In the framework agreement, the client demands the transfer of exclusive usage rights to all the delivered content. The agency signs the agreement as usual. However, part of the images and texts come from an AI tool. By doing so, the agency promises something that it may not even have, and the client pays for exclusivity that does not legally exist.
Such constellations are now commonplace, and they are only noticed when someone asks a question or when a warning letter is issued. Three questions determine the risk: Who owns the output, what can you allow third parties to do with it, and who is liable if an outcome infringes on someone else's rights.
clience.legal is a digital business law firm for entrepreneurs. We support the legally compliant use of AI and in the process, bring usage guidelines, contracts, and customer commitments into line. This post answers the three questions and shows which rules actually help in the workplace.
The most important information at a glance
- Purely machine-generated results are not protected: Copyright protection requires a personal intellectual creation that a machine does not provide.
- The human factor decides: Selection, processing, and arrangement can constitute protection; the mere input of a prompt usually does not.
- You can only admit what you have: Without a protective right, there is no exclusive license; however, there are contractual obligations under civil law.
- The liability remains with the user: The company that uses the output that infringes on a copyright is not the creator of that output, but the tool that is used to create it.
- Failure to act applies without fault: A claim for damages exists regardless of whether you were able to recognize the violation.
- One can protect one's own content: A machine-readable usage clause can restrict the use of your content for training.
Whose output belongs to whom?
The starting point is § 2, paragraph 2 of the UrhG. Personal intellectual creations are protected. A machine does not create personal creations, which is why a purely machine-generated result, in the overwhelming majority of cases, does not enjoy copyright protection. It is therefore in the public domain in the sense that no one can claim copyright over it.
Things are different once a human contribution to creation is added. Three constellations are practically relevant: first, the substantial reworking: anyone who rewrites, supplements, and structures a generated text can acquire its own rights insofar as the reworking is creative; second, the selection and arrangement of several elements that can enjoy protection as a collective work; and third, its use as a basis for an independent work.
On the other hand, simply formulating a prompt is usually not enough, because the prompt does not specify the result in its concrete terms. A very detailed prompt can in itself be a work of language, but it does not transfer that protection to the generated image or the generated text. Those who need protection should document their own contribution, for example through milestones and editing steps.
Why lack of protection isn't always a disadvantage
The finding that a result is not protected is usually interpreted as bad news. For many applications, it is not.
Anyone who creates content for their own needs, such as images for the website, illustrations in presentations, or text snippets for internal communication, does not need a copyright. The only thing that matters is that the use is permitted and does not violate any third-party rights. The lack of protection has practically no effect here.
It becomes relevant in three situations: when the content itself is the product, when customers expect exclusivity, and when a trademark is intended to be created, such as a logo or a figurative mark. In these cases, the effort to contribute with one’s own creative input and document it is worthwhile. Alternatively, one can avoid other intellectual property rights by registering a logo as a trademark independently of copyright law, which creates protection on a different basis.
What you can grant to third parties
The lack of protection leads to a practical problem. Rights of use can only be granted to protected works. Anyone who promises a customer exclusive rights to purely machine-generated content is promising something that does not exist legally. The customer cannot then prohibit third parties from using the same or very similar content.
Contractually, this can be addressed. Possible contractual commitments include not using the content elsewhere or not sharing it with third parties. This is not an exclusivity in the sense of copyright law, but it does create a difference between the parties. It is important to openly mention the difference instead of hiding it in a standard clause. The forms of wording that apply here are shown in the key clauses in contracts and in the Drafting of contracts and T&Cs.
Also, check the conditions of the tool used. Providers usually license the output insofar as it exists; however, they often attach usage restrictions and, in some cases, the right to use the input for training. For companies, the latter is often the more important point.
Who is liable for a legally infringing output?
This question has the greatest economic significance. If a tool produces a result that infringes on someone else's rights, and you use it, claims against your company will be filed. The tool is not a liability shield.
- Copyright: An obligation to refrain from infringing under § 97(1) of the UrhG exists regardless of fault. Compensation for damages depends on intent or negligence, with high standards of care being imposed.
