The short answer
Whether you can copyright AI-generated content depends primarily on how much human creativity contributed to the final work and which country’s law applies. In the United States, material produced entirely by an artificial-intelligence system generally is not protected by copyright because copyright requires human authorship. A person may, however, obtain copyright in human-authored elements of an AI-assisted work, such as original writing, creative editing, selection, arrangement, or substantial modifications.
This distinction answers the related questions—can you copyright AI art, are AI-generated images copyrighted, and who owns AI-generated content? Purely machine-generated output may be usable by anyone if no other rights apply, but the person who generated it does not automatically receive an exclusive copyright merely because they entered a prompt, paid for a service, or possessed the resulting file. Ownership of a file, a license to use an output, and copyright protection are separate issues.
The law is not identical worldwide. Some jurisdictions recognize computer-generated works through special rules or attribution mechanisms, while others emphasize human creativity in ways similar to the United States. Contracts, employment arrangements, privacy rights, trademark law, publicity rights, and the terms of the AI service can also affect what a person may do with an output. For an important commercial, publishing, or legal use, review by a qualified lawyer in the relevant jurisdiction is prudent.
Copyright and artificial intelligence are different questions
Several questions are often combined under the phrase “copyright AI-generated content,” but they should be separated:
- Is the output protected by copyright? This concerns whether the law recognizes an exclusive copyright in the material.
- Who owns any copyright that exists? This concerns the person or organization entitled to exercise the rights.
- Does the user have permission to use the output? This is often determined by a platform’s contract or license.
- Does the output infringe someone else’s rights? A work can be unprotected by copyright and still create problems involving trademarks, privacy, publicity, confidentiality, or other laws.
- Can the work be registered or enforced? Registration systems may require disclosure of AI-generated material and identification of the human-authored portions.
For example, an AI service might give a user contractual permission to use an image commercially. That permission does not necessarily mean the image is copyrightable, and it does not guarantee that the image is free from claims by third parties. Conversely, a user may own copyright in an original human-written article containing AI-assisted research or editing, even though the underlying AI-generated phrases may not be independently protected.
Copyright protects expression, not every useful result or idea. It normally does not protect a bare idea, a general concept, a fact, a short phrase, a method, or a style as such. AI systems can produce content that resembles a style or combines familiar elements without giving the user a copyright monopoly over the style or concept.
The human-authorship principle
In the United States, copyright protection is tied to original human authorship. A work must contain creative expression attributable to a human author. The law does not generally treat an AI model as an author, and an automated output created without meaningful human creative contribution is unlikely to qualify for copyright protection solely because a person initiated the generation process.
The central practical question is therefore not simply whether AI was used, but what the human actually contributed. Human involvement can range from negligible to extensive:
| Human involvement | Likely copyright significance in the United States |
|---|---|
| Clicking a generation button with no meaningful creative decisions | Usually does not create copyright in the resulting machine-generated expression |
| Entering a short, general prompt | Usually insufficient by itself to make the output human-authored |
| Developing detailed instructions and repeatedly selecting among results | May support protection for the human’s selection or arrangement, but not necessarily for every generated element |
| Combining AI output with original human text, drawings, photographs, or music | Copyright may protect the human-created contributions and the original compilation |
| Substantially editing, rewriting, painting over, or transforming the output | May create protection for the human modifications and the resulting human-authored expression |
| Creating the work manually and using AI only for limited assistance | The final work may be protected to the extent its expressive elements are human-authored |
These categories are not mechanical rules. A highly detailed prompt can reflect significant creative thought, but the prompt is an instruction to a system rather than necessarily the human expression embodied in the output. A court or copyright office may examine the actual process, the nature of the decisions, and the relationship between the user’s choices and the final expressive details.
Prompts themselves
A prompt can sometimes be copyrightable as text if it contains sufficient original expression. That is a separate question from whether the image, video, or passage generated in response to the prompt is protected. A short instruction such as “a red bird on a branch” ordinarily contains too little original expression to receive meaningful copyright protection, while a long, literary prompt could potentially contain protectable wording.
The existence of a copyright in a prompt would not normally give the author ownership of every output produced from it. Copyright does not generally allow someone to control all later works based on an idea or instruction. The output must be analyzed independently.
Selection, arrangement, and curation
Human selection can matter even when the individual AI outputs are not protected. For example, a person might generate hundreds of images, choose particular images, crop and order them, and combine them with original captions in a creative presentation. Copyright may protect the human-authored selection, arrangement, captions, and other contributions, although it may not protect the underlying machine-generated images as standalone works.
The protection of a compilation is usually limited. It does not automatically prevent others from using unprotected components or making a different arrangement. The more the final work reflects deliberate, creative human choices, the stronger the argument for protection in those choices.
Can you copyright AI art and AI-generated images?
Purely AI-generated art is generally not copyrightable in the United States when no human has contributed sufficient original expression. This applies whether the output is an illustration, photograph-like image, painting, logo concept, 3D model, animation, or other visual work.
The person who supplies the prompt does not automatically own copyright in the resulting image. Payment for an AI subscription also does not itself create copyright. A service’s terms may grant the user contractual rights to use, reproduce, or sell an output, but those contractual rights are not the same as an exclusive copyright enforceable against the public.
AI-assisted art may contain copyrightable human contributions. Examples include:
- A human-made sketch that is transformed or expanded with AI while the final expressive choices remain substantially human-directed.
- An original photograph that a person edits, composites, masks, colors, or combines with generated elements.
- A generated image that is extensively repainted, redrawn, or modified by a human artist.
- A human-created collage or series whose selection, sequencing, layout, and accompanying text are original.
