You asked ChatGPT to write a blog post, a product description, or a batch of social captions. It delivered polished copy in 30 seconds. You hit publish. But here is the question that rarely comes up until it matters: who owns that text, can you stop a competitor from copying it, and could you be infringing someone else without knowing it?
In 2026, the answer is clearer than it was two years ago — and less generous than most AI terms of service make it sound.
Why Copyright Still Requires a Human Author
U.S. copyright law does not protect everything that looks creative. It protects original works of authorship fixed in a tangible medium, and the U.S. Copyright Office has consistently read "authorship" to mean human authorship.
That principle has been tested at every level:
- The Copyright Office said no. Starting with the Zarya of the Dawn graphic novel decision in 2023, the Office found that images generated by Midjourney from text prompts were not copyrightable, even though the human applicant wrote detailed prompts and curated the output. The text and the human-arranged layout were registrable. The images alone were not.
- A federal court agreed. In Thaler v. Perlmutter (D.D.C. 2023), Judge Beryl Howell affirmed the Office, holding that human authorship is a "bedrock requirement of copyright" and that a work generated entirely by an AI system with no human creative input cannot be registered.
- The appeals court affirmed. In March 2025, the D.C. Circuit affirmed 2-0 in Thaler v. Perlmutter, 130 F.4th 1039, emphasizing that the Copyright Act as written requires a human author and that any expansion to machine authors must come from Congress, not the courts.
- The Supreme Court left it standing. In October 2025, the Supreme Court denied certiorari, leaving that human-authorship requirement intact.
The through line is simple: the machine can be a tool, but it cannot be the author.
What the Copyright Office Actually Said in January 2025
On January 29, 2025, the Copyright Office released Part 2 of its Copyright and Artificial Intelligence report, focused squarely on copyrightability. It is the most detailed guidance small businesses have, and its conclusions are practical rather than theoretical.
Prompts alone are not enough
The Office concluded that, given currently available generative AI technology, "prompts essentially function as instructions that convey unprotectable ideas" and "do not provide sufficient human control to make users of an AI system the authors of the output." Even highly detailed, iterative prompting does not produce the predictable control that copyright law associates with authorship.
That surprises many owners who spend an hour engineering a prompt. The Office's analogy is closer to hiring a commissioned artist and describing what you want than to painting the image yourself. Direction is not the same as expression.
Human contribution must be perceptible in the final work
Copyright can still attach when human creativity is perceptible in the output itself. The report analyzes this case by case and outlines three scenarios where protection remains available:
- AI as an assistive tool. Using AI to brainstorm, correct grammar, upscale an image, or remove a background does not strip a human-authored work of protection, as long as the human shapes the final expression.
- Human-authored work that references AI output. If you use AI to generate ideas or research summaries but write the final post yourself without copying AI text verbatim, the resulting work is fully yours.
- AI-generated material with perceptible human authorship. If you take AI output and creatively select, arrange, or modify it — rewriting passages, combining multiple outputs with original analysis, adding expressive editorial judgment — the human-contributed elements can be protected. The AI-generated portions themselves remain unprotected, but your contributions are.
In other words, the Office is not asking whether you used AI. It is asking what a reader can see on the page that came from your creative choices.
Existing law is enough — for now
The Office explicitly declined to recommend a new sui generis right for AI-generated content or a separate registration track for AI works. It found that existing doctrines — human authorship, originality, and the idea-expression distinction — are adequate and that every determination should be made on its individual facts.
What This Means for Your Marketing Content
Translate that law into day-to-day decisions for a small business that uses ChatGPT, Claude, Gemini, or similar tools:
1. Pure AI output has no copyright to enforce
If you paste an AI draft directly to your blog without meaningful human changes, you likely have no copyright in that text. That has two consequences:
- You cannot register it with the Copyright Office. In the Office's 2023 guidance on AI-generated material, applicants must disclose AI use and disclaim AI-generated portions that are more than de minimis. Misrepresenting the work can invalidate the registration.
- You cannot easily stop a competitor from copying it. Copyright's infringement remedy depends on owning a valid copyright. If the text was never protectable, copying it — even verbatim — is not copyright infringement, though it may still violate platform terms or constitute unfair competition in narrow cases.
The common misconception comes from AI terms of service. OpenAI's terms, for example, state that you own the output and assign any rights the company might have to you. That assignment matters for the contract between you and OpenAI, but it does not create a copyright where the law says none exists. Ownership of a file and ownership of a copyright are different things.
2. Light edits are not the same as creative authorship
Adding a headline, fixing typos, or reordering two paragraphs is unlikely to cross the threshold of perceptible human authorship. The Office looks for expressive choices — the kind of decisions a writer or designer makes about voice, structure, examples, argument, and curation.
Practical steps that do add protectable authorship:
- Substantially rewriting AI prose in your own voice, adding original examples from your business, client work, or data you collected
- Combining research from multiple sources with your own analysis, conclusions, or recommendations rather than summarizing a single AI answer
- Creating original selection and arrangement, such as curating a comparison chart, adding custom photography or illustrations you directed, or building a narrative arc the AI did not propose
Think of AI as a first-draft intern: useful for speed, but the copyright lives in the revision, not the prompt.
3. Your infringement risk runs in the other direction, too
Small businesses tend to worry only about being copied. The larger legal risk in 2026 is unknowingly publishing AI output that itself copies someone else.
