Can you copyright AI-generated art?

No, not if it is generated solely by AI. The US Copyright Office has rejected multiple registration attempts for works created entirely by AI, including the famous 'A Recent Entrance to Paradise' painting created by the DABUS AI system. The Supreme Court declined to take up the AI authorship case in March 2026, settling the matter: AI cannot be an author. However, you CAN copyright AI-assisted art if there is meaningful human authorship. The key is the degree of human creative control. If you write a detailed prompt, iterate multiple times, make substantive selections, and creatively modify the output, the human-authored elements (your editing, arrangement, modification) may be copyrightable. The AI-generated portions must be disclosed in the copyright application. To maximize protectability: (1) Use AI to modify or adapt existing human-authored materials rather than generating from scratch. (2) Document your creative process and human contributions. (3) Substantially edit AI outputs. (4) Disclose AI-generated material in copyright applications. (5) For flagship brand materials, consider using licensed AI tools like Adobe Firefly (trained on licensed images) to reduce infringement risk.

Is using copyrighted works to train AI fair use?

This is the most contested question in AI copyright law in 2026. Over 70 lawsuits have been filed in US courts addressing whether using copyrighted works to train AI models constitutes fair use. The US Copyright Office's 2025 report on Generative AI Training provides an analytical framework but does not settle the question. Fair use is assessed on four factors: (1) Purpose and character of the use — transformative uses favor fair use, but commercial uses weigh against it. (2) Nature of the copyrighted work — factual works receive less protection than creative works. (3) Amount and substantiality used — using entire works weighs against fair use. (4) Effect on the market — if AI outputs compete with the original works, this weighs heavily against fair use. AI companies argue training is transformative. Creators argue their entire bodies of work are used without consent to produce competing content. Key cases: New York Times v. OpenAI (3 years pending), Getty Images v. Stability AI (settled with licensing deal). The outcome of these cases will have global consequences. The EU, Japan, and Singapore have text and data mining exceptions; the US relies on fair use; the UK is developing its approach.

Can AI be listed as an inventor on a patent?

No. The Federal Circuit settled this in Thaler v. Vidal, holding that an inventor under the Patent Act must be a natural person. The Supreme Court declined to take up the case, and in March 2026 turned away a related AI authorship case, so this is settled law. A patent application must name a human inventor — you cannot list the AI even if the AI designed the invention. However, you can use AI tools to help build your invention without hurting patent eligibility. The USPTO issued updated AI guidance in July 2024 confirming that examiners judge AI inventions on the same footing as any other technology. You can patent a specific technical method your AI system performs (a novel training technique, a specific architecture improvement, a new inference method), but you cannot patent 'an AI model' or 'software that uses a neural network' — that is too abstract under the Alice Corp v. CLS Bank test. The human inventor must have contributed to the conception of the invention, not just operated the AI tool.

How can enterprises manage IP risk with AI-generated content?

Enterprises should manage AI IP risk through five strategies: (1) Choose AI tools with licensing transparency — Adobe Firefly (trained on licensed images), Getty Images AI (licensed), and tools with indemnification programs reduce infringement risk. Avoid tools that trained on scraped data without licensing. (2) Document human contributions — carefully document the respective contributions of human authors and AI systems, including when, how, and to what extent AI-generated content is modified, curated, or incorporated into the final work. (3) Use AI for appropriate assets — AI-generated content is suitable for background, low-value, or short-lived assets where exclusivity is not critical. For flagship brand materials, national advertising, or proprietary visual identities, use human-authored content or substantially modify AI outputs. (4) Add review layers — implement IP review and approval processes for AI-generated content, checking for infringement, right of publicity, and consumer protection risks. (5) Use AI to modify existing human-authored works — rather than generating from scratch, using AI to adapt or modify existing materials can help preserve enforceable copyright in the underlying work. Consult legal counsel on your specific use cases, especially for public-facing commercial content.

