability AI, Midjourney, and DeviantArt. Their complaint ran to forty-four pages. The central argument: their life's work had been scraped from the internet without consent, compressed into model weights, and was now being commercially exploited to generate images in their styles — replacing demand for their actual labor.
In the same month, Getty Images filed separately, noting that some Stable Diffusion outputs displayed distorted versions of Getty's watermark — apparent evidence that images had been reproduced so faithfully that even the watermark transferred.
When an AI company says it "trained" on a dataset, it means the model processed millions or billions of examples to learn statistical patterns — which pixel colors follow which, which words follow which, which musical notes follow which. The question before the courts is: does ingesting a copyrighted work to extract statistical patterns constitute the kind of "copying" copyright law prohibits?
There are two serious legal theories in play:
Fair Use (for AI training): U.S. copyright law permits "fair use" of copyrighted material in certain circumstances. Courts weigh four factors: (1) the purpose and character of the use — is it transformative? (2) the nature of the copyrighted work, (3) the amount used, and (4) the effect on the market for the original. AI companies argue training is "highly transformative" — no specific work is reproduced, only patterns are learned. Critics argue the fourth factor weighs heavily against AI: if AI output can substitute for original human work, it harms the market.
The "memorization" problem: Research by researchers at Google DeepMind and elsewhere has shown that large language and image models can sometimes reproduce training data nearly verbatim when prompted in specific ways. This undercuts the "no reproduction occurs" argument.
The New York Times filed suit against OpenAI and Microsoft alleging that GPT-4 was trained on millions of Times articles without authorization. The complaint included striking examples where GPT-4 could reproduce verbatim passages from Times journalism when prompted. The Times argued this both infringes copyright and substitutes for Times content (the fourth fair use factor). The case is ongoing and widely considered the most consequential AI copyright case in U.S. history.
Copyright does not protect artistic style — only specific expression. You cannot copyright "impressionist painting" or "cyberpunk aesthetic." However, when AI systems can generate images "in the style of [living artist]" with enough fidelity to suppress market demand for that artist, many creators argue the legal framework fails to provide adequate protection even when technically correct.
In August 2023, the Authors Guild — representing over 13,000 authors — sent a letter to major AI companies asking them to seek permission and pay licensing fees for training data. Signatories included Margaret Atwood, Roxane Gay, and Dan Brown. The Authors Guild filed a class action in September 2023, alleging OpenAI trained on hundreds of thousands of copyrighted books without authorization.
Several companies have created opt-out mechanisms for creators:
Google: Added an opt-out meta tag standard (robots.txt-style) that websites can use to signal they don't want their content used for AI training. Critics note this puts the burden on creators, not scrapers.
DeviantArt: Added an "noai" tag system to let artists flag work as off-limits for training. Effectiveness depends entirely on whether AI companies honor the tags.
Spawning's "Have I Been Trained?" tool: Lets artists check if their work appeared in LAION-5B (a major training dataset) and submit opt-out requests. As of 2024, Stability AI has committed to honoring these requests for future model versions.
The fundamental tension: opt-out systems mean consent is assumed unless you act, which inverts most people's intuitions about property rights. The EU's approach in the AI Act and Digital Single Market Directive requires opt-in licensing for certain AI training use — a significantly different baseline.
Opt-out default: scraping is presumed acceptable unless content owner objects. Fair use doctrine may protect training. Courts still deciding. Burden on creators to protect themselves.
Article 4 permits text and data mining unless rights holders have "reserved" their rights. Research and cultural institutions have broader TDM rights. More creator-protective but complex in practice.
If you are a creator: use available opt-out mechanisms (Spawning, noai tags, robots.txt). If you are an AI user building commercial products: understand that the tools you use may be entangled in litigation whose outcomes could affect your rights to those outputs retroactively.
You're going to steelman both sides of the training data debate. First, build the strongest possible argument that AI training on copyrighted work is fair use. Then, build the strongest case against it. The assistant will push back and deepen your reasoning.
After the Zarya of the Dawn decision, the Copyright Office published a broader statement of policy: AI-generated content with "sufficient human authorship" could be registered, but applicants must disclose the use of AI and describe the human creative contribution. The office was deliberately not drawing a bright line — it would evaluate cases individually.
This created both an opening and a strategy: document the human creative choices. The more a creator could point to specific decisions — which of 300 outputs to select, how they were arranged, what was edited, what original elements were added — the stronger the registration case.
The single most valuable thing you can do to protect AI-assisted creative work is document your process in real time. This means:
Save prompt iterations: Keep a record of the prompts you tried, in order. A hundred-word highly specific prompt demonstrates more creative input than a five-word one. Screenshot or export your conversation history.
Keep rejected outputs: If you generated 200 images and chose 12 for your book, the selection process itself is a creative act. Save evidence you made those choices. A folder of 200 images with 12 flagged is documentation.
Record your edits: Every Photoshop layer, every text edit, every element you added on top of the raw AI output is evidence of human creative contribution. Save layered files, not just flattened outputs.
Write a process note: A short written description of how you created the work — what AI tools you used, what you decided, what you changed — is useful both for Copyright Office registration and for defending against infringement claims.
The Copyright Office has begun processing registrations for AI-assisted works. As of 2024, applicants must: (1) disclose which elements were AI-generated, (2) describe the human authorship being claimed, and (3) disclaim the AI-generated portions. This creates a patchwork registration covering only the human-created elements. Registration is still highly recommended — it establishes a public record and is required to sue for infringement in U.S. federal court.
Several technical standards are emerging for proving the provenance of creative work:
C2PA (Coalition for Content Provenance and Authenticity): An open technical standard backed by Adobe, Microsoft, Google, Intel, and others. C2PA embeds cryptographically signed metadata into image, audio, and video files recording when they were created, with what tools, and by whom. Adobe's Content Credentials system implements C2PA — when you export from Photoshop with Content Credentials enabled, a verifiable record is attached.
OpenAI's DALL-E 3 watermarking: DALL-E 3 images include metadata indicating AI generation. As of 2024, OpenAI has begun testing invisible watermarking in generated images. These can potentially survive basic editing operations.
The limitation: Metadata can be stripped. Any JPEG saved through a basic converter loses embedded metadata. Watermarking is a provenance tool, not an enforcement tool — it helps you prove what you made, not prevent others from copying it.
When delivering AI-assisted work to clients, your contract matters more than any copyright claim. Consider:
Disclosure clauses: Many clients now require disclosure of AI use. Failing to disclose when asked may constitute misrepresentation. Build disclosure into your client agreements proactively.
Warranty limitations: Rather than warranting that your work is 100% original (difficult to guarantee with AI), warrant that you have the rights to use the tools you used, that you've complied with applicable ToS, and that you're not aware of specific third-party claims. This is more honest and legally defensible.
Indemnification caps: If a client insists on IP indemnification for AI-assisted work, negotiate a cap tied to the project fee rather than open-ended indemnification. The legal exposure on AI work is genuinely uncertain.
The legal landscape is changing fast. As of mid-2025, the following are in play:
Perfect protection doesn't exist yet, but imperfect protection is still valuable. Register what you can. Document your process. Choose your tools carefully. Use contracts wisely. Monitor developments — this area of law is moving faster than any other in the creative industries, and the rules that apply today may change significantly within a year.
You've been commissioned to create a series of AI-assisted illustrations for a published book. Work with the assistant to build a complete protection strategy — documentation, copyright registration approach, contract terms, and disclosure practices. Make it specific to your actual situation if you have one.