- A book, game, or advertising campaign in which AI-generated images are combined with original human writing, design, characters, and art direction.
In these cases, the likely result is partial protection: copyright may cover the human-authored material and the original arrangement, while excluding purely generated portions. A registration or infringement claim that treats all of the output as human-created may be vulnerable if the AI contribution was not disclosed or if the claimed authorship is inaccurate.
A person can also own the physical or digital copy of an image without owning exclusive rights in its expressive content. Possessing a file, receiving an output through an account, or having a license to display an image does not necessarily prevent others from copying the same unprotected image.
Who owns AI-generated content?
The answer depends on what “owns” means.
Ownership under a service’s terms
An AI provider may state that the user owns, receives, or is licensed to use outputs. These terms can address commercial use, redistribution, sublicensing, confidentiality, account restrictions, and the provider’s rights to operate or improve the service. They may also reserve rights where outputs are similar because many users entered similar prompts.
Service terms can give a user permission to use an output even where copyright does not exist. They can also impose obligations that limit use. Terms change over time and may differ by product, location, account type, or subscription plan, so the applicable version and any incorporated policies should be checked when the output matters commercially.
A contract cannot necessarily create a copyright against unrelated third parties where copyright law does not recognize one. It can, however, govern the relationship between the user and the provider and may create obligations concerning confidentiality, indemnity, prohibited uses, or dispute resolution.
Ownership by an employer or client
If a human employee creates copyrightable material as part of employment, the employer may own the copyright under applicable work-made-for-hire or employment rules. A commissioned freelancer usually does not transfer copyright merely because the client paid, unless an effective written agreement or applicable law provides otherwise. The parties should specify who owns human-authored additions, whether AI tools may be used, who bears infringement risk, and what disclosures are required.
An agreement cannot turn purely machine-generated expression into human-authored expression. It can still allocate contractual rights, warranties, payment rights, and responsibilities for using the output.
Multiple users and similar outputs
AI systems may generate similar or identical results for different users. If the output lacks copyright protection, neither user necessarily has an exclusive copyright in it. Even if one user has copyright in a human-created arrangement or modification, that protection normally does not extend to an independently created similar work.
Registration, disclosure, and recordkeeping
In the United States, an applicant for copyright registration is expected to identify material that was generated by AI when that material is more than a trivial part of the work. The applicant should claim only the human-authored portions. A work that includes both human and machine-created material may still be registerable, but the application must accurately describe the contributions.
The exact treatment can depend on the work and the extent of AI involvement. A human-authored novel containing a small amount of generated text, a collection of generated images with original curation, and a heavily modified AI illustration raise different questions. Failure to disclose relevant AI-generated material or claiming authorship in material not created by a human can create registration and enforcement problems.
Creators and businesses should retain useful records, such as:
- Original sketches, drafts, photographs, source files, and project files.
- Prompts and major revisions to prompts.
- The dates and versions of generated outputs.
- Human editing, compositing, painting, rewriting, or arrangement steps.
- The applicable provider terms and license information.
- Records showing which parts were created by employees, contractors, or third-party contributors.
These records do not guarantee protection, but they can help establish the human contribution and clarify contractual rights.
AI training and infringement are separate from output ownership
The question of who owns an AI output is different from whether an AI provider lawfully obtained or used training material. Models may be trained on large collections of text, images, code, audio, or other data. Whether particular training practices are lawful can depend on copyright exceptions, licenses, contracts, jurisdiction, and unresolved or changing case law.
Likewise, an output can raise infringement concerns if it reproduces protected expression from a particular source, even if the user did not know the source or if the output itself would otherwise be difficult to protect. A generated image can also resemble a protected character, brand logo, photograph, or distinctive design. The absence of copyright in the generated output is not a blanket defense to every claim involving third-party rights.
Creators should be especially careful with outputs that:
- Reproduce recognizable characters, logos, product designs, or photographs.
- Contain identifiable people whose privacy or publicity rights may be implicated.
- Use confidential business information or personal data in prompts.
- Closely imitate a living artist’s specific work rather than merely drawing on a broad artistic tradition.
- Are used as trademarks, packaging, commercial endorsements, or evidence in regulated contexts.
Trademark, trade secret, right-of-publicity, privacy, consumer-protection, and contractual rules can apply independently of copyright.
Differences outside the United States
There is no single worldwide rule for AI-generated content. Many copyright systems are built around human intellectual creation, which can make entirely autonomous output difficult to protect. Other legal systems have provisions for certain computer-generated works or use different tests for originality and authorship. The identity of the author, the location of relevant acts, the country where protection is sought, and the applicable contract can all matter.
For example, a work may receive different treatment in the United States, the United Kingdom, the European Union, Canada, Australia, or another jurisdiction. International conventions help coordinate certain rights but do not create one uniform test for AI authorship. A creator seeking worldwide commercial protection should not assume that a conclusion reached in one country automatically applies elsewhere.
Regional rules can also differ regarding text-and-data mining, moral rights, training data, disclosure, database rights, and contractual restrictions. “Copyright-free” is therefore an unsafe worldwide description unless the legal basis and jurisdiction are clearly identified.
Practical approach for creators and businesses
A sensible workflow is to treat AI output as a component whose rights must be investigated rather than as automatically owned intellectual property.
First, identify the jurisdiction and the intended use. Personal experimentation, publication, merchandise, client work, advertising, and use as a brand asset present different levels of risk. Next, read the AI provider’s current terms, including rules for commercial use and output ownership. Then document the human contribution and preserve the underlying files.
When the work is important, separate the claimed components: human-written text, human-made artwork, generated material, third-party material, and the final arrangement. Use contracts that address AI assistance, approvals, confidentiality, indemnification, ownership of human additions, and responsibility for clearance. Before release, review outputs for recognizable third-party material, personal information, and branding conflicts.
