Can I patent an invention created using AI?
There are several provisions of patent law that do not easily mesh with AI-generated inventions. These are:
the notion of an inventor;
the inventive step requirement (referred to as the “nonobviousness” test in certain jurisdictions); and,
disclosure requirements.
The notion of an inventor
Different jurisdictions have different criteria for inventorship. In the US, the “threshold question in determining inventorship is who conceived the invention. Unless a person contributes to the conception of the invention, he is not an inventor.” Conception is the “the formation in the mind of the inventor of a definite and permanent idea of the complete and operative invention as it is thereafter to be applied in practice". After conception, a person having ordinary skill in the subject matter of the invention should be able to use the invention without undue experimentation.
In the UK, an inventor is the “actual deviser” of an invention. There is a two-step approach to determining inventorship: first, the inventive concept of an application must be identified, after which it must be determined who devised that concept. The “inventive concept” is “concerned with the identification of the core (or kernel, or essence) of the invention – the idea or principle, of more or less general application which entitles the inventor’s achievement to be called inventive”. The inventive concept varies by application, but may reside in an idea, a means of realizing an idea or the combination of individually known elements. Ultimately, the inventive concept must be determined on a case-by-case basis, but superficial contributions, including to patent claims, do not qualify for inventorship.
The inventive step requirement
Generally, for patents to be granted, an invention must be new (i.e., the novelty requirement), involve an inventive step and have industrial application. It is the inventive step requirement that makes it difficult for AI-generated inventions to meet the traditional threshold.
An invention can be said to include an inventive step if the invention is not obvious to a person skilled in the relevant field, as measured against the current state of the field or the “art”. This is different from the novelty requirement in patent law, which is a relatively objective assessment that asks the question: “did someone else get here first?”
By contrast, for the inventive step requirement, the core comparison is done with the prior art. The core question is: “does the prior art get us close enough to the invention that it is an obvious invention to come up with?” To put it another way: “would this invention seem obvious to a person skilled in this field?”
This test exists so that only those inventions that represent significant advances qualify for patent protection. That is because patents can have significant social costs, and too low a bar would allow for excessive patenting, which would ultimately be counterproductive.
For AI-generated inventions, the difficulty of the inventive step requirement lies in the definition of a person “skilled in the art”. According to the EPO guidelines, a person skilled in the art is “presumed to be a skilled practitioner in the relevant field of technology, who is possessed of average knowledge and ability and is aware of what was common general knowledge in the art at the relevant date".
The challenge is that an AI system could have access to far more than “average” or “common general” knowledge in the “relevant field of technology”. This means the scope of “obviousness” for AI is much larger than it is for a natural person. This challenges definitions of both “skilled in the art” and what counts as “obvious”.
The use of AI in research and development (R&D) is gradually becoming the norm, though it is not yet an essential and integrated element of all fields. In any case, once “inventive” AI is a standard part of R&D in a particular field, it may be challenging for a patent examiner or court to determine what an AI would find “obvious.” Going forward, will our definition of a person skilled in the art need to account for their access to AI systems? This is a tricky question, and it may force patent offices to refocus on different factors, such as: objective; economic facts about how an invention was received in the marketplace; whether others tried and failed to generate the same invention; and whether there was a long-felt need for the invention.
Disclosure requirements
One of the central requirements of patent law is that the patent filing must disclose the invention in sufficient detail so that a person skilled in the art can comprehend it. Essentially, the description of the invention should be clear and complete enough for the person skilled in the art to reproduce it without too much difficulty (according to the EPO Boards of Appeal, a “reasonable amount of trial and error is permissible”).
For AI-based inventions, this raises important practical question: what type of information and what level of detail need to be provided in the patent application about the AI system itself? To what extent should the application explain the AI model’s architecture, training data and methods? How can applications satisfy the disclosure requirement while navigating the inherent “black box” nature of many AI systems, where the developers themselves may not fully understand how the model arrived at specific outputs?
"Disclosure" in the AI context carries a second, equally important meaning related to transparency about the inventive process itself. Current patent law in most jurisdictions does not require applicants to explain how an invention was made: only what the invention is and how to reproduce it. This means the use of AI in the inventive process can remain hidden.
This creates a significant policy challenge. If jurisdictions decide that a certain threshold of human contribution is required for patentability (as suggested by the rejection of the DABUS applications in most countries), then patent examiners would need to know not just if but how AI was used in the inventive process. Introducing disclosure requirements to support this would represent a fundamental shift in patent practice, requiring applicants to disclose their methods of inventing. Such a change could have far-reaching ripple effects on how the patent system operates and how innovation is documented and protected.