Generative models are trained on vast corpora that include copyrighted books, articles, lyrics, and images. Several high-profile lawsuits allege that training and output reproduce protected works:
- The New York Times v. OpenAI and Microsoft, filed in late 2023, alleges that ChatGPT was trained on millions of Times articles and can generate near-verbatim excerpts. In March 2025 a federal judge denied OpenAI's motion to dismiss, allowing the core infringement claims to proceed.
- Parallel suits by authors, publishers, and image creators raise the same question: does training on copyrighted works without permission, or generating outputs that are substantially similar to those works, constitute infringement?
Courts have not yet produced a definitive fair-use answer that applies to every model and every use, and different jurisdictions are reaching different preliminary conclusions. An Indian court's July 2025 interim order in ANI v. OpenAI, for example, treated training as falling under fair dealing on those facts, while U.S. cases continue to litigate the issue. That patchwork means a small business cannot safely assume that AI output is automatically clear.
The practical takeaway is not to panic, but to adopt verification habits similar to any other outsourced writing:
- Do not publish AI output that contains long quoted passages, distinctive phrasing, or factual narratives that read like they were lifted from a single source. Run a quick web search on suspiciously polished quotations or statistics.
- Treat AI citations as unreliable until you check them. Models hallucinate sources and can invent quotations that look authentic.
- Keep records of your prompts, the raw output, and your edits. That paper trail helps demonstrate independent creation and the scope of your human contribution if ownership is ever questioned.
A Practical Ownership Checklist for Small Businesses
Before you publish AI-assisted content, run through these five questions:
Who actually wrote the final sentences?
If the answer is "the model, with minimal touch-up," assume no copyright. If you can point to paragraphs, transitions, examples, or visual choices that reflect your judgment, those portions are protectable even if the underlying AI draft is not.
Did you disclose AI use where required?
The Copyright Office requires disclosure of AI-generated material that is more than de minimis in an application. If you ever register a work that incorporated AI, be prepared to describe what is human-authored and disclaim the rest. Accuracy at registration affects enforceability later.
What do the tool's terms say about input and confidentiality?
Most commercial tools grant you ownership of output but retain a license to use your inputs to improve services unless you opt out, and enterprise tiers may offer different data-use commitments. If you paste client data, unreleased product details, or employee information into a public chatbot, you may be making a disclosure you cannot undo. Use business or enterprise settings that commit to not training on your data, or keep sensitive prompts local.
Could this output be substantially similar to someone else's work?
For text, images, and code, ask whether the output reproduces distinctive expression rather than just facts or ideas. Facts, titles, and short phrases are not copyrightable, but extended verbatim or near-verbatim reproduction of expressive content can be. When in doubt, rewrite with original phrasing and cite the idea rather than copying the expression.
How will you prove authorship later?
Copyright disputes often turn on evidence created long before the dispute. For important assets — a flagship guide, a course, a brand book — save version history showing your drafts, AI outputs, and human revisions. A dated, version-controlled record is far more persuasive than memory.
The Bookkeeping Angle: Track What You Spend to Create Content
Content creation has real costs that belong in your books, especially if you are investing in AI-assisted production at scale:
- Separate content costs from general software. Subscriptions to generative AI tools, stock image licenses for human-created visuals, freelance editing fees, and registration fees are distinct expenses. Booking them to a single "software" account obscures the true cost per asset.
- Capitalize where appropriate. If you commission a substantial evergreen resource that will generate leads for more than a year — a detailed guide, a video series, or a registered copyright — discuss with your accountant whether it should be treated as a capitalized intangible and amortized rather than expensed immediately.
- Log licenses and permissions. Keep a simple register of every image, font, music cue, and data source you paid to license, with renewal dates. AI-era infringement claims often start with "we never licensed that image" — a register prevents that scramble.
- Document contractor contributions. If a freelancer substantially revises AI output on your behalf, your agreement should state in writing that the work is a work made for hire and that copyright is assigned to your company. Without that language, the freelancer may retain rights even if you paid for the work.
Clear records do not just help at tax time. They establish the chain of title that makes your copyright enforceable and your expense deductions defensible.
Bottom Line: Use AI for Speed, Keep Human Judgment for Ownership
The law in 2026 rewards the same behavior that produces better content for your audience: use generative AI to accelerate research and drafting, then add the human decisions that make the piece unmistakably yours — original examples, lived experience, curated structure, and a point of view that reflects your expertise.
If you do that, two good things happen at once. Your readers get something a model alone could not produce, and your business retains the copyright that lets you control how that work is reused.
If you publish raw AI output without that layer, you gain speed but trade away enforceability and inherit the infringement risk baked into any model trained on the open web. For a blog post that drives SEO and lead flow, that is a lopsided bargain.
Keep Your Content and Your Books Aligned
As you systematize AI-assisted content production, systematize the financial tracking that goes with it. Separating content costs, license fees, and contractor assignments in your books makes tax categorization cleaner and gives you the documentation you need if ownership is ever questioned.
Beancount.io provides plain-text, version-controlled accounting that keeps that history transparent — every entry dated, every change traceable, and every report reproducible without a black box. If you want financial records that are as auditable as the content trail you just built, get started for free and see why teams that care about provenance choose plain-text accounting.