Copyright questions around AI-generated and AI-assisted art are often framed as a single yes-or-no issue: is the image protected or not? That framing is misleading. Copyright offices and courts are not evaluating how an image looks. They are evaluating how it was made, who made the decisions that produced it, and whether those decisions meet a legal threshold called originality. Two images that are visually indistinguishable can have completely different copyright status depending on the process behind them.
This article summarizes the current state of the law in the United States, the European Union, and the United Kingdom as of 2025, along with the separate and frequently confused issue of training data. None of this constitutes legal advice. Copyright law is being actively litigated and amended in multiple jurisdictions, and creators with a specific commercial dispute or registration question should consult an intellectual property attorney licensed in their relevant jurisdiction.
Why Copyright Status Depends on Process, Not Appearance
Copyright protects "original works of authorship fixed in a tangible medium," to use the language of United States law. Two elements matter here: authorship, meaning a human being made creative choices, and originality, meaning those choices were not purely mechanical or dictated by the tool. A photograph is protected not because a camera exists, but because a photographer chose framing, timing, lighting, and composition.
Generative AI systems complicate this because the final pixel arrangement often results from a statistical process the user does not directly control. If a creator types a short prompt such as "a fox in a forest, watercolor style" and accepts the first output, the legal analysis in most jurisdictions treats the human contribution as too thin to support authorship over the image itself. If the same creator generates fifty variations, selects one, inpaints specific regions by hand, adjusts color curves in post-processing, and composites elements from multiple generations into a new layout, the human contribution is substantially different, even though the starting technology is identical.
This is why appearance is not the test. A painterly, highly detailed AI image can receive no protection, while a simple AI-assisted sketch with heavy manual revision can qualify. The deciding factor is documented creative control, not visual sophistication.
Current Rules in the United States
The U.S. Copyright Office has stated consistently, including in its Compendium of U.S. Copyright Office Practices, Third Edition (Section 313.2), that copyright protection requires human authorship. Works produced without any human creative input are not registrable.
Two cases clarify the practical boundaries. In Thaler v. Perlmutter (D.D.C. 2023, affirmed on appeal in 2025), the Office and the courts rejected an application listing an AI system as sole author, reinforcing that a machine cannot hold a copyright. In the Zarya of the Dawn decision (February 2023), the Office granted registration for the written text and the overall selection and arrangement of images in a graphic novel, but refused protection for the individual AI-generated images themselves, because the applicant had used Midjourney and could not demonstrate sufficient control over each image's specific appearance.
Following this, the Office issued formal guidance in March 2023 requiring applicants to disclose the use of AI-generated content in new applications and to identify which portions were human-authored. Failure to disclose can result in cancellation of a registration after the fact. A further Copyright Office report, released in stages through 2025, confirmed this human-authorship requirement while leaving open how much editing or arranging is "enough" to qualify, a question decided case by case.
Current Rules in the European Union and United Kingdom
The European Union has no AI-specific copyright statute addressing authorship of AI outputs. Instead, the existing originality standard applies, built mainly from Court of Justice of the European Union rulings. In Infopaq International v. Danske Dagblades Forening (2009), the Court held that protection requires the work to be "the author's own intellectual creation." In Painer v. Standard Verlags
The United Kingdom is unusual in having an explicit statutory category for works with no human author. Section 9(3) of the Copyright, Designs and Patents Act 1988 assigns authorship of a "computer-generated work" to "the person by whom the arrangements necessary for the creation of the work are undertaken." This provision predates modern generative AI by decades and was written with earlier computer-assisted works in mind. Protection under this section lasts 50 years from creation, not the standard life-plus-70-years term applied to human-authored works. The UK government ran a consultation on AI and copyright in 2022 and again in late 2024 into 2025, and as of this writing no legislative change has been finalized, so Section 9(3) remains the operative, if contested, framework.
| Jurisdiction | Human authorship required for full-term protection | Special provision for no-human-author works | Term for AI-only output (if protected) |
|---|---|---|---|
| United States | Yes | No; unprotected if no human author | Not applicable |
| European Union | Yes, per CJEU case law | No EU-wide statute | Not applicable |
| United Kingdom | No, under Section 9(3) | Yes, authorship assigned to the arranger | 50 years from creation |
The Separate Question of Training Data
Authorship of an output is a different legal question from whether the training process that built the model was lawful. A model can be trained on copyrighted images without permission, and the resulting outputs can still raise authorship questions independently.