Can I copyright an AI-generated work?
When AI started appearing prominently in the IP discourse, it was primarily in the context of AI-generated outputs. The central question was whether there is an author of the outputs of these complex systems and, if so, could they trigger a legitimate copyright claim? Academic literature had already been grappling with this issue before the AI explosion of the last decade. The landscape was already complex when machine learning systems began to challenge the allocation of authorship. It has become even more complex considering the widespread availability of GenAI tools.
The two crucial IP concepts in this discussion are “authorship” and “originality".
But before discussing these concepts, we need to address an important underlying point. When trying to determine who should be considered the "author" of AI-generated works and whether these works deserve copyright protection, we must examine the rationale for why copyright law exists in the first place.
There are two main theories that justify copyright protection. The first is called "natural rights justification" and includes ideas like personality rights (works express the creator's personality) and Lockean labor theory (people deserve to own what they create through their work). These theories require a human creator, who is largely missing when an AI system generates content.
The second theory is the "utilitarian justification", which argues that copyright's main purpose is to benefit society by encouraging people to create new works and share intellectual content with the public. The idea is that copyright protections motivate creators to keep creating by granting them certain rights, such as the right to exclusively profit from their work.
So, if we cannot identify a specific person or entity who would be motivated by these copyright incentives, then it can be argued that there is no good reason to randomly assign copyright protection for the sake of it, without any solid theoretical basis.
Authorship
When determining whether AI-generated content can be protected by copyright, the first key question is: who or what counts as an "author?"
The Berne Convention for the Protection of Literary and Artistic Works (the main international instrument governing copyright) does not provide a clear definition of an author. Similarly, European law does not offer a definition of an author, except in two specific technical areas: software and databases.
However, despite the lack of an explicit definition, much of copyright law is human-centric and assumes human authorship. Because copyright law assumes human authorship, an AI-generated work can only receive copyright protection if we can identify a human being who can be called the "author". That person would then own both the economic rights (such as the right to sell copies of the work) and moral rights (such as the right to be credited) associated with the work, where that right is available.
Originality
The second key requirement for copyright protection is "originality". Courts have different understandings of the criterion of originality, depending on the jurisdiction and the type of work.
In the US, the landmark Supreme Court case Feist Publications, Inc. v. Rural Telephone Service Co. established that originality requires two things: independent creation (the work was not copied) and a "modicum of creativity" (at least a tiny spark of creative expression).
In the EU, the situation was chaotic for years because each country had its own definition of originality. While economic rights were harmonized as part of the InfoSec Directive in 2001, the basic question of what works deserve protection varied from country to country. This changed in 2009 with the Infopaq case, where the European Court of Justice created a unified standard: a work is only protected if it represents "the author's own intellectual creation". The Court explained this means the work must reflect the author's "free and creative choices".
Regardless of the legal test used, it is broadly understood that to meet the originality standard, a work must be the product of human intellectual creation involving genuine creative choices.
This creates a major challenge for AI-generated content. When we break down how machine learning and neural networks work (see Chapter 1), we see that a trained AI model is essentially a huge collection of mathematical functions linked together by numeric weights. In such a highly technical, algorithmic process, it is very difficult to identify the "free human creative choices" and "modicum of creativity" that copyright law requires.
However, as the MPI study points out, these AI systems do not operate in a vacuum. They all require some degree of human intervention to get started and guide the process, even if the actual creation is driven by algorithms.
Several high-profile AI-generated works have tested the boundaries of our traditional concepts of originality and authorship, forcing us to reconsider these fundamental principles. Two specific instances of works that were rejected for registration by the US Copyright Office are worth mentioning: Zarya of the Dawn and Théâtre D'opéra Spatial. Both works were generated with Midjourney, a text-to-image GenAI tool.
Zarya of the Dawn is an 18-page comic book. The work was registered at the United States Copyright Office, but this was later revoked as the applicant had not disclosed the use of AI in the application. The applicant, Ms. Kristina Kashtanova, argued that she had provided “hundreds of thousands of descriptive prompts” to Midjourney until the “hundreds of iterations [created] as perfect a rendition of [her] vision as possible". She sustained that she was the author of that work. The US Copyright Office was not convinced, concluding that Midjourney generates images in an unpredictable way and the person who provides the prompts does not “actually form” the generated images and as such is not the “mastermind” behind them. That said, the text and arrangement of the images in the comic are protected by copyright, as they are considered the creative work of Kashtanova and not AI.
