AI Tools and Licensing: Who Owns the Output?

Intellectual property law is based on the idea of human originality, and ownership of creative works. But the creation of artificial intelligence (AI), and its ability to generate creative works in tandem with humans, has raised a number of questions about how these laws apply.

Legal questions have also arisen in relation to who owns the output of generative AI models, how models deal with copyrighted input, and what models are trained on.

This article covers what AI-generated works are, what IP law is, how it applies to AI models, who owns the input for these models, who owns the output, and what AI companies have set out in their Terms and Conditions. Finally, we'll go through some tips for how you can protect your own IP when using generative AI.

Let's begin.

What are AI-Generated Works?

AI-generated works are outputs from AI models, particularly creative works such as texts, audio files, video files, images, or computer code. These models can also generate synthetic data sets, designs for products, and other types of content that could be original or valuable

AI-generated works are created when a user inputs information and prompts into an AI system, which then provides an output in response

Many of these AI-generated works are creative in nature, such as texts, images, or computer code. Some of these works and creations would usually be covered by intellectual property law. However, since these works are created in collaboration with AI models, the question of who owns the output becomes a slightly more complex one.

What is Intellectual Property Law?

Intellectual property law is a set of legal rules that covers "intellectual creations", i.e. original ideas that people have, that are realised in some physical or digital form. This could include literary works, artistic works, designs for a new technology or tool, inventions, business marks and logos, musical works, or commercial secrets or knowledge

The main types of intellectual property are:

• Patents
• Trademarks
• Copyrights
• Trade secrets

Intellectual property law protects these types of property from infringement or theft by other people. The owner of the intellectual property is usually the creator of the work. When you create work in tandem with generative AI, such as using an AI model to create an image or a piece of writing, it can become confusing about who the "creator" of the work is.

Let's take a look at how intellectual property law has been applied to works created with generative AI in different countries.

How Does Intellectual Property Law Apply to AI Generated Works?

Intellectual property law varies by jurisdiction. In most countries, intellectual property laws only protect works by human creators

This creates complications when another "entity", in this case, an AI system, takes additional creative steps using human input. Particularly when training data or other input could be itself copyrighted, the legal consequences are complex

Let's take a look at IP laws in some jurisdictions and how they can apply to AI generated works.

IP and AI in the EU

At the moment, works created by AI are not able to be protected by copyright in the EU

The International Association for the Protection of Intellectual Property (IAPPI) explains that the Court of Justice of the European Union found in the legal case Infopaq International A/S v. Danske Dagblades Forening (C-5/08, 2009), that a work is considered original if it is "the author's own intellectual work", as a result of their "personal contribution."

The European Parliament also released a research study in 2025 which found that human authorship is a critical aspect of eligibility for copyright protection. The study concluded that "EU copyright law only protects works that are the result of a human's intellectual creation". In addition, it explicitly excluded AI works, stating that "fully autonomous AI-generated outputs, without meaningful human input, are excluded from protection."

The words "meaningful human input" are important, however. The same study also found that "legal eligibility depends on the degree of human involvement." This means that AI tools used "under human creative control may lead to protectable works"

The study's key conclusions are below:

The threshold for "meaningful human input" can be higher than many assume. In the United States, the Copyright Office has clarified that simply entering a short prompt, even if refined over multiple attempts, rarely meets the human-authorship requirement. Instead, protection is more likely when the human contribution involves substantive creative choices, such as editing, combining AI outputs with original material, or directing the AI's role in a broader creative process.

Businesses relying on AI-generated content should therefore document their creative steps beyond prompt submission to strengthen any future copyright claim.

The question of how much "human creative control" is required for a copyright claim is still unknown. However, one case, Football Dataco Ltd v. Yahoo! UK Ltd and Others (C-604/10, 2012), found that eligibility for copyright protection depended on "originality"

Whether something is original was determined by whether the "author expresses his creative ability in an original manner by making free and creative choices"

This was contrasted with a process that is governed only by "technical considerations, rules or constraints which leave no room for creative freedom."

As AI models use technical processes, rules, and constraints to produce outputs, it could be argued that their outputs are not truly original and therefore not able to be copyrighted.

The wording of the case is below:

In addition, the European Patent Office (EPO) already rejected two patent applications which listed an artificial intelligence model (DABUS) as an inventor. Their reasoning was that "an inventor designated in the application has to be a human being, not a machine."