If seeking registration or asserting infringement, describe AI involvement accurately. Do not claim the entire output as human-authored merely because a human selected the prompt or paid for the tool. A lawyer or qualified intellectual-property professional can assess borderline cases, especially where the output is central to a business, contains third-party material, or will be distributed across multiple countries.
The practical rule
For most users, the safest general rule is:
AI assistance does not automatically prevent copyright, but purely AI-generated expression does not automatically receive copyright either. Copyright generally follows the original human creative contribution, not the mere act of operating an AI tool.
A user may have a contractual license to use an AI-generated image or passage without owning copyright in it. A human may own copyright in original edits, writing, artwork, selection, or arrangement without owning the underlying generated portions. And neither contractual permission nor lack of copyright eliminates separate issues involving infringement, privacy, publicity, trademarks, confidentiality, or the terms under which the AI system was used.
The Legal Framework of Human Authorship and AI Output
Under current statutory frameworks and judicial precedents across most major jurisdictions, purely AI-generated content cannot be copyrighted. Copyright law is fundamentally designed to incentivize and protect human creative labor. When an artificial intelligence model generates text, an image, code, or musical composition based solely on an automated prompt, the resulting output lacks the necessary human authorship required for copyright protection. In the United States and many other legal systems, such unassisted AI outputs immediately enter the public domain upon creation, meaning they can be freely copied, distributed, and used by anyone without legal infringement.
However, the boundary between unprotectable machine output and protectable human creative work is nuanced. While raw machine outputs cannot hold copyright, human-created works that incorporate AI-generated elements can be copyrighted, provided there is sufficient original human expression, selection, arrangement, or post-generation modification. In these cases, copyright protection extends exclusively to the human contributions and arrangements—not to the underlying AI-generated assets themselves.
+-------------------------------------------------------------------------+
| SPECTRUM OF COPYRIGHTABILITY |
+-------------------------------------------------------------------------+
| PURE AI OUTPUT AI-ASSISTED HYBRID HUMAN ORIGINAL|
| (Prompt Only) (Iterative Human Work) (Manual Tool) |
| |
| • No human execution • Human arrangement/layout • Direct human|
| • Output is random/machine • Substantial digital paint authorship |
| • Public Domain • Text/story by human • Full |
| copyright |
| [ CANNOT BE COPYRIGHTED ] [ PARTIAL PROTECTION ] [ FULL PROT. ]|
+-------------------------------------------------------------------------+Understanding whether AI content is protectable requires separating three distinct issues:
- Copyrightability: Does the work meet statutory standards for intellectual property protection against the public?
- Ownership via Contract: What rights does an AI platform's Terms of Service (ToS) grant to the prompter?
- Infringement Liability: Does using or training on copyrighted material to generate new outputs violate existing third-party rights?
The Authorship Doctrine: Why Prompts Do Not Confer Copyright
The central pillar of copyright law in the United States, the European Union, and international treaties like the Berne Convention is the human authorship requirement (nexus between human mind and creative expression).
The Legal Distinction Between Prompts and Authorship
A common misconception is that crafting a complex, detailed prompt constitutes sufficient creative labor to secure copyright over the resulting image or text. Copyright offices and federal courts have rejected this argument on foundational grounds:
- Prompts as Ideation rather than Expression: Copyright protects the specific expression of an idea, not the idea itself (the idea-expression dichotomy). A prompt like
"oil painting of an astronaut in a Renaissance cathedral, dramatic volumetric lighting, 8k"is legally treated as an instruction or an idea. The generative engine—not the user—determines how every pixel, color gradient, shadow, and line is rendered. - The Commissioning Analogy: Legally, a prompter functions like a patron commissioning an artist. If a patron gives detailed instructions to an independent painter, the patron does not become the legal author of the painting; the human artist who wields the brush is the author. Because the AI model executing the prompt is a machine without legal personhood, no legal author exists, rendering the raw output public domain.
- Lack of Direct Predictability and Control: In Burrow-Giles Lithographic Co. v. Sarony (1884), the U.S. Supreme Court ruled that photography is copyrightable because the human photographer exercises direct control over framing, lighting, pose, timing, and lens selection. By contrast, generative AI tools rely on probabilistic diffusion or neural network weights where the user cannot predetermine the exact placement of elements, meaning direct human creative execution is absent.
Key Precedents and Regulatory Guidance
Administrative rulings and court challenges have solidified the limits of AI copyrightability.
U.S. Copyright Office Guidance (March 2023)
In March 2023, the United States Copyright Office (USCO) published formal policy guidance on works containing AI-generated materials. The USCO established that:
- Applicants have an explicit duty to disclose the inclusion of AI-generated content in copyright applications.
- AI-generated content that is more than de minimis (more than a trivial or background element) must be disclaimed.
- Examiners will evaluate whether the AI contributions are the result of "mechanical reproduction" or whether the work reflects the applicant's own mental conception.
