Who Owns AI’s Output?

Albert Einstein called creativity “intelligence having fun.” Historically, we’ve rewarded human creativity with intellectual property rights. Article 1, Section 8, Clause 8 of the U.S. Constitution empowers Congress to “promote the Progress of Science and useful Arts” by granting exclusive rights—patents and copyrights—to human creators. But what happens when the creative output comes from a machine? Who owns it, and how can it be used?

This post explores the core questions surrounding artificial intelligence (AI)-generated content ownership. We’ll examine: (1) the key legal issues; (2) their impact on particular industries; (3) practical strategies for navigating this new terrain; (4) international divergence in AI intellectual property (IP) norms; and (5) emerging trends and gray areas. As always, I welcome comments and insights!

1.     The Core Legal Questions

• Who owns AI-generated works? In most major jurisdictions (U.S., E.U., China), copyright requires three elements: (1) human authorship + (2) a creative work + (3) fixation in a tangible medium. While humans create AI’s algorithms and the chips that run its computations, the actual output is machine-generated. This precludes copyright protection. Patents—and trademarks in most jurisdictions—require filing. Newly-generated AI content thus enjoys no intellectual property protection. AI companies acknowledge this reality—terms of use (See, e.g., OpenAI ToU, Anthropic Legal and compliance, and Google Gemini terms) give content rights to the user.
• AI as an inventor or co-inventor. Users of AI don’t get free rein to make AI-generated content their own, however. Patent legislation restricts the ability to name AI systems as “co-inventors”—and thus the ability to patent AI-assisted inventions. One recent controversy involved the DABUS (Device for the Autonomous Bootstrapping of Unified Sentience), an AI system created by Stephen Thaler. The Federal Circuit in the US, according to Morgan Lewis, and the UK Supreme Court, according to JUVE Patent, interpreting their respective patent acts, emphasized that the term “inventor” is restricted to natural persons, leaving no room for AI as an inventor of AI-generated inventions (e.g., DABUS litigation in the U.S., U.K., and Australia).
• AI training data and copyright. Ownership of the information that goes into AI models is a more vexing question. There has been considerable dispute as to whether using books and other information to train AI models constitutes fair use. In June 2025, a California Federal court ruled that Anthropic’s use of books to train its Claude model constituted fair use. However, there is still litigation pending over the use of news content to train LLM models—which may be a greater test to the limits of fair use.

2.     Industry-Specific Impacts

• Creative industries. AI-generated content affects different industries in different ways. Creative industries will face the most immediate impact. First, AI is good—and getting better—at composing music, poetry, imagery, videos, and even novels. As we discussed above, none of this enjoys IP protection when it is generated. There is no human authorship (and thus no copyright). It is also not registered, and thus initially enjoys no patent or trademark protection. What about when humans tweak, register, or use AI-generated content in commerce? Here is a quick look:
  • Copyright: Tweaking AI-generated content can be sufficient to establish human authorship and thus copyright protection. However, this is a fact-specific question. There is also a question of whether AI attribution (“created with [Insert AI tool]” or watermarks) is enough to defeat copyright protection.
  • Patents: The creative industry is less directly patent-heavy than other industries. Nonetheless, patented technologies could have a major impact on creative output.
  • Trademarks: AI-generated logos and brand names can generally be trademarked. The USPTO simply requires use in commerce. There is no requirement for establishing human authorship. The ability for users to easily ideate, design and tweak their own trademarks—and then file them—will disrupt a number of creative industries, including graphic design and advertising.
• Software & code. AI is extremely good at generating—and re-generating—code. This poses two related challenges: (1) reproduction of existing code without a license and (2) ownership of new code.
  • Reproduction. The key question of whether reproducing existing code constitutes fair use. Because code is inherently functional and has direct economic value, courts have been reluctant to find fair use absent significant transformation. While much remains to be seen, it is unlikely that LLM makers will be given carte blanche to reproduce code without a license. A related challenge concerns GPL content, which imposes viral licensing obligations that models cannot readily disclaim if reproduced.
• • Ownership. Ownership of newly-generated code seems straightforward, but a few wrinkles remain. Copyright requires human authorship and originality. Absent meaningful human contribution, model-generated code lacks copyrightability, even if fixed in a tangible medium. Human-tweaked code can qualify, but only where the tweaks themselves are sufficiently creative.
  • Software Patents & AI. Software patents raise similar questions. While algorithms generated by AI cannot be patented in the name of the model, human inventorship is still required. Companies seeking patent protection for AI-assisted code should carefully document the human contribution to the inventive step.
• Fashion & design. Fashion is set for major disruption with AI-generated designs. It is still uncertain whether AI-generated clothing and accessory designs can enjoy trade dress or design patent protection. Here is a look at each category:
  • Design Patents. Design patents in the United States can be rejected if they are not novel. Given the plethora of AI-generated designs, there may be an uptick in rejections for novelty. The flood of AI-generated designs also makes prior art searches extremely difficult for both applicants and examiners.
  • Trade Dress. Trade dress refers to the legal protection of the overall look and feel of a product or service, including its design, packaging, and overall presentation. AI can both create and infringe upon trade dress. Given that AI relies on pattern-matching, it is likely to recommend looks that have worked well in the past, increasing the chances of infringing designs.