IP and AI in the US

In the US, similar positions as the EU have been taken in court cases and by regulatory bodies. For example, the U.S. Copyright Office released a three-part report in January 2025 called "Copyright and Artificial Intelligence"

This report came to a number of conclusions about how copyright applies to AI-generated material. For example, copyright protects original expression created by a human author, even if it also includes AI-generated material. However, copyright does not extend to purely AI-generated material.

In addition, case law such as Thaler v. Perlmutter (2023) affirmed that human authorship is a necessary requirement for works to be able to be copyrighted. The court stated that "many of the Copyright Act's provisions make sense only if an author is a human being."

Here's a portion of the case, which identifies authors as human beings, and machines as "tools" used by humans. Tools are used by humans in the creative process, and are not creators themselves.

Similarly to the EU, the U.S. Patent and Trademark Office (USPTO) also would not allow DABUS, the artificial intelligence model, to be registered as an inventor. This was litigated in court, in the case Thaler v. Vidal (2022). The court stated that " there is no ambiguity: the Patent Act requires that inventors must be natural persons; that is, human beings."

The court's wording below:

Like in the EU, in the US works must be authored by humans, or have substantial human influence and creative control to be able to be copyrighted. If an AI system is used as a tool by a human, the ability to protect the output with intellectual property laws will depend on many individual, case-by-case factors.

IP and AI in the UK

In the UK, the use of computer generation to create works has been explicitly covered in the UK's Copyright, Designs and Patents Act. Under section 9(3) of the Act, it states that the author of a computer-generated work is "the person by whom the arrangements necessary for the creation of the work are undertaken".

However, the question of who "the person" is, who has undertaken arrangements necessary for the creation of the work, is yet to be tested. This could be the user of the AI program, or potentially the developer of the AI program.

In January 2025, the Intellectual Property Office (IPO) in the UK released new guidelines on patent applications relating to artificial intelligence. These guidelines were in relation to the Patents Act and the Patents Rules in the UK

The question for the IPO was less about whether AI-generated works can be patented, but rather whether AI programs themselves can be patented. This is because under the Patents Act, there is a so-called "computer exclusion". This means, in the words of the IPO, that "UK patent law excludes from patent protection inventions relating solely to a mathematical method or a program for a computer." The IPO decided that an AI program can be patented if it "makes a technical contribution to the state of the art."

Other than this, there is no ICO guidance as yet around how AI technologies and IP law interact in the UK.

IP and AI in other parts of the world

In China, Beijing Internet Court has recognized copyright in AI-generated works with substantial human input.

Israel allows certain fair-use exceptions for AI training data.

Now let's take a look at IP law relating to inputs and outputs for AI technologies.

New and Upcoming Laws on AI and IP

As the development of AI is relatively new and also rapid, lawmakers are scrambling to try to create legislation that regulates its use

In relation to the use of generative AI and copyright laws, some legislation has been developed specifically with these issues in mind, specifically, the EU AI Act in the EU, and the Generative AI Copyright Disclosure Act in the US

The EU AI Act is partially in force, but not completely, while the Generative AI Copyright Disclosure Act is not yet in force. Rather, it is a Bill that has been introduced into the US Congress

The way these laws will be interpreted by courts is still to be seen, as the EU AI Act is still new, and the Generative AI Copyright Disclosure Act is still in an early stage.

Let's take a look at each of these.

EU AI Act

The EU AI Act was passed into law in 2024. Since then, its provisions have slowly come into force. The Act contains two obligations related to copyright as part of its provisions, neither of which specify who owns the output of an AI model

However, it does state in Article 53 that providers of general-purpose AI (GPAI) models must take several compliance steps, two of which you can see below:

In Article 53(1)(c) providers of GPAI models must "put in place a policy to comply with Union law on copyright and related rights."

In addition, they have to "draw up and make publicly available a sufficiently detailed summary about the content used for training of the general-purpose AI model."

These two provisions relate to how copyright should be dealt with when releasing (and using) AI models. This is because these models usually need to be trained on large datasets that often contain copyrighted material. To find this information, models usually use "web crawling" approaches, collecting data from the open internet. This puts copyright holders at risk of their work being used by these models without their consent

In relation to GPAI, the European Commission has also released a guide, called "The General-Purpose AI Code of Practice". Chapter 2 of this guide covers copyright. It states that GPAI model providers must "reproduce and extract only lawfully accessible copyright-protected content when crawling the World Wide Web," and should "mitigate the risk of copyright-infringing outputs."

However, the guide does not relate to who owns the output of generative AI or GPAI models.

Generative AI Copyright Disclosure Act

The Generative AI Copyright Disclosure Act was introduced as a Bill in Congress in 2024. It also focuses more on how AI deals with copyrighted content, rather than figuring out who owns the works produced by AI models.