The cover page of Zarya of the Dawn generated in 2022 by Kristina Kashtanova using the generative AI model Midjourney.A similar application was registered by Jason M. Allen for the work Théâtre D’opéra Spatial. Allen sustained that he put 624 text prompts into Midjourney to arrive at the initial version of the work, which he then modified further with Adobe Photoshop and another AI tool called Gigapixel AI. The Office found that the work was not eligible for copyright protection as the human creative input was de minimis, with AI-generated elements dominating.
In 2023, the Beijing Internet Court came to the exact opposite conclusion in a landmark copyright infringement case, Li v.. Liu. Plaintiff Mr. Li sued Defendant Ms. Liu for copyright infringement after she allegedly used an image he generated using Stable Diffusion without permission. Mr. Li claimed this violated his right of authorship and dissemination. The court determined that Li's intellectual investment in creating the image through prompts and parameter adjustments met the requirements for a copyrighted work, and ruled that Ms. Liu had indeed infringed copyright. This sets a bold precedent for future cases involving AI-generated content in China.
Also in 2023, the US Copyright Office maintained its restrictive approach by granting only partial copyright protection to another AI-assisted artwork entitled "Rose Enigma", also created by Kristina Kashtanova. The Office registered copyright solely for the human-created sketch that served as the basis for the AI generation, while explicitly excluding all AI-generated or AI-modified portions. As with "Zarya of the Dawn", this case seems to indicate that while humans can retain copyright in their original contributions to AI-assisted works, the AI-generated elements do not qualify as copyrightable expression.
Kristina Kashtanova's original sketch for the work "Rose Engima", 2023 , alongside the final work generated by Midjourney.In January 2025, a work entitled "A Single Piece of American Cheese" received copyright protection in the US. It was created by Mr. Kent Keirsey, CEO of Invoke AI. Using Invoke AI’s tool, Keirsey iteratively modified an initial AI-generated image 35 times, making specific creative decisions about colour, composition and placement of elements. After an initial rejection, Invoke AI successfully appealed by providing documentation demonstrating Keirsey's active creative role, including a time-lapse video. The US Copyright Office determined that the final work contained "sufficient human original authorship,” comparing it to a copyrightable collage where an artist compiles elements created by others into an original composition. The original AI-generated image was excluded from protection.
"A Single Piece of American Cheese" (2025) by Kent Keirsey (right) alongside the AI-generated image that provided the starting point for the work (left).What should I look for regarding IP in AI terms of service?
As AI companies face lawsuits and new regulations, they are using their terms of service (ToS) agreements to shift legal risks away from themselves and onto their users. This was demonstrated by a 2024 study looking at the terms of service of 13 major AI companies and covering text, image, audio and video generators. Three aspects of copyright law were studied in-depth:
ownership of inputs and outputs;
liability if outputs infringe copyright; and,
content moderation with the goal of reducing copyright infringement.
Ownership of input and outputs
While the study found that there is barely any reference in the ToS to ownership of user inputs into GenAI tools, more recent anecdotal evidence shows that AI companies are quite clear about the fact that they use inputs provided by users to train their AI models.
The study also identified a trend where providers assign ownership of outputs to the user, but also grant themselves a broad license to reuse the outputs. Some models (Stable Diffusion in particular) operate on an open-source license in this regard. If you are creating using AI, this means you technically "own" your AI-generated output, but the company can also use it however it wants, essentially nullifying any exclusive rights you might hold.
Liability for copyright infringement
If an AI-generated output potentially infringes on copyright, the AI provider’s ToS typically place liability for this squarely on the shoulders of the user. The providers position their products as neutral intermediaries rather than active participants in content creation.
Content moderation and prompt filtering
Almost all companies use content moderation systems, including:
filtering prompts to prevent users from requesting copyrighted content; and
"notice and takedown" systems (similar to YouTube's copyright strikes).
This is consistent with broader trends in internet law: the more control a platform exercises over content, the more legal responsibility it is considered to have. The Digital Services Act in the EU has been pioneer legislation in this area, imposing serious transparency and reporting obligations and fines for noncompliance. By actively monitoring and removing content, companies are trying to demonstrate responsibility in this regard while still maintaining some protection.
In conclusion, users should be aware that most ToS for widely-used AI products are one-sided contracts that:
put the bulk of the risk on the user;
give the company maximum control over both inputs and outputs; and
frame companies as neutral platforms rather than service providers.