Several lawsuits address the training side specifically. Getty Images v. Stability AI, filed in both the UK and the US, alleges that millions of Getty's licensed photographs were used without authorization to train Stable Diffusion. Andersen v. Stability AI, Midjourney, and DeviantArt, a class action filed in California in 2023, makes similar claims on behalf of visual artists. The New York Times v. OpenAI and Microsoft, filed in late 2023, focuses on text rather than images but tests the same underlying question: does training on copyrighted material without a license constitute infringement, or does it qualify as fair use?
The jurisdictions differ structurally on this point. The EU's Digital Single Market Directive (2019/790), Articles 3 and 4, provides a text and data mining exception that permits mining copyrighted works for research and, under Article 4, for commercial purposes as well, unless the rights holder has explicitly reserved their rights in a machine-readable format. The United States has no equivalent statutory exception; training-data disputes there are litigated under the four-factor fair use test in 17 U.S.C. § 107, and as of 2025 no U.S. Court of Appeals has issued a final, binding ruling settling whether large-scale AI training qualifies as fair use. Creators should treat this area as unsettled rather than resolved in either direction.
Documenting Human Input for Legal Purposes
Because copyright offices evaluate process, not appearance, creators who want the strongest possible claim to authorship should keep records while they work, not after a dispute arises.
- Keep layered working files. A Procreate, Photoshop, or Krita file with visible layer history is stronger evidence of manual contribution than a flattened export.
- Save prompt and iteration logs. Record the number of generations reviewed, the prompts used, and which output was selected, along with the date.
- Record the editing session. A screen recording or timestamped autosave history showing manual inpainting, color correction, or compositing work can substantiate a claim of substantial human revision.
- Write a short authorship statement. A plain-language description such as "AI-generated base image, followed by manual repainting of the face, hands, and background over approximately six hours" is useful both for registration disclosure and for licensing negotiations.
- Disclose AI involvement at the time of registration where required, particularly in the United States, since later discovery of undisclosed AI use can lead to cancellation of an existing registration.
Common Situations and Their Copyright Outcomes
| Scenario | Likely outcome (US/EU) | Likely outcome (UK) |
|---|---|---|
| Single prompt, first output accepted unmodified | Not protected; human contribution too thin | May qualify under Section 9(3), reduced 50-year term |
| Many generations reviewed, one selected, no further editing | Selection alone generally insufficient for protection of the image | Arranger may hold rights under Section 9(3) |
| AI output substantially repainted, composited, and color-graded by hand | Likely protected; human-authored elements registrable | Likely protected as a human-authored work |
| AI used only for early sketches, final piece painted traditionally | Fully protected as a human-authored work | Fully protected as a human-authored work |
| Upscaling or minor cleanup applied to an entirely human-made painting | Protected; AI tool used in a non-authorship-altering role | Protected; AI tool used in a non-authorship-altering role |
Common Mistakes
Creators frequently assume that visual detail or technical polish implies legal protection; it does not. Others assume a platform's terms of service automatically grant them full ownership of outputs, when in fact a license from the tool provider and a copyright under national law are separate matters, and some providers explicitly disclaim any ownership claim over outputs. A third common error is treating the EU's text and data mining exception as a blanket permission that also covers how outputs may be used commercially; the exception addresses training legality, not output ownership or licensing.
Practical Next Steps
Creators working with AI tools should keep process documentation as a habit rather than an afterthought, check the specific disclosure requirements of the copyright office relevant to their country before filing a registration, and read the terms of service of each generative tool they use, since licensing terms vary between providers and can change between product versions. For any work with meaningful commercial value, a short consultation with an intellectual property attorney before publication or sale is a reasonable precaution given how unsettled this area of law remains, and it is far cheaper than resolving a dispute after the fact.
