AI AI and the law: the rules that reach ordinary users
The EU AI Act and general purpose models, in plain English
What the EU AI Act asks of general purpose AI model providers, the transparency and copyright duties, and what it means for a business that only uses these models.
The short answer
- The EU AI Act puts the general purpose model duties on the companies that build and release the models, so a business that only uses them is mainly affected through its supplier and through how it uses the output.
- Every general purpose model provider owes technical documentation, information for downstream developers, a policy for respecting EU copyright law, and a public summary of the content used for training.
- Models judged to carry systemic risk take on extra duties including adversarial evaluation, risk mitigation, incident reporting and cybersecurity of the model itself.
- The Act reaches providers anywhere in the world who put a model on the EU market, and deployers outside the EU whose system output is used inside it.
- Companies using AI are expected to ensure their staff have a level of AI literacy suited to the role, which is one of the few duties that lands directly on ordinary users.
- This is general information about how the law is structured rather than legal advice about any particular product or company.
The EU AI Act puts almost all of its duties for general purpose AI models on the companies that build and release them, not on the businesses that use them. If you are a provider of a general purpose model, you owe technical documentation, information for the developers who build on your model, a policy for respecting EU copyright law, and a public summary of the content you trained on. If your model is powerful enough to be classed as carrying systemic risk, you owe several more duties on top. If you are an ordinary company that pays for a chatbot or calls an API, you are a deployer, and this part of the law mostly reaches you second hand: through what your supplier tells you, and through the separate rules that apply to how you use the output.
This is general information about how the law is structured, not legal advice about your situation. If a decision turns on it, get advice from someone who can read your contracts and your use case.
How the law is put together
The AI Act has two halves that people often blur into one. The first half regulates AI systems by how risky the use is: some practices are banned outright, a defined set of uses is treated as high risk and carries heavy duties, some uses only trigger transparency duties, and the rest is unregulated by the Act.
The second half regulates general purpose AI models as a thing in themselves, regardless of what