┌─────────────────────────────────────────────────────────────────────────┐
│ LANDMARK AI COPYRIGHT RULINGS │
├──────────────────────────┬─────────────────────────┬────────────────────┤
│ Case / Work │ Key Ruling │ Legal Principle │
├──────────────────────────┼─────────────────────────┼────────────────────┤
│ Zarya of the Dawn │ Registration cancelled │ Compilation and │
│ (Kashtanova, 2023) │ for raw Midjourney art; │ text protected; │
│ │ text & layout upheld │ images disclaimed │
├──────────────────────────┼─────────────────────────┼────────────────────┤
│ Théâtre D’opéra Spatial │ Copyright rejected │ Prompt engineering │
│ (Jason Allen, 2023) │ despite 600+ prompts │ does not equal │
│ │ and Photoshop tweaks │ human authorship │
├──────────────────────────┼─────────────────────────┼────────────────────┤
│ Creativity Machine │ Pure AI output denied │ Human authorship │
│ (Thaler v. Perlmutter) │ copyright by US District│ is bedrock law │
│ │ Court (D.D.C. 2023) │ under Title 17 │
└──────────────────────────┴─────────────────────────┴────────────────────┘Zarya of the Dawn (Kristina Kashtanova)
In late 2022, artist Kristina Kashtanova registered a copyright for a graphic novel titled Zarya of the Dawn, created using images from Midjourney. Upon discovering the images were AI-generated, the USCO partially cancelled the registration in February 2023:
- Protected: Kashtanova’s original written text, as well as the unique selection, coordination, and arrangement of the images and text on each page.
- Unprotected: The individual visual images generated by Midjourney, which were deemed uncopyrightable machine outputs.
Théâtre D’opéra Spatial (Jason Allen)
Artist Jason Allen won first place at the 2022 Colorado State Fair with a digital art piece created using Midjourney, upscaled with Gigapixel AI, and refined in Adobe Photoshop. Allen appealed the USCO's refusal to register the work, arguing that inputting over 624 prompts, adjusting parameters, and performing digital touch-ups constituted authorship. The USCO Review Board affirmed the refusal, holding that prompt-driven generation remains fundamentally automated and that the claimed human modifications were insufficient to grant copyright over the entire image without disclaiming the AI-generated core.
Thaler v. Perlmutter (Creativity Machine)
Computer scientist Stephen Thaler listed an AI algorithm ("The Creativity Machine") as the sole author of an artwork titled A Recent Entrance to Paradise, attempting to challenge the human authorship requirement directly. In August 2023, the U.S. District Court for the District of Columbia upheld the Copyright Office’s rejection, stating plainly that "human authorship is an essential component of copyright."
What Can Be Protected: Hybrid Works and Modifications
Although raw generative output cannot be protected, creators can secure copyright over projects involving AI through three distinct legal avenues.
HUMAN INPUT & MODIFICATION PATHWAYS
[Raw AI Generation] ──────► Unprotected / Public Domain
│
├───► + Human Selection & Arrangement ───► [Compilation Copyright]
│ (e.g., Anthologies, Graphic Novels) (Protects layout & order)
│
├───► + Substantial Human Painting ───► [Derivative Copyright]
│ (e.g., Overpainting, Retouching) (Protects human edits only)
│
└───► + Original Human Text / Music ───► [Integrated Work Copyright]
(e.g., Human lyrics over AI audio) (Protects human elements)1. Human Selection, Coordination, and Arrangement (Compilations)
Under Section 103 of the U.S. Copyright Act, an author can obtain a compilation copyright by gathering and organizing pre-existing materials (even uncopyrightable ones) in an original way.
- Example: A web designer who generates 50 individual icons using an AI tool, selects the best 10, edits their color palettes for harmony, and arranges them into a cohesive user-interface design system can protect the layout and overall compilation, though third parties could theoretically copy individual raw icons without infringing the compilation.
2. Substantial Post-Generation Human Modification
If a creator uses an AI output as a rough draft, sketch, or underpainting, and then substantially modifies the work using traditional digital painting, physical media, or complex manual compositing, the new creative additions are protectable.
- Standard of Originality: Minor edits—such as basic color correction, automatic upscaling, or global contrast adjustments—do not meet the threshold. The human must add original expressive elements (e.g., manually drawing new characters, altering anatomy, painting intricate backgrounds, or compositing multiple disparate elements into an entirely new composition).
3. Integrated Mixed-Media Works
In multimedia fields such as film, video games, and music production, AI may be used for isolated assets while humans handle other core components:
- Film/Animation: Human-written screenplays, directed voice acting, and choreographed pacing remain fully copyrighted, even if background concept art used AI.
- Music: Human-written melodies and original lyrics remain protected, even if an AI-generated synthesizer patch or background beat was used during production.
Terms of Service Ownership vs. Statutory Copyright
A critical area of confusion for businesses and creators is the difference between contractual ownership granted by an AI vendor and statutory intellectual property rights under the law.
┌────────────────────────────────────────────────────────────────────────┐
│ CONTRACTUAL OWNERSHIP vs. STATUTORY COPYRIGHT │
├──────────────────────────┬─────────────────────────────────────────────┤
│ Characteristic │ Contractual Ownership (ToS) │
├──────────────────────────┼─────────────────────────────────────────────┤
│ Legal Mechanism │ Private bilateral contract │
│ Enforceability │ Only enforceable against the platform │
│ Commercial Use │ Dictates whether the platform lets you sell │
│ Protection vs. Copycats │ None (does not prevent public copying) │
├──────────────────────────┼─────────────────────────────────────────────┤
│ Characteristic │ Statutory Copyright (Federal Law) │
├──────────────────────────┼─────────────────────────────────────────────┤
│ Legal Mechanism │ Federal/International intellectual property │
│ Enforceability │ Enforceable against any third party globally│
│ Commercial Use │ Exclusive right to monetize, license, adapt │
│ Protection vs. Copycats │ Grants federal infringement litigation power│
└──────────────────────────┴─────────────────────────────────────────────┘When platforms like OpenAI (ChatGPT, DALL-E), Midjourney, or Anthropic state in their Terms of Service that "You own all output generated by your prompts," they are making a contractual assignment. This means:
- The platform will not claim an equity or royalty stake in your creations.
- The platform grants you permission to use the generated output commercially.
- The platform agrees not to sue you for using the output.