- Trademark law: Generated logos, names, or slogans may infringe on the rights of other trademark owners. Therefore, research is essential before using them.
- Personal rights: Portraits of real people and voices infringe on the right to one's own image and the general right to personal identity.
- Competition law: Misleading statements in generated texts, such as regarding properties or test results, are vulnerable, even if they were created unintentionally.
The fact that protected content can contain results is not a theoretical concern. The Munich I Regional Court ruled in a judgment of November 11, 2025 that the storage of protected song lyrics in the parameters of a speech model and their playback in the output are legally inadmissible. The decision is not legally binding, but it shows that even simple inputs can lead to the output of protected content.
Practically, this leads to a verification routine that can be completed in just a few minutes. For images, this includes the question of whether recognizable people, brands, or structures are depicted, and whether the style is so closely related to an existing work that a transfer is likely. For texts, it involves quotes, numbers, and test results that need a source, as well as formulations that appear strikingly specific. Obvious discrepancies can be verified with a simple search, which in most cases is sufficient.
What is important is that this examination is assigned to someone and time is planned for it. A duty to undergo an examination, which is stated in the directive, but has no place in practice, will not be carried out and will not be of any use in an emergency.
Special case program code
With Code, an additional layer is added. Assistive systems are trained on publicly available source code, which is mostly under open-source licenses. If a suggestion identifies recognizable protected passages, the terms of the respective license may apply, including attribution requirements and, in the case of copyleft licenses, the obligation to make the own work open under the same conditions.
For companies that distribute their own software, this is a significant risk because it could affect the business model. Three measures are practical: choosing a tool that provides source information, checking the generated code for discrepancies, and keeping a list of the components used. We explain why caution is also advised when signing contracts below. why contracts should not be drafted with AI tools such as ChatGPT.
A second point concerns the evidence. If you are accused of a violation of the law, it is of great help if you can provide information about the tool used, the input, and the time when the content was created, and the examination that took place. These details can be easily recorded, for example in a simple log field in the editorial system. They do not change the claim for injunction, but can make a difference when determining liability and thus the amount of damages.
Protecting your own content
The reverse direction is often overlooked. Your website, your expert articles, and your product images can, in turn, become training material. § 44b UrhG allows text and data mining, but only if the rightsholder has not declared a restriction on use. For content accessible online, this restriction must be in machine-readable form.
In practical terms, this means technically recording the reservation and additionally explaining it in the website’s terms of use. Whether and to what extent such a reservation takes effect is the subject of ongoing proceedings. The effort is minimal, so the explanation is also useful even if the legal questions have not yet been conclusively resolved.
Identification with customers
Since August 2, 2026, the transparency requirements of the AI Regulation have been applicable. Artificial generated or modified content must be marked, and in systems that interact with humans, the machine nature must be recognizable. The technical, machine-readable marking was postponed to December 2, 2026. Regardless of this, a lack of labeling can be legally offensive if it leads to confusion, for example in the case of purported customer photos or testimonials. We describe the broader context in more detail below. AI: Opportunities and pitfalls.
The blind spot: agencies and service providers
Even if your company strictly regulates the use of AI, risks arise from third parties. Today, agencies, freelancers, and production service providers mostly work with the same tools, often without it being mentioned in the offer.
This shifts the question to the supplier contracts. Three regulations are useful: first, a disclosure obligation: the service provider must disclose whether and for what purposes it uses generative tools. Second, a guarantee that the delivered content does not violate any third-party rights, combined with a disclaimer in the event of claims being made. Third, a regulation on labeling, so that you can fulfill your own transparency obligations.
Conversely, the same applies if you yourself are acting as a service provider. Anyone who signs such a waiver without knowing their own processes assumes a risk that can only be managed with an internal audit routine.
Seven rules for practice
- List of permitted tools: Only authorized tools should be used, the conditions of which have been verified.
- Input limitations: No personal data and no business secrets without verified basis.
- Check results: Before publication, a designated person must review the results, especially when it comes to images, names and numbers.
- Research on trademarks: Check the names, logos and slogans created before using them against existing rights.