July 13, 2026

TL;DR — AI copyright in 2026: Works generated solely by AI are not copyrightable (US Copyright Office requires human authorship). 70+ lawsuits pending on training data fair use. Key cases: NYT v. OpenAI, Getty Images v. Stability AI (settled with licensing). AI cannot be a patent inventor (Thaler v. Vidal, Supreme Court declined March 2026). Enterprise guidance: Use licensed AI tools (Adobe Firefly, Getty), document human contributions, use AI for low-value assets, add IP review layers.

AI Copyright Ownership in 2026: Who Owns AI-Generated Content, Training Data, and IP Rights

The intersection of AI and copyright law is the most contested legal area in technology in 2026. Over 70 lawsuits are pending in US courts addressing whether using copyrighted works to train AI models constitutes fair use. The US Copyright Office has issued multiple reports and guidance documents. The Supreme Court declined to take up AI authorship cases in March 2026, settling key questions (copyright.gov 2025, nortonrosefulbright 2026).

This guide covers copyright ownership of AI-generated content, training data fair use, patent ownership, and practical strategies for enterprises.

Copyright Ownership of AI-Generated Content

The Human Authorship Requirement

Under current US law, works generated solely by AI are not eligible for copyright protection. The US Copyright Office has repeatedly held that copyright requires human authorship (copyright.gov 2025):

"The term 'author,' used in both the Constitution and the Copyright Act, excludes non-humans. If a work's traditional elements of authorship were produced by a machine, the work lacks human authorship and the Office will not register it."

— US Copyright Office, AI Registration Guidance

Scenario	Copyrightable?	Who Owns It
AI generates image from a simple prompt	No	Public domain
AI generates text, human publishes without editing	No	Public domain
Human writes detailed prompt, iterates, selects, edits	Partially	Human owns their edits/arrangement
Human creates original work, uses AI to modify	Yes (human-authored elements)	Human owns original + modifications
Human arranges AI-generated elements into a larger work	Yes (arrangement)	Human owns the arrangement

Source: copyright.gov (2025), debevoise.com (2026).

Key Cases on AI Authorship

Case	Year	Ruling	Significance
Thaler v. Perlmutter	2023	AI cannot be an author	Copyright Office correctly rejected DABUS-created work
Thaler v. Vidal	2022, affirmed 2024	AI cannot be an inventor	Patent Act requires natural person inventor
Supreme Court declines AI authorship	March 2026	Settled	AI authorship question is resolved — AI cannot be author or inventor
Zarya of the Dawn	2023	Partial copyright	Human owns arrangement, AI-generated images are public domain

The UK Approach

The UK government is proposing to remove copyright protection for wholly computer-generated works while retaining protection for AI-assisted works. Most consultation respondents agreed that works created solely by AI should not be protected (gov.uk 2026).

Training Data and Fair Use

The Central Question

Can AI companies use copyrighted works to train their models without permission? This is the most contested question in AI copyright law. The US Copyright Office's 2025 report on Generative AI Training provides an analytical framework but does not settle the question (copyright.gov 2025):

"The stakes are high, and the consequences are often described in existential terms. Some warn that requiring AI companies to license copyrighted works would throttle a transformative technology. Others fear that unlicensed training will corrode the creative ecosystem."

The Four Fair Use Factors

Factor	What It Considers	AI Training Implication
1. Purpose and character	Is the use transformative?	AI companies argue training is transformative; creators disagree
2. Nature of copyrighted work	Factual vs. creative	Creative works receive more protection
3. Amount and substantiality	How much was used?	Using entire works weighs against fair use
4. Effect on market	Does it compete with original?	AI outputs competing with originals weighs heavily against fair use

Major AI Copyright Lawsuits

Case	Plaintiff	Defendant	Status (2026)	Key Issue
New York Times v. OpenAI	NYT	OpenAI, Microsoft	3 years pending	Article reproduction, training data
Getty Images v. Stability AI	Getty	Stability AI	Settled (licensing deal)	Image training without license
Authors Guild v. OpenAI	Authors	OpenAI	Pending	Book training data
Universal Music v. Anthropic	Record labels	Anthropic	Partial settlement	Lyric training data
Andersen v. Stability AI	Artists	Stability AI, Midjourney	Pending	Artist style imitation
Concord Music v. Anthropic	Music publishers	Anthropic	Settled	Lyric reproduction

Sources: aibusiness.com (2026), builtin.com (2026), nortonrosefulbright.com (2026).