However, a private terms of service agreement cannot override federal statutory copyright law. While OpenAI assigns whatever rights it has in the output to you, if the law states that purely AI-generated outputs have zero copyright protection, you own zero statutory rights against third-party copycats. If a competitor downloads an AI-generated image directly from your website, you cannot successfully sue them for copyright infringement under federal law if the image is a raw, unedited AI output.
Global Legal Perspectives
While the human authorship requirement is widespread, approaches to AI-generated works vary across different international jurisdictions.
| Jurisdiction | Human Authorship Required? | Statutory Recognition for Computer-Generated Works | Notable Legal Position / Precedent |
|---|---|---|---|
| United States | Yes | No | USCO guidance firmly rejects pure AI outputs; Thaler v. Perlmutter requires human origin. |
| European Union | Yes | No | CJEU standard requires work to be the "author's own intellectual creation," expressing personal creative choices. |
| United Kingdom | Qualified | Yes (CDPA s. 9(3)) | Grants 50-year protection to "computer-generated works" with the author defined as "the person by whom the arrangements necessary for the creation of the work are undertaken." Generative AI applicability remains actively debated. |
| China | Mixed | No | The Beijing Internet Court (Li v. Liu, 2023) granted copyright to an AI-generated image based on extensive prompt customization and parameter settings, though other Chinese courts have ruled differently. |
| Japan | Yes | No | Focuses heavily on whether the output reflects human "creative intent" and "creative contribution." Generative inputs alone are generally unprotectable. |
| Australia | Yes | No | Telstra Corp v Desktop Marketing Systems and IceTV established that works generated without direct human intellectual effort cannot hold copyright. |
The United Kingdom's Computer-Generated Works (CGW) Provision
Section 9(3) of the UK’s Copyright, Designs and Patents Act 1988 (CDPA) is one of the few statutory exceptions globally that explicitly provides protection for works generated by computers without direct human authors. It assigns authorship to the person who made the "arrangements necessary for the creation of the work." However, UK legal scholars and intellectual property courts continue to debate whether simply typing natural language prompts into an interactive commercial AI model meets the threshold of "making the necessary arrangements," or if the provision was originally intended for proprietary software developers and data pipeline engineers.
China’s Li v. Liu Landmark Decision
In November 2023, the Beijing Internet Court diverged from Western consensus in Li v. Liu. The plaintiff used Stable Diffusion to create an image of a woman, carefully adjusting prompt weights, seed numbers, and generation parameters. The court held that the continuous adjustment of parameters and aesthetic selection reflected the plaintiff's original intellectual input, thereby granting copyright protection. This case highlights how different jurisdictions may define the boundary of "creative control" differently in the coming years.
Infringement Risks: Inputs vs. Outputs
When assessing whether AI content is legally viable for commercial deployment, organizations must evaluate two separate exposure surfaces: input infringement and output infringement.
INPUT INFRINGEMENT VS. OUTPUT INFRINGEMENT
[Training Data Scrape] ──────────────► INPUT LIABILITY
• Web scraping billions of images • Fair Use defense (US)
• Lawsuits: NYT v. OpenAI, • TDM Exceptions (EU, Japan)
Getty v. Stability AI • Potential statutory damages
[User Prompts & Generates] ──────────► OUTPUT LIABILITY
• Model hallucinates/memorizes art • Substantial Similarity test
• Output looks identical to Disney, • Direct, vicarious, or
Getty, or living artists contributory infringement1. Input Liability (Training Data)
Multiple class-action lawsuits and high-profile litigations (e.g., The New York Times v. Microsoft & OpenAI, Andersen v. Stability AI, Getty Images v. Stability AI) target the scraping of copyrighted works to train neural models. Defendants typically assert the Fair Use doctrine (under U.S. law) or Text and Data Mining (TDM) exemptions (under EU and Japanese law), arguing that training is transformative because it extracts mathematical patterns and relationships rather than distributing copies of the original works.
2. Output Liability (Substantial Similarity)
Regardless of how a model was trained, if an end-user prompts an AI engine in a way that generates an output substantially similar to an existing copyrighted character, protected artwork, or specific textual excerpt, the end-user (and potentially the enterprise publishing the work) can be sued for direct copyright infringement.
- Memorization and Overfitting: When models are over-trained on specific iconic images (e.g., well-known comic book characters or photographic portraits), they can regurgitate near-exact duplicates of copyrighted works even in response to generic prompts.
- Commercial Indemnification: In response to corporate liability fears, major enterprise providers (including Adobe Firefly, Microsoft Copilot, and Google Cloud) offer commercial indemnification clauses, agreeing to defend and compensate enterprise customers against third-party copyright claims resulting from generated outputs, provided the user did not intentionally attempt to generate infringing material.
Practical Framework for Creators and Enterprises
For businesses, independent creators, and software developers seeking to utilize AI tools while maximizing their intellectual property protection and minimizing legal exposure, the following operational steps are critical:
1. Document the Creative Workflow
To prove human authorship during copyright registration or in litigation, maintain an audit trail of the creative process:
- Save intermediate files: Keep layered
.PSDfiles, vector project files, digital audio workstation (DAW) project stems, and version histories that showcase human modifications over time. - Log human additions: Document where AI output was used strictly as a reference, wireframe, or initial baseline, and where human manual labor altered the final expression.
2. Apply the Layering Strategy
To secure enforceable copyright on commercial deliverables, treat raw AI generation as raw material rather than a final deliverable:
- Visual Arts: Use AI for brainstorming or texture generation, then manually paint, composite, relight, and re-illustrate the primary focal elements.