If you make a training data set or alter one, and use it to build a generative AI system, you have to submit certain documents to the "Register". In this case, the Register means the Register of Copyrights.

You are required to provide:

• A sufficiently detailed summary of copyrighted works used in the training dataset or to alter it
• The URL for the dataset

The purpose of the law is to increase transparency around copyrighted inputs used in AI training. However, the law does not tackle the question of who owns the outputs.

Let's go further into these issues now.

Who Owns the Input for AI Technologies?

Ownership of input into AI models is an important question to consider, as much training data or other input data is also copyrighted material, as highlighted by the two new and proposed EU and US laws above.

The material used as AI inputs can vary from text, images, whole data sets, computer code, or anything else that can be uploaded to a model or tool.

If you upload your own information, such as your thoughts, computer code, photographs and so on, you would still own the copyright in that data

However, it's important to read the Terms and Conditions set out by different AI companies before you upload data to their models. This is because some models will use your information for training, while others won't

In some cases you can explicitly opt out, while other tools do not have this option unless you pay for the model

When uploading material that doesn't belong to you, you could be violating the copyright of someone else. Or, when AI companies use training data or scrape the internet for text, they are also potentially violating the intellectual property of others. This issue has already led to a number of lawsuits.

For example, the New York Times sued Open AI and Microsoft, claiming that Chat GPT and Bing had been trained on copyrighted material. The lawsuit claimed that these "generative artificial intelligence ("GenAI") tools rely on large-language models ("LLMs") that were built by copying and using millions of The Times's copyrighted news articles, in-depth investigations, opinion pieces, reviews, how-to guides, and more."

The lawsuit below:

They alleged that the "Defendants repeatedly copied this mass of Times copyrighted content, without any license or other compensation to The Times." In contrast, OpenAI relied on the defense of "fair use". Harvard Law School discussion of the case noted the factors going into the "fair use" defense:

The factors include:

• Whether the infringing work is transformative or just copying
• Whether the original work has a particularly original, creative character
• Whether a large or small amount was used
• The impact on the market for the original

The case is not yet decided.

In another court case, Getty Images v Stability AI, Getty sued Stability AI for using copyrighted Getty Images content to train its model. As a result of the training, Stability AI produced a number of images with similar-looking watermarks as Getty. Here's an excerpt from the case below:

This was particularly problematic because many of the images produced by Stability were explicit or inappropriate images. Getty claimed that Stability's AI-generated images would cause "detriment to the repute of [Getty's] trade mark".

These issues around copyright and other IP infringement when training AI models are increasing, and continue to be litigated.

What Do AI Companies Say About Ownership?

Alongside explicit IP laws and the litigation of IP issues, different AI companies have also produced their own Terms and Conditions, in which intellectual property is outlined

It's important to understand what you are agreeing to in relation to generative AI programs, and the inputs and outputs you are working with

While these Terms and Conditions shape the contractual relationship between the AI provider and the user, their enforceability is not absolute. Some clauses, especially those granting the provider broad rights over user content, may face challenges if they conflict with local consumer protection laws or copyright principles.

For companies integrating AI into core products or workflows, relying solely on click-wrap terms is risky. Instead, consider negotiating tailored contractual provisions or data processing agreements with providers, ensuring that rights, usage limitations, and confidentiality obligations are expressly defined in a binding commercial contract.

Let's take a look at some clauses from a few common generative AI models, including:

• ChatGPT
• Claude
• Gemini
• Grok

In each section, we'll go through Terms and Conditions on who owns the output of the AI model, how ownership (going in) is dealt with, and how the AI company makes use of inputs for training.

ChatGPT

First, let's take a look at ChatGPT's Terms and Conditions.

In the section below, Open AI states that users retain their ownership rights to input, and also own the output. In addition, Open AI assigns their "right, title, and interest, if any", in and to any output, to the user

You can also see below that content (input or output) can be used to provide, maintain, develop, and improve the AI model services

There is also an opt-out provision so that users can specify that they don't want their content used to train models. However, this is an opt-out approach, so users have to take an active step

Finally, if someone believes that Open AI has infringed their copyright, the person can contact the company

The consequences may be that infringing content may be deleted or disabled. This does not mean that a copyright infringement claim could not still be brought, however.

Claude

For the Claude model, made by Anthropic, similar terms are present in its Terms and Conditions.

Users are responsible for inputs, and must ensure that they have the intellectual property for those inputs. This includes rights, licenses and permissions

You can also see that users retain rights, title, and interest in any inputs that they may have already had rights to. In addition, like Open AI, Anthropic assigns any of their rights, title or interest in any outputs, to the user.

