Skip to content
NLEN
Illustration: AI and intellectual property: rights to output

AI and intellectual property: rights to output

By Ivo Donker — compiled with AI support (Claude & Gemini) · Last updated: 6 August 2026

When deploying generative artificial intelligence for business purposes, organizations quickly face the question of who owns the rights to generated text, source code, images, and data analyses. Where traditional software development and content creation rely on clear-cut copyright frameworks, the use of large language models (LLMs) introduces an interplay of IP law, license terms, and contractual agreements. Many organizations wrongly assume that receiving an AI response automatically means they acquire an exclusive intellectual property right.

To keep legal and operational risks manageable, organizations need to move away from simplistic assumptions about ownership. The issue is best unpacked by distinguishing between the scope of statutory protection, contractual transfer, and operational control. This article analyzes how the rights surrounding generative AI output interlock, and which measures an organization should take within its own development and management processes.

The three legal layers of an AI system

When an organization uses an LLM or a generative AI application, three separate layers of intellectual property become intertwined. To determine where which rights lie, these layers must be assessed strictly separately.

1. Rights to the training data

The first layer concerns the dataset used to train a model. This data consists of copyrighted books, articles, code repositories, and web pages. The legal question at this layer concerns the lawfulness of scraping, storing, and processing this material by the model builder. For the business end user, this is primarily an indirect risk: if the training data was collected unlawfully, infringement claims could affect the continuity of the service or the lawfulness of the model itself.

2. Rights to the model and the weights

The second layer is formed by the software architecture, the trained parameters (weights), and the specific model files. For proprietary models, the intellectual property rests entirely with the vendor. For open models, the weights are made available under a specific license. The user only receives a right to use the functionality of the software or the weights, but doesn't become the owner of the underlying AI technology.

3. The status of the generated output

The third layer is the result that comes out of the model after entering a prompt: the output. This is the layer organizations work with on a daily basis. The question of which rights apply to this specific output is separate from who owns the model or who holds the rights to the training data. A license on a model doesn't automatically mean the generated output is protected under IP law, or that the user is the only one who can lay claim to that result.

Ownership versus contract: what do vendor terms actually regulate?

In practice, the question "who owns the AI output?" is rarely a pure ownership question in the IP-law sense, but rather a contractual arrangement between the customer and the AI vendor. Copyright arises automatically under specific conditions by operation of law, and a vendor cannot "create" it if the law provides no basis for it. A vendor can, however, contractually waive its own claims to the output and transfer any rights to the customer.

Most business SaaS terms for AI services contain a clause in which the provider declares that it transfers all rights, title, and interest in the generated output to the user, insofar as those rights legally exist. The crux of the matter lies in that last qualifier. When a vendor states that "all rights belong to the customer," it's only handing over the rights it can actually assert itself. If no copyright exists on the output under the law, the vendor legally transfers nothing except a contractual promise that it won't sue the customer for using it.

Note: A contractual transfer from the vendor to the customer offers no protection against third parties if the output infringes on existing copyrights. For the precise structuring of commercial agreements, we refer to the guide on AI contracts and SLA negotiations.

Moreover, terms differ significantly by service tier. With free or consumer-facing interfaces, providers often retain the right to use entered prompts and generated output to retrain their models. With business enterprise subscriptions, this kind of processing is usually excluded. Choosing the right contract form is therefore a first requirement when selecting an AI vendor.

Being allowed to use versus holding an exclusive right

Organizations frequently conflate two concepts: the right to use a work and the possession of an exclusive right to exclude others from that work. This distinction has direct operational consequences.

When an organization generates marketing copy, a logo design, or a software module via a generative AI tool, the organization is free, under the terms of the user agreement, to use, publish, or sell that result. However, this right of use is not the same as an exclusive copyright. Without an exclusive right, an organization cannot legally force a competitor to stop using an almost identical text or image.

This risk manifests because AI models are deterministic and statistical in nature. Two different users who enter a similar or identical prompt with the same provider can receive outputs that hardly differ in content or form. Because the AI provider grants both users the same right of use, a situation arises in which multiple parties lawfully use the same content without any of them being able to claim sole rights.

The role of human creative input

International and European copyright law rests on a central standard: a work only qualifies for copyright protection if it is the result of a human author's own intellectual creation. There must be free and creative choices made during the work's creation process.

For AI-generated output, this means that purely machine-generated material isn't protected by copyright. Simply entering a short text instruction (prompt) generally doesn't count as sufficient creative direction. A prompt such as "Write a press release about a new software product" contains insufficient formative elements; after all, the actual expression of the text is determined by the algorithm.

A copyright-protected work can only arise when the human role is dominant and directive. This can take shape in three ways:

The legal dividing line between 'unprotected AI generation' and an 'AI-assisted copyrighted work' is not sharply defined. For now, organizations should assume that raw AI output is free of rights in the public domain, and that protectability can only be built up through demonstrable human editing. You can read more about the societal and legal debate on this topic in the article on AI and copyright on nieuws.llmnet.nl.