3.     Commercial & Compliance Considerations

• Licensing strategies. One way to deal with the onslaught of AI is to preemptively license content for use in AI models. News Corp., the parent of the Wall Street Journal, struck a deal with OpenAI for use of its content to train its models. Other companies, such as the New York Times, have brought litigation, seeking better royalty terms for use with models. While specific strategies are dictated by the content mix and its uses, the key to success is to be proactive. Businesses should not assume that AI models don’t pose a legitimate threat to their existing revenue streams.
• Risk allocation in contracts. It is 2025—AI should be covered in nearly all contracts. Warranties and indemnity clauses should explicitly cover AI content creation.
  • Warranties. If your company is concerned about AI use in content development, you can mandate in the warranty section that: (i) your counterparty does not use AI; or (ii) has sufficient human oversight and review of AI-generated content.
• • Indemnity. Companies that are being licensed or acquiring rights to content should ensure that the IPR indemnity clause covers any AI-generated content. This prevents the producing party from pointing the finger at a third-party vendor.
  • Vendor Strategies. Vendors who are using AI, by contrast, should seek to disclaim warranties on third-party AI content featured in their deliverables and seek indemnity from their vendors. They should also expand the force majeure definition to cover AI-related factors.
• Trade secrets. AI poses obvious risks to protecting trade secrets. Employees can leak information through use of LLMs and other platforms. There is also the question of whether methods developed through the use of AI are eligible for trade secret protection (it depends). This is where careful planning comes in:
  • Trade Secret Development. Businesses should make sure that their inputs are not used to train models for access by third parties—this defeats both the “reasonable measures” and the “independent economic value” requirements for trade secret protection.
  • NDAs & Maintaining Confidentiality. Smart companies should also revise their employee and partner NDAs to cover use of AI models to prevent leakage of proprietary information.

4.     International Divergence

• EU AI Act & Copyright Directive. The EU AI Act requires disclosure and watermarking of AI-generated content, which has a direct knock-on effect on copyright claims. If disclosure is mandatory, it becomes harder to argue that a given work had sufficient human authorship to qualify for copyright. The EU Copyright Directive already tightened platform liability and licensing obligations. This framework could expand to cover text and data mining for training AI—raising the bar for companies that want to use copyrighted material without permission. AI-generated designs may face stricter novelty rejections for design patents/registered designs in the EU, given the flood of AI outputs.
• China’s Generative AI Measures. China requires watermarking of AI content and impose obligations for protecting “network data.” This can impact copyright by making AI authorship more transparent and easier to challenge. The regulations explicitly focus on preventing data leakage. For businesses, that means stronger alignment between trade secret management and AI compliance. China hasn’t recognized AI as an inventor, but regulators have signaled openness to considering AI-assisted inventions if human inventorship is clearly documented.
• U.S. Copyright Office guidance. The U.S. Copyright Office has been the most explicit: no copyright for AI-only works, but human modifications can qualify. Its recent reports emphasize human authorship as the dividing line. The Federal Circuit’s DABUS decision underscores that inventorship must be tied to a natural person. Regulators haven’t yet carved out exceptions for AI-assisted inventions. The U.S. has avoided broad AI laws and taken a hands-off approach to AI regulation. However, litigation (e.g., NYT v. OpenAI, Getty v. Stability AI) is pushing courts to decide how training data and fair use apply to IP.