The "output ownership" question will be particularly important to watch in the future, especially as AI systems become more capable and help users create more sophisticated and commercially valuable content.
What disclosure obligations do I have when using an AI-assisted creation?
In commercial contexts, disclosure obligations vary significantly by industry and jurisdiction, but transparency is increasingly becoming both legally prudent and ethically expected. Many professional bodies and sectors (such as academia, law and creative agencies) now require explicit disclosure of AI assistance in any work produced. Even where not legally required, failing to disclose AI assistance in commercial work can create liability risks around misrepresentation, breach of client expectations or violation of professional standards.
Additionally, as mentioned above, if an AI-assisted creation is found to violate copyright, the ToS typically place liability entirely on the user, which could have severe consequences in a commercial context.
Some major AI providers, such as OpenAI and Microsoft, have made commitments to cover some of their users’ legal fees arising from copyright issues with AI-generated outputs. These commitments only apply to commercial and enterprise-level users and are hedged with significant conditions. In any case, they have yet to be tested in practice, as litigation in this space has so far been directed at the AI providers themselves, and direct claims against end users have not materialized.
Beyond formal requirements, commercial disclosure serves important business purposes: it manages client expectations, demonstrates technological competency, and can even be a competitive advantage when positioned as innovation and efficiency. However, the level of detail required varies: you may need to disclose that AI was used and for what purpose, but not the exact prompts or technical details (though it may be prudent to document these in any case). When in doubt, err on the side of transparency, as undisclosed AI use discovered later can damage relationships and professional credibility more than upfront disclosure ever could.
How can I demonstrate originality in an AI-assisted work?
Two landmark decisions – Li v. Liu in China in 2023 and "A Single Piece of American Cheese" in the US in 2025 – provide some guidance for creators who use AI heavily in their creative process. These cases indicate that demonstrating originality in commercial AI-assisted work requires showing substantial human creative input and decision-making throughout the process. For creators, this could involve documenting in detail:
your conceptualization of the original idea;
strategic prompting decisions;
selection criteria for AI outputs and editorial choices; and,
how you combined, refined or transformed AI-generated elements.
The key is proving that the final work represents human intellectual judgment and creative vision, with AI serving as a sophisticated tool rather than the primary creator.
The focus should be on areas where human creativity is most evident: strategic planning, conceptual framework, editorial curation and the integration of AI elements into a cohesive whole. To support this, it would be useful to maintain work files that show iterations, rejected alternatives, reasoning for creative decisions, or even video documentation of the creative act, as was provided for in "A Single Piece of American Cheese".
How can I protect my business's IP when using AI tools?
Given the uncertain legal landscape, the burden of protection remains squarely with users for the time being.
If you are using AI tools in a commercial context, regardless of jurisdiction, protecting your business IP requires a multilayered approach starting with careful vendor selection and contract negotiation. Choose AI services that offer robust data protection guarantees, including explicit commitments not to use your data for model training, clear data retention and deletion policies, and compliance with relevant data protection regulations. Negotiate specific contractual protections around data handling, including data residency requirements, encryption standards and the right to audit data handling practices. Many leading AI providers now offer professional and enterprise-grade tiers with enhanced privacy protections designed for commercial use.
Internally, it is wise to implement robust data governance policies that classify information by sensitivity level and restrict what can be input into AI systems. Never input highly sensitive information like trade secrets or confidential business strategies into AI tools, unless you have explicit contractual protections. Consider using on-premises or private cloud AI solutions for the most sensitive work, and establish clear employee guidelines about what data can and cannot be shared with AI systems. Regular security audits and monitoring of AI tool usage can help ensure compliance with your data protection policies and identify potential risks before they become problems.
How can I prevent my business's IP from being used to train an AI model?
Preventing your business IP from being used for AI training requires both technical and legal measures.
Some jurisdictions maintain opt-out provisions for copyright holders who do not wish to have their works used in TDM activities. So, from a legal perspective, clear copyright notices and licensing terms that explicitly prohibit use for AI training may provide some protection. Some companies are exploring digital watermarking and other technical solutions to track their content and prove unauthorized use. Technical barriers, including robots.txt files, can explicitly disallow AI crawlers, as can access controls on public-facing content and clear terms of use that prohibit scraping or automated data collection.
Be prepared to issue takedown notices when you discover unauthorized use, and consider working with legal counsel to send cease-and-desist letters to AI companies that may be using your content. While the legal landscape is still very much in flux globally, taking proactive steps to signal your intent not to participate in AI training strengthens your position in any future disputes and may deter casual scraping of your content.