The risk of infringing third-party rights

Besides the question of whether an organization can claim rights to AI output itself, there's the risk that the generated output unintentionally infringes on third-party rights. Because language and image models are trained on billions of existing works, there's a danger that a model reproduces patterns, passages, or stylistic elements literally or almost literally.

This risk isn't equally significant for all types of output. For lengthy texts, the chance of an identical paragraph is relatively small, unless the training material contained a specific source extremely frequently. The risk peaks, however, with short, recognizable, and strongly protected forms:

Copyright infringement is an objective fact: the fact that an organization didn't know the AI tool had reproduced an existing work doesn't absolve the organization of liability when the output is made public or commercially exploited.

Indemnifications from AI providers: what do they cover and what don't they?

To reassure business customers, several major AI vendors offer so-called IP indemnification or indemnification arrangements. Such arrangements promise that the vendor will cover legal costs and any damages if a customer is sued by a third party for copyright infringement resulting from the use of AI output.

Although these indemnifications sound commercially attractive, organizations need to thoroughly analyze the fine print. These guarantees usually come with strict conditions attached:

Aspect Standard promise Contractual reality & limitations
Built-in filters Indemnification against infringement claims. Only applies if all built-in safety and copyright filters remained enabled.
Prompt restrictions Coverage for generic output. Lapses as soon as the user deliberately sought to reproduce existing works.
System integration Compensation in the event of claims. Applies only to the direct model output, not to the end product in which the output is incorporated.
Commercial limits Full legal coverage. Often capped at the total amount the customer has paid for the service over the past 12 months.

Indemnifications provide an important safety net, but they don't cover reputational damage, the cost of operational delay, or having to pull an already-launched product from the market. An organization remains responsible for its own risk assessment.

Open models and license structures

When using open models (whose model weights are run locally or on the organization's own cloud infrastructure), a different dynamic applies than with proprietary SaaS services. Here, the model's software license determines which rights and obligations the organization has.

It's crucial to understand the distinction between the license on the model weights and the status of the output generated with the model:

Licenses on weights

The license on the weights (such as Apache 2.0, MIT, or a customized commercial license like the Llama license structure) governs what an organization can and cannot do with the model files themselves. Think of modifying the architecture, redistributing the model, or deploying the model for commercial purposes above a certain number of active users.

Rights to generated output

Most open-source licenses don't impose ownership claims on the generated data. The model's creator generally doesn't claim any rights to what users get out of the model. However, usage restrictions (Acceptable Use Policies) may be woven into the license. These may, for example, prohibit using the output for specific medical, military, or deceptive purposes.

For a deeper analysis of the documentation accompanying open models, we refer to the guide on reading model cards and licenses on hub.llmnet.nl, as well as the article on the licensing debate around open models on nieuws.llmnet.nl.

Specific risks in software development and source code

Using generative AI in software development (via code assistants and integrated LLMs) carries specific legal risks. Source code is inherently highly structured and subject to clear licensing models, which makes infringement easier to establish.

The biggest risk with AI-generated code is license contamination. If a model has been trained on repositories with a so-called 'copyleft' license (such as GPL or AGPL), and the model generates a code fragment that substantially matches that trained source code, the question can arise whether the proprietary software into which this code is integrated must also fall under that open-source license.

To prevent license contamination and IP risks in software projects, organizations should not check source code management after the fact, but build it into the development pipeline:

Practical measures for the organization

To ensure the use of AI tools doesn't become a legal time bomb under an organization's IP portfolio, concrete operational measures are necessary. The point is that an organization must always be able to demonstrate how content or code came into being.

1. Determine where generative output may end up

Not all business assets require exclusive copyright. It's wise to make a clear distinction within the organization between categories of documents and media:

Documenting these categories forms an essential part of drafting internal AI policy.

2. Establish an audit trail and provenance record

When an organization wants to claim copyright on a work created partly with the help of AI, the human contribution must be demonstrable. This requires documenting the 'provenance' (origin):

3. Conduct periodic IP reviews of AI processes

Regularly evaluate which AI tools are being used within the teams, and check whether the vendor terms for those tools have changed in the meantime. Vendors regularly update their terms of service and data processing policies, which can directly affect the ownership and confidentiality of the data entered.

Questions to ask a lawyer

Setting up responsible AI use is a multidisciplinary task. An AI consultant or IT architect helps with technical safeguards and process design, but specific risks require the expertise of an IP lawyer or corporate counsel. Put the following concrete questions to the organization's legal advisor:

  1. Do our current customer contracts provide sufficient coverage if we deliver work products (code, advice, designs) that were partly generated with AI?
  2. Does our way of documenting and editing satisfy current case law well enough to claim copyright on the end products?
  3. Are the indemnification terms of the enterprise AI vendors legally enforceable under the applicable law of the agreement?
  4. How do our AI activities relate to the confidentiality obligations (NDAs) we have agreed with our business clients?
  5. What specific risks do we run when processing confidential customer data in self-hosted open-source models?

By answering these legal questions in advance and embedding them in the operational workflow, an organization builds an AI strategy that is not only efficient but also legally sound in the long term.

Further reading