OpenAI's defense: OpenAI generally denies claims but admitted to several factual allegations concerning datasets used to train GPT-2 and GPT-3. OpenAI asserted affirmative defenses including fair use and de minimis copying (nortonrosefulbright 2026).

Global Approaches to AI Training Data

Jurisdiction	Approach	Key Feature
USA	Fair use defense	Case-by-case, 70+ cases pending
EU	Text and Data Mining (TDM) exception	Opt-out mechanism for rightsholders
Japan	TDM exception	Broad exception for AI training
Singapore	TDM exception	Computational data analysis exception
UK	Developing	Proposing transparency + licensing framework
China	Sector-specific	Algorithm recommendation regulations

Source: gov.uk (2026).

Patent Ownership and AI

AI Cannot Be an Inventor

The Federal Circuit settled this in Thaler v. Vidal: an inventor under the Patent Act must be a natural person. The Supreme Court declined to take up the case, and in March 2026 turned away a related AI authorship case (usip.law 2026).

Question	Answer	Legal Basis
Can AI be listed as inventor?	No	Thaler v. Vidal (Federal Circuit)
Can AI be listed as author?	No	Thaler v. Perlmutter (DC District)
Can you patent AI inventions?	Yes (if specific technical method)	Alice Corp v. CLS Bank test
Can you use AI to help invent?	Yes	USPTO July 2024 guidance
Can you patent open-source AI?	Yes (your novel method, not the framework)	Copyright ≠ patent

Patenting AI Inventions

You cannot patent "an AI model" or "software that uses a neural network" — that is too abstract. You CAN patent a specific technical method your AI system performs (usip.law 2026):

Patentable	Not Patentable
A fine-tuning method that cuts training cost by reordering data flow	"An AI that predicts things"
A real-time neural-network anomaly detection method	"Software that uses a neural network"
An AI speech separation technique with specific steps	"Using AI to solve a business problem"

The Alice test: (1) Is the claim aimed at an abstract idea? (2) Does it add a real technical improvement? If yes to (1) and no to (2), the patent is invalid (usip.law 2026).

Enterprise IP Risk Management

flowchart TD Start["Enterprise wants to\nuse AI-generated content"] --> Q1{"What is the\ncontent's purpose?"} Q1 -->|"Flagship brand,\nnational ad, proprietary"| Human["Use human-authored content\nor substantially modify AI output\nDocument human contributions\nFile copyright for human elements"] Q1 -->|"Background, low-value,\nshort-lived"| Q2{"Which AI tool?"} Q2 -->|"Licensed training data\nAdobe Firefly, Getty AI"| Safe["Lower infringement risk\nUse with standard review"] Q2 -->|"Scraped training data\nUnknown provenance"| Risk["Higher infringement risk\nAdd IP review layer\nCheck for output similarity\nto existing works"] Q2 -->|"Open AI tools\nChatGPT, DALL-E, Midjourney"| Medium["Medium risk\nCheck terms of service\nModify outputs substantially\nDocument human authorship"] Safe --> Review["IP review process:\n1. Check for infringement\n2. Verify right of publicity\n3. Consumer protection check\n4. Document AI + human contributions"] Risk --> Review Medium --> Review Review --> Q3{"Commercial use?"} Q3 -->|"Yes"| Legal["Consult legal counsel\nReview terms of service\nConsider indemnification\nFile copyright if human-authored"] Q3 -->|"Internal only"| Deploy["Deploy with documentation"] Legal --> Deploy Human --> Deploy