5.     Emerging Trends & Gray Areas

• Moral rights. There is considerable question as to whether AI’s mimicry of style infringes on the personal rights of authors in jurisdictions with strong moral rights protections. There was considerable debate earlier this year regarding ChatGPT-generated images in the style of Studio Ghibli, South Park, and other well-known animators. These images don’t appropriate any copyrighted work—or any other IP such as trademarks or trade dress—but they clearly utilize a unique artistic style. It remains to be seen what steps governments and artists can and will take to limit this practice.
• AI deepfakes & publicity rights. We are still in the early stages of the use of deepfakes. While the focus has been on safety and security, AI-generated deepfakes pose considerable risk to a person’s publicity rights. Using AI to replicate a celebrity’s voice or image can considerably reduce the amount of income a celebrity can earn through endorsements. It can also unfairly sway consumers. Tools—including personality right legislation and unfair competition law—exist but are largely inadequate.
• Open-source models & derivative works. Open-source models pose a unique risk to IP protection. Unlike closed-source models in which the financial incentives are clear, open-source models are open to the public free of charge. This creates considerable opportunity to generate derivative works free of limitation. To the extent that such models borrow from existing works to create derivatives, there is considerable risk of weakening existing protections. Unlike with closed-source, proprietary model makers, open-source models don’t offer creators a chance to monetize by proactively licensing their content.

6.     Practical Takeaways

The challenge posed by AI to existing intellectual property norms is tremendous. Businesses will need to rethink their approach to both generating and using ideas. Einstein called creativity ‘intelligence having fun.’ For companies, the fun lies in creatively applying their intelligence to manage AI strategy. Here are a few practical takeaways:

• Human Oversight. To preclude the argument that a creative work lacks authorship, make sure that humans contribute to any creative process. Ideas generated purely by AI may not enjoy copyright protection, but humans can use AI as a tool. To use Steve Jobs’ memorable phrase, think of AI as a bicycle for the mind, not an alternative to it.
• Proactive Licensing. Companies in the content-generation industry need to consider a licensing strategy. The reluctance to offer crown jewels to an arriviste competitor is understandable. But one doesn’t want to be like a buggy maker that didn’t adapt to the rise of the automobile. By proactively licensing content to makers of LLMs, content producers ensure a steady revenue stream in the face of declining web traffic. This revenue can be used to fund strategic pivots and investment in technical upgrades.
• Contract Clauses. Any contract worth its salt in 2025 must deal with AI. The desired terms may vary by industry and role, but you’ll want to ensure that you’re protected against AI-related risks. If you’re on the buy-side, this means indemnity, human oversight, and ultimate contracting-party responsibility from the seller. If you are on the sell-side, this means getting protection from your vendors, disclaiming warranties and expanding the definition of force majeure.
• Trade Secrets. AI poses unique challenges to trade secrets, both to your existing ones and the new ones you seek to generate. Robust contract terms—both with your business partners and employees—are key. Companies should ensure that their employees avoid disclosure of information to AI models in ways that preclude trade-secret status or give away valuable information. Contracts with vendors should also regulate how information is disclosed to LLM models and who has access to it.
• Global Strategy. Approaches to AI vary considerably across jurisdictions. This divergence includes not just how AI is regulated, but who owns content that it generates. Businesses that rely on AI to generate content need to consider these divergences and adopt a multi-jurisdictional approach to protecting their rights. A good rule of thumb is that a show of human input will be needed to enjoy IP protection. While much is uncertain, companies with a proactive AI IP strategy will be better placed to ride the wave of AI-assisted content generation.

Disclaimer: This blog is for informational purposes only and does not constitute legal advice. Reading or interacting with this content does not create an attorney–client relationship. You should consult a qualified attorney for advice regarding your specific situation. Mehaffy, PLLC disclaims all liability for actions taken or not taken based on this blog.