- Software Engineering: Use AI coding assistants (e.g., GitHub Copilot) for routine syntax or boilerplate, while ensuring the broader software architecture, novel algorithms, and modular structure are designed and integrated by human engineers.
- Publishing: Use AI for outlining, topic research, or proofreading, while keeping the structural narrative, unique phrasing, character voices, and critical analysis authored by human writers.
RECOMMENDED IP ASSET REGISTRATION STRATEGY
+-------------------------------------------------------------+
| 1. AUDIT ASSETS |
| Identify all AI-generated vs. human-authored components. |
+------------------------------┬------------------------------+
▼
+-------------------------------------------------------------+
| 2. APPLY SUBSTANTIAL HUMAN LABOR |
| Overpaint, rearrange, composite, edit, and orchestrate. |
+------------------------------┬------------------------------+
▼
+-------------------------------------------------------------+
| 3. FORMULATE USCO APPLICATION |
| • Claim: Original text, arrangement, and human edits. |
| • Disclaim: AI-generated backgrounds, raw image files. |
+------------------------------┬------------------------------+
▼
+-------------------------------------------------------------+
| 4. PROTECT VIA SUPPLEMENTAL MECHANISMS |
| Trademarks, trade secrets, contracts, and ToS. |
+-------------------------------------------------------------+3. Disclose Correctly on Registration Filings
When filing applications with the U.S. Copyright Office or international equivalents:
- Do not list an AI program or engine as an author or co-author.
- Complete the "Limitation of Claim" section of the copyright application.
- Explicitly disclaim the unedited AI-generated visual, textual, or audio material that exceeds de minimis thresholds.
- Formulate the claim around the human contributions: "Selection, coordination, and arrangement of AI-generated assets alongside original human text and digital modifications."
4. Protect Commercial Assets Beyond Copyright
Because copyright cannot protect raw generative output, organizations should rely on alternative legal mechanisms to protect proprietary AI-driven assets:
- Trademark Law: Logos, brand names, and slogans can be registered as trademarks regardless of whether AI assisted in the design phase, because trademark law protects source identification and brand reputation in commerce, not creative authorship.
- Contract and Trade Secret Law: Access to proprietary model weights, specialized internal datasets, fine-tuning scripts, and proprietary prompt databases can be protected using non-disclosure agreements (NDAs), trade secret protections, and strict end-user license agreements (EULAs).
The short answer: copyright depends on human authorship and jurisdiction
Whether you can copyright AI-generated content is not answered simply by asking who typed the prompt or paid for the tool. In many important copyright systems—most clearly in the United States—copyright protects original human expression. An image, text, song, video, or other output created entirely by an autonomous generative system is generally not protected by copyright as a newly authored work if no human made sufficiently creative expressive choices in the result.
That does not mean AI-generated art is necessarily free of all legal restrictions, nor does it mean that nobody can own or use it. Several distinct questions are often incorrectly collapsed into one:
- Is the output copyrightable? This concerns whether a new copyright arises in the output.
- Who owns any copyright that does arise? Usually the human author or an employer/commissioning party under applicable law or contract—not the AI itself.
- Who has permission to use the output? This may be determined by the AI provider's terms of service, a workplace agreement, a client contract, or a platform license.
- Does the output infringe someone else's rights? An output can be unprotected itself while still being too similar to an existing copyrighted work, a trademark, or a person's protected likeness.
Accordingly, the most accurate general answer to “who owns AI art?” is: an AI cannot ordinarily be a copyright owner, but a human or organization may own copyright in the human-authored parts of an AI-assisted work and may have contractual rights to use the output. The scope of copyright, if any, depends on the human contribution and the law of the relevant country.
Copyright ownership, platform permission, and infringement risk are separate issues. A tool may grant a user broad commercial-use rights without making the output copyrightable, and a copyrightable AI-assisted work may still contain material that creates legal risk.
Why human authorship matters
Copyright is intended to protect original works of authorship: expressive works such as paintings, photographs, books, music, films, software, and graphics. It does not protect abstract ideas, methods, facts, styles, or merely mechanical production.
In the United States, the Copyright Office and courts have consistently treated human authorship as a foundational requirement. The Office has stated that it will register a work only where it was created by a human author, and it has rejected registrations for material generated by AI without sufficient human creative control over its expressive elements. Courts have also declined to recognize a nonhuman entity as an author under the Copyright Act.
This principle is not a judgment that a generated image lacks aesthetic value. It is a legal rule about the source of authorship. If a model determines the image's specific visual expression—its exact forms, placement, lighting, faces, textures, wording, or composition—then a user who merely requested a result may not be considered the author of those details.
A useful comparison is commissioning an artist. A client can describe a desired scene in general terms—“a red house on a cliff at sunset”—yet the artist makes many expressive decisions in turning that instruction into a particular image. A conventional generative AI system likewise supplies much of the specific expression, although through computational processes rather than human judgment. Whether the user's instructions crossed the line from a request to genuine authorship is highly fact-dependent.
Prompts are not automatically authorship
Writing a prompt can require imagination, technical knowledge, and effort. But effort alone is not the copyright test. The question is whether the prompt caused the user to control the work's protected expression with enough specificity and predictability.
For example, a prompt such as:
A cinematic watercolor of a fox reading beside a rainy window, warm lamplight,
blue and amber palette, highly detailedmay express an idea, mood, and requested features. The system still makes the decisive visual choices. It may place the fox differently on each generation, change the pose and props, or create details the user neither selected nor anticipated. Under the prevailing U.S. approach, that alone is unlikely to make the user the author of the generated image.
By contrast, a creator may begin with their own detailed drawing, use software features to alter selected parts, repeatedly revise the result, paint over it, add original typography, arrange it in a larger layout, or combine it with human-created material. Those steps can create copyrightable human expression even if some individual AI-generated portions remain unprotected.