Five Strategies for Managing AI IP Risk

Choose AI tools with licensing transparency — Adobe Firefly (trained on licensed images), Getty Images AI (licensed content), and tools with indemnification programs reduce infringement risk. Avoid tools that trained on scraped data without licensing (artandmedialaw 2026).
Document human contributions — Carefully document the respective contributions of human authors and AI systems, including when, how, and to what extent AI-generated content is modified, curated, or incorporated into the final work (debevoise 2026).
Use AI for appropriate assets — AI-generated content is suitable for background, low-value, or short-lived assets where exclusivity is not critical. For flagship brand materials, use human-authored content or substantially modify AI outputs (debevoise 2026).
Add IP review layers — Implement IP review and approval processes for AI-generated content, checking for infringement, right of publicity, and consumer protection risks (debevoise 2026).
Use AI to modify existing human-authored works — Rather than generating from scratch, using AI to adapt or modify existing materials can help preserve enforceable copyright in the underlying work (debevoise 2026).

AI Tool Licensing Comparison

Tool	Training Data	Indemnification	Copyright Ownership	Infringement Risk
Adobe Firefly	Licensed (Adobe Stock, public domain)	Yes (enterprise)	User owns outputs	Low
Getty Images AI	Licensed (Getty library)	Yes	User owns outputs	Low
ChatGPT/DALL-E 3	Mixed (some licensed, some scraped)	Limited (enterprise)	User owns outputs (per ToS)	Medium
Midjourney	Scraped (disputed)	No	User owns outputs (per ToS)	Higher
Stable Diffusion	Scraped (LAION dataset)	No	Open source model	Higher

Sources: artandmedialaw.com (2026), debevoise.com (2026).

Best Practices

Assume AI-only outputs are public domain — If you generate content solely with AI and publish it without meaningful human modification, no one owns the copyright. Plan accordingly (copyright.gov 2025).
Document your creative process — Record prompts, iterations, selections, edits, and human modifications. This documentation supports copyright claims for human-authored elements (debevoise 2026).
Disclose AI-generated material in copyright applications — The Copyright Office requires applicants to disclose AI-generated material and explain human contributions (copyright.gov 2025).
Use licensed AI tools for commercial content — Adobe Firefly and Getty Images AI are trained on licensed data and offer indemnification, significantly reducing infringement risk (artandmedialaw 2026).
Check AI outputs for similarity to existing works — AI models can reproduce training data. Before publishing, check if outputs are substantially similar to existing copyrighted works (debevoise 2026).
Review terms of service — Different AI tools have different terms regarding output ownership, commercial use, and indemnification. Read the fine print.
Consult legal counsel for high-stakes content — For flagship brand materials, national advertising, or proprietary visual identities, consult IP counsel before using AI-generated content (debevoise 2026).
Monitor ongoing litigation — The 70+ pending cases will shape AI copyright law. The outcomes of NYT v. OpenAI and similar cases will have global consequences. Stay informed (aibusiness 2026).

For related topics, see our AI ethics framework, AI deepfakes detection, AI accountability, AI misinformation prevention, and AI bias detection guides.

FAQ

What happens if AI generates content similar to my copyrighted work?

If an AI model generates output that is substantially similar to your copyrighted work, you may have an infringement claim. There are two theories of infringement: (1) Training data claims — the AI model was trained on your copyrighted work without authorization, and the ingestion itself constitutes infringement. (2) Output claims — the AI-generated output reproduces or is substantially similar to your protected work. To pursue a claim, you need to demonstrate that the AI model had access to your work (likely, if it was publicly available and used in training) and that the output is substantially similar. Document the similarity with side-by-side comparisons. Several lawsuits are testing this theory, including Andersen v. Stability AI (artists claiming AI outputs imitate their style). Note that style imitation is generally not copyrightable — copyright protects specific expression, not style or ideas. However, if the AI reproduces specific elements of your work (character designs, specific compositions, exact text), you may have a stronger claim. Consult IP counsel to assess your specific situation (debevoise 2026, nortonrosefulbright 2026).

Can I sell AI-generated content?