In addition, Anthropic uses "Materials", i.e. inputs and outputs, to provide, maintain, and improve their services. In addition, they may be used to develop other services.

However, it explicitly states that data is not used to train the model, except for when information is publicly available, when direct feedback is provided, or when an output is flagged for a trust and safety review

Now let's take a look at a couple more models.

Gemini

Like other models, Google Gemini has a similar set of Terms and Conditions. For the creation of original content, Google explicitly states that it won't claim ownership over that content. This remains the case if this is created by the Gemini model.

When it comes to how Google uses data, there is a difference between paid and unpaid services. In their section below, when "Unpaid Services" are used, Google uses the content users submit to AI tools

This content is then used to provide, improve, and develop Google products, including machine learning technologies like generative AI models

For Google "Paid Services", however, prompts and responses (inputs and outputs) are not used by Google to improve its products

This creates a distinction between how the model interacts with different user accounts (paid vs. unpaid) in terms of content usage and further development of the model.

Grok

Finally, let's take a look at Grok, made by xAI. Grok also states that users are responsible for user content, including that it doesn't violate any laws. This includes copyright and other intellectual property laws

You can also see that users retain any rights they have to their content. With regard to outputs, xAI asks to be attributed as a generator of the output

Below you can also see how xAI deals with user content. Unlike the other models, xAI asks for an irrevocable, perpetual, transferable, sublicensable, royalty-free and worldwide right to the content. This is a right to use, copy, store, modify, distribute, reproduce, publish, display, make derivative works of, and aggregate.

The purpose of this content use is to provide and maintain the Grok service, improve xAI's products, and for other business purposes

Below you can also see how xAI deals with data for product development or model training

The section outlines how users can select whether or not their content can be used for model training. Any private chats and user content requested to be deleted, will also be deleted within 30 days.

So Who Owns the Output of AI Technologies?

The question of "who owns the output" of AI models is still evolving, and depends significantly on the specific factors of each individual case

In many cases, the Terms and Conditions of publicly-available AI models specify that users continue to own both their (previously owned) input into any AI model, and any output that is produced.

In addition, copyright and other intellectual property laws around the world, such as in the US, UK, and EU, generally maintain that content produced alone by an AI model is not able to be protected by intellectual property law

However, content produced in collaboration between a human and a generative AI, could be protected with copyright or patent, when filed by the human individual who worked on creating it.

Some models contain terms that allow AI companies to use both input and output for model training, however. Others allow you to opt out of this usage, or, in the case of Google Gemini for example, to choose a paid model in which your data is not used for model training. Consider this when using these models, as once your data has been input, your control over the outputs or training uses is much less.

How Can You Protect Your IP When Using AI?

When using AI models, regardless of whether you are an individual, business owner, or software developer, there are a number of steps you should take to protect your IP when using generative AI models

1. Review the Terms and Conditions: First, make sure you review the Terms and Conditions of whatever model you are using. Whether it's Chat GPT, Gemini, or Grok, you need to understand whether you retain the rights to your inputs, and who owns the outputs. Make sure you check any updates to Terms, as they could change over time.
2. Avoid uploading third-party data: Make sure you don't upload any third-party content that you don't have the copyright to. While this isn't your IP, you could be exposing yourself to unnecessary risk by sharing protected data with a generative AI model.
3. Opt-out of data use for training or other processes: If possible, opt out of your data being used for model training or other business purposes. If you input data that belongs to you, and it is used for model training, it could appear later as an output, in either a modified or copied form. Many AI systems remain "black boxes", i.e. their inner workings are not easily comprehensible. This means you cannot be sure how copyrighted information you enter could come out again later
4. Record your work: If you want to work in collaboration with an AI, separately record your contributions and ideas. This makes it easier to determine what work was done through your creative effort and originality, and what was supported or created by the generative AI model.
5. Use paid versions if possible: Like Google Gemini, sometimes paid versions of generative AI tools come with better conditions and more protections for users. If you can afford it, pay for these protections to help protect your data, as well as your IP rights.

Summary

The complexities of intellectual property law when it comes to generative AI are still evolving. In multiple jurisdictions, including the EU, US, and the UK, laws and advice from regulators are beginning to tackle these questions.

In addition, court cases are being brought to determine the finer points of when data can be used, what happens with copyrighted data being entered into AI models, and who owns the output

Make sure you always check the Terms and Conditions when using an AI model, avoid uploading copyrighted content, and protect your own IP with clear recording, paid models, and opting out of data for model training where possible