There is no reliable universal rule based on prompt length, number of iterations, or percentage of AI use. A long prompt is not necessarily authorship, and a work containing substantial AI material is not necessarily wholly uncopyrightable. The relevant issue is which expressive elements came from the person and which came from the system.
Copyrightability, ownership, and licensing compared
The following distinctions help resolve most confusion about AI-generated content.
| Question | What it asks | Typical answer for a purely AI-generated output |
|---|---|---|
| Copyrightability | Does a new copyright exist in the output? | Often no in the United States, because human authorship is absent. |
| Authorship | Who created the protected expression? | The AI cannot ordinarily be an author; the user may not be one if they did not control expression. |
| Ownership | Who holds the copyright or related rights? | There may be no output copyright to own; contracts can still allocate use rights. |
| License | Who is allowed to copy, sell, publish, or adapt it? | The provider's terms, employment terms, or a client agreement may control. |
| Infringement | Does it unlawfully copy protected material or violate another right? | Possible, regardless of whether the output itself has copyright protection. |
A provider can say, in its contract, that it assigns to a user whatever rights it has in an output or permits commercial use of that output. That clause may be commercially important, but it cannot create copyright in a work that copyright law regards as lacking human authorship. Nor can it transfer rights that the provider does not own.
Likewise, calling an image “royalty-free” generally concerns the provider's license terms; it does not guarantee exclusivity, copyright registration, originality, noninfringement, or that another user will not receive a very similar image.
When an AI-assisted work may receive copyright protection
AI is not categorically incompatible with copyright. It can be used as one tool among many, much as creators use cameras, image-editing programs, synthesizers, spell-checkers, procedural effects, and other technologies. The law generally focuses on the human-created contribution.
Human selection, coordination, and arrangement
A compilation or design may be protected when a person makes original choices in selecting and arranging material, even if individual components are not protected. For example, a designer might choose among generated elements and arrange them with original text, photography, spacing, color relationships, and visual hierarchy in a poster. Copyright may cover that human-authored composition, though not necessarily each generated element independently.
The protection is usually thin: it prevents copying of the creator's original arrangement, not use of the underlying unprotected components in another arrangement.
Human modification and transformation
A creator can add copyrightable expression through meaningful editing. Relevant actions can include:
- Drawing, painting, sculpting, or retouching original details into an output;
- Reworking the composition or combining multiple sources in an original way;
- Writing original text, dialogue, narrative, captions, or code around generated material;
- Directing and editing footage or audio with creative choices that are reflected in the final work;
- Creating original source materials, then using AI only for limited assistance.
Merely cropping, resizing, applying routine filters, correcting a small artifact, or selecting one output from many may not be enough by itself. The stronger the identifiable human contribution to the final expressive result, the stronger the basis for claiming copyright in that contribution.
A practical spectrum
| Situation | Likely copyright position in a human-authorship jurisdiction |
|---|---|
| One prompt produces an image; the user publishes it unchanged | The generated image may lack copyright protection as a new work. |
| User generates alternatives and chooses one | Selection may have limited significance; the image itself may remain unprotected. |
| User uses a generated background in a human-designed book cover with original illustration and typography | The human-authored cover design and text may be protected; the background may not be. |
| Artist supplies a sketch, uses AI during drafting, then substantially redraws and revises the image | The artist's resulting original expression can be protected, subject to proof and scope. |
| Writer uses AI for brainstorming but writes the article independently | The human-written article is ordinarily copyrightable. |
| Writer asks AI to generate a full article and makes only superficial changes | Protection may extend only to the person's original edits, if any. |
This analysis applies across media. AI-generated prose, code, music, and video raise the same core question, though the relevant human contribution differs. A human editor's original revisions to AI text may be protected; the unedited model-generated passages may not be. A filmmaker's creative direction, choice of shots, editing, soundtrack, and human-performed elements may be protected even where certain visual effects were generated.
Registration and disclosure in the United States
For U.S. works, registration is not required for copyright to arise, but it is generally required before filing an infringement action for a U.S. work and provides important procedural benefits in some circumstances. Registration of a work containing AI-generated material requires particular care.
The U.S. Copyright Office has instructed applicants to disclose material that is more than de minimis and was generated by AI, exclude it from the claim, and describe the human-authored material being claimed. The appropriate treatment depends on the facts. An application that claims an entire AI-generated image as human-authored may be refused, limited, or later challenged.
An applicant should be able to identify, in concrete terms:
- the original human-created text, artwork, editing, arrangement, or other expression;
- the portions generated by the AI system;
- the nature of the human creative modifications; and
- the sequence of creation, where relevant.
Keeping source files, earlier drafts, sketches, layer files, prompt records, edit histories, and versioned exports can help establish what a person actually created. These records do not transform machine-generated portions into human authorship, but they can support a claim for genuine human contributions.
Registration practices and agency guidance can change, and registration rules differ outside the United States. Significant commercial works should be reviewed by a copyright lawyer familiar with the jurisdiction and the registration process.
Who owns AI-generated content in practice?
The word “owns” may refer to several kinds of control. The following parties may have different interests at once.
The person using the tool
A user may have a contractual right to download, publish, sell, or adapt outputs. They may also own copyright in their own human-authored inputs and later modifications. However, a user does not automatically obtain exclusive copyright in the raw generated output merely because they entered prompts, paid a subscription, or initiated generation.
The AI provider
The provider's terms may grant or reserve broad rights involving outputs, inputs, feedback, service improvement, safety review, or commercial use. Terms vary by provider, account type, region, and date. They should be read as a contract, especially before commercial publication.