Yes, you can sell AI-generated content, but with important caveats. (1) Ownership — if the content was generated solely by AI, you do not own the copyright, meaning others can freely copy and use it. This reduces its commercial value for exclusive assets. (2) Terms of service — most AI tools (ChatGPT, DALL-E, Midjourney) grant users ownership of outputs per their terms, but this does not override the Copyright Office's human authorship requirement. You can sell the content, but you cannot prevent others from copying it. (3) Infringement risk — if the AI tool was trained on scraped data, outputs may inadvertently reproduce copyrighted material, exposing you to infringement claims. Use licensed tools (Adobe Firefly, Getty AI) for commercial content to reduce this risk. (4) Disclosure — some jurisdictions may require disclosure that content is AI-generated (EU AI Act transparency obligations). (5) Best practice — for commercial content where exclusivity matters, substantially modify AI outputs with human creativity, document your contributions, and file for copyright on the human-authored elements. For content where exclusivity doesn't matter (stock images, background graphics, social media posts), AI-generated content is fine to sell (debevoise 2026, artandmedialaw 2026).

Does the EU AI Act address copyright?

The EU AI Act primarily addresses safety, transparency, and fundamental rights rather than copyright directly. However, it includes relevant provisions: (1) Transparency obligations — AI systems that generate synthetic content (deepfakes, AI-generated media) must disclose that the content is AI-generated. (2) Training data documentation — general-purpose AI models must document their training data and comply with EU copyright law. (3) The EU's Copyright Directive (Article 4) includes a Text and Data Mining (TDM) exception that allows AI training on copyrighted works unless the rightsholder has explicitly opted out (machine-readable opt-out). This means EU-based AI training must respect opt-outs, while US-based training relies on fair use. The UK is developing its own framework with transparency and licensing requirements. The interaction between AI regulation (EU AI Act) and copyright law (Copyright Directive) creates a complex compliance landscape for multinational AI companies. Consult legal counsel on compliance with both regulatory frameworks (gov.uk 2026, copyright.gov 2025).

Can I copyright a prompt I wrote for an AI?

The prompt itself may be copyrightable as a literary work if it contains sufficient original, creative expression. A simple prompt like 'a cat sitting on a chair' is too short and generic to be copyrightable. However, a detailed, creative prompt with specific artistic direction, unique descriptions, and original composition instructions may meet the threshold for copyright protection as a literary work. The output generated from that prompt, however, is not copyrightable to the prompt author unless they have made meaningful human modifications to the output. The US Copyright Office has clarified that prompting alone — even detailed prompting — does not constitute the kind of human authorship required for copyright protection of the generated output. The human must exercise creative control over the final work through editing, selection, arrangement, or other modification. In practice: copyright your detailed prompts as literary works if they contain original expression, but do not assume that prompting alone gives you copyright in the AI-generated output. For maximum protection, modify the output substantially and document your creative process (copyright.gov 2025).

What is the difference between AI copyright and AI patent law?

AI copyright and AI patent law protect different things and have different rules regarding AI. Copyright protects original works of authorship (text, images, music, software code). Patents protect inventions (new and useful methods, processes, machines). Key differences: (1) AI as author/inventor — AI cannot be an author (copyright) or inventor (patent) under US law. Both require a human. (2) AI-generated output — AI-generated content is not copyrightable (no human authorship). AI-assisted inventions ARE patentable if a human contributed to conception (USPTO July 2024 guidance). (3) Protection scope — Copyright protects specific expression; patents protect novel methods. (4) Registration — Copyright registration is optional but provides enforcement benefits; patent registration is required (via USPTO application). (5) Duration — Copyright lasts life of author + 70 years; patents last 20 years from filing. (6) Training data — Copyright law addresses use of copyrighted works in training (fair use); patent law does not directly address training data. In practice: patent your novel AI methods (specific technical improvements, not abstract ideas), copyright your human-authored content (including modifications to AI output), and do not try to copyright or patent AI-only output (usip.law 2026, copyright.gov 2025).

Want a self-hosted AI company brain that does all of this out of the box?
Book a demo →