- Documenting contribution: Retain the editing steps when protection for the result is required.
- Adjust customer commitments: Do not promise exclusive rights that do not exist; instead, agree on liability regulations.
- Treat code separately: Source verification and license overview for each use in our own products.
These rules apply to one page and can be combined with the data protection requirements in the same directive. We describe how this looks like in the following section. AI and data protection and Liability, copyright and data protection when using AI.
Related: Legal certainty in the digital enterprise and Using AI in the company.
Please tell us the tools used, the type of content produced, and the commitments you make to your customers in your contracts. We will examine where your contracts promise more than is legally possible, create the terms of use, and adapt the clauses accordingly. To do this, arrange an appointment. free initial consultation from 15 to 20 minutes. The professional framework is formed by our advice Compliance and managing director liability, We take on the ongoing monitoring as external legal department.
FAQs
Are AI-generated texts and images protected by copyright?
According to the overwhelming majority of opinions, mechanically generated results are not protected because § 2, paragraph 2 of the UrhG requires a personal intellectual creation. Protection can arise when a person makes a discernible creative contribution, for example through significant post-processing or through a creative selection and arrangement of several elements. The mere input of a prompt is generally not sufficient for this purpose.
Can I grant my client exclusive rights to AI content?
Not for purely machine-generated content, because without a copyright, no license can exist. Guarantees under tort law are possible, such as that you do not use or share the content elsewhere. These have an effect between the parties to the contract, but do not prevent third parties from creating and using similar content. The difference should be explicitly stated in the contract and not hidden in a standard clause.
Who is liable when an AI result infringes on someone else's rights?
The company that uses the output. A claim for damages under § 97(1) UrhG exists regardless of fault; the amount of compensation depends on intent or negligence. In addition, trademark, personality, and competition law claims may arise. The provider of the tool is not always liable for this, as is also stipulated in its terms and conditions.
What did the Munich I Regional Court decide in the GEMA case?
With a judgment of November 11, 2025, the court ruled that the storage of protected song lyrics in the parameters of a language model as well as their playback in the output are both legally inadmissible. In the court’s view, this does not justify the restriction on text and data mining. The judgment is not legally binding. For users, it means that results must be carefully examined before use.
Do I have to label AI-generated content?
The transparency obligations of the AI Regulation apply from August 2, 2026, and require, among other things, that artificially generated or modified content is identifiable. The technical, machine-readable marking was postponed to December 2, 2026. Regardless of this, the lack of labeling can be legally offensive, for example in the case of purportedly real customer images or reviews.
How do I protect my own content from being used as training material?
About a usage clause under § 44b of the German Copyright Act, which must be stated in machine-readable form for content accessible online. A technical deposit is advisable, supplemented by an explicit statement in the website’s terms of use. The scope and effectiveness of such clauses are the subject of ongoing proceedings. The effort is minimal, which is why the statement is also useful in open legal questions.
What applies to program code from AI assistants?
In addition to general copyright law, the license terms of the underlying code must be observed. If recognizable protected code is used, naming obligations and, in the case of copyleft licenses, the obligation to disclose one’s own work may apply. For companies that distribute software, this may affect the business model. A source check and an overview of the components used are useful.
May I enter customer data into an AI tool?
Only with a legal basis and appropriate contractual basis with the provider. For personal data, an order processing agreement is regularly required, and the processing location and any use of the training data must be examined. Business secrets lose their protection if appropriate confidentiality measures are lacking. A checklist with clear input rules is the most practical way.
How do I document my own creative contribution?
Interim results, the steps of the process, and a brief note of the decisions made are helpful. For texts, version numbers are suitable; for images, the source file and the edited file. The goal is to be able to prove in the event of a dispute that the result was not adopted unchanged. The effort is minimal if it is incorporated into the workflow.
When is it worth hiring a lawyer when it comes to AI content?
As soon as you commercialize the generated content or agree to grant customers rights to it. We review your contracts for promises that are not legally valid, create the usage policy, and clarify the treatment of images, trademarks, and code. The economic leverage lies less in individual lawsuits than in the question of whether your customer agreements hold up what they promise.