A provider may have a license to use outputs for operational purposes even when users may commercially exploit them. Conversely, a provider may restrict certain uses, require attribution in some circumstances, or treat free and paid accounts differently. Such clauses do not decide the statutory copyrightability of an output, but they can determine whether a user is contractually permitted to use it.
An employer or client
If content is created in employment, the employer may own the copyright in an employee's human-authored work under work-made-for-hire principles in the United States or under local employment rules. A client agreement may also assign rights. But an agreement cannot assign nonexistent copyright in a wholly machine-generated element; it can instead allocate contractual rights, risk, confidentiality, and ownership of files or deliverables.
Creators working for clients should state clearly whether deliverables include AI-generated material, what human work was performed, which tool terms apply, and whether exclusivity can realistically be promised.
The people whose works were used in training or referenced in prompts
The legal status of training AI systems on copyrighted works has been the subject of litigation and policy debate in several jurisdictions. That separate issue is not automatically resolved by the status of any one output. A user normally does not acquire rights in training data merely by using a model.
More immediately, a generated output can create risk if it reproduces protected expression from a recognizable source. Asking for “in the style of” a living artist is not, by itself, the same as copying a particular copyrighted work under every legal system, since style is generally not protected by copyright. Still, the output may be problematic if it is substantially similar to a specific work, incorporates identifiable protected characters or logos, or creates consumer confusion.
AI art is not automatically in the public domain everywhere
It is tempting to say that uncopyrightable AI art is “public domain.” That shortcut can be misleading.
In the United States, if an output contains no protectable human expression and no other applicable right restricts it, there may be no copyright preventing others from copying it. In that limited sense, it can be freely reproducible under U.S. copyright law. But practical restrictions may still arise from a contract, trademark law, rights of publicity, privacy law, database rights, moral rights, or the copyright in incorporated human-authored material.
International law is less uniform. Some jurisdictions have provisions addressing computer-generated works. The United Kingdom, for example, has historically provided a statutory authorship rule for computer-generated literary, dramatic, musical, and artistic works where there is no human author, identifying the author as the person who made arrangements necessary for creation. How that older framework applies to modern generative AI, and the precise scope of protection it provides, can require legal interpretation. Other countries may emphasize human intellectual creation, recognize different related rights, or have not yet developed settled rules for generative AI.
The relevant law may depend on where a work is created, published, used, registered, or litigated. An answer valid for a U.S. registration application should not be assumed to settle rights in every market.
Risks beyond copyright
Publishing or selling AI-generated material involves more than determining whether the user owns copyright.
Copying and derivative-work claims
If an output is substantially similar to a copyrighted source, its use can draw an infringement claim. Intent is not always required. A user may not know what data a model was trained on, but they should not ignore obvious resemblance to a particular artwork, photograph, character, lyric, or illustration.
Avoid presenting generated material as an exact recreation of a named work or using it in place of licensed assets where it visibly replicates protected details. For high-value uses, conduct a human review and consider reverse-image or similarity checks as risk-reduction measures, not guarantees.
Trademarks and false endorsement
Logos, brand names, product trade dress, and marks that suggest a commercial source can raise trademark issues. An AI-generated image of a branded product or a logo-like symbol may be unsuitable for advertising, packaging, app icons, or other contexts where viewers could infer affiliation or sponsorship.
Likeness, privacy, and defamation
Realistic images, voice clones, and videos involving identifiable people can implicate rights of publicity, privacy, consent requirements, and defamation law. The rules vary widely by jurisdiction. The risks are especially serious for advertising, political communications, adult content, and depictions that could harm a person's reputation.
Confidentiality and data handling
Prompts, uploaded drafts, client designs, source code, medical information, and unreleased product plans may be confidential. Before using them in an AI service, organizations should assess the provider's data terms, retention practices, account settings, and internal policies. Copyright ownership does not solve a breach of confidentiality or privacy obligation.
Sensible practices for creators and organizations
Creators who want the strongest possible copyright position should treat AI as an assistive component of a documented human creative process rather than as a source of final, untouched deliverables.
- Create and preserve human-authored source material. Sketches, drafts, storyboards, layouts, and written revisions make the human contribution clearer.
- Make substantive expressive choices in the final work. Use AI outputs as ingredients, then compose, rewrite, redraw, edit, and refine them through your own judgment.
- Keep records proportionate to the work's value. Save editable files, dated drafts, layers, and descriptions of material changes.
- Read the applicable tool terms at the time of use. Confirm commercial-use rights, output restrictions, data-use provisions, and any required disclosures.
- Do not promise exclusivity casually. Other users may receive similar outputs, and unprotected portions may be available for others to copy.
- Review high-risk content before release. Scrutinize material resembling known art, brands, copyrighted characters, real individuals, or confidential source material.
- Be precise in contracts and registrations. Describe the human-authored work accurately and distinguish it from generated elements.
For ordinary personal use, these distinctions may have little practical effect. For publishing, licensing, advertising, games, film, software, merchandise, or investment-backed creative projects, they can materially affect the ability to enforce rights and to give warranties to partners.
The central limitation
Copyright law around generative AI is developing through agency guidance, legislation, and court decisions. There is no universal formula that converts a percentage of human involvement into ownership. The legally important evidence is the actual creative process and the original human expression visible in the claimed work.
Therefore, the best general rule is not “AI art can never be copyrighted” or “the prompter automatically owns it.” It is this: copyright can protect the original human-authored elements of an AI-assisted work, while purely AI-generated expressive elements may receive no copyright protection in jurisdictions requiring human authorship. The right to use an output and the risk that it infringes someone else's rights must be evaluated separately, under the relevant tool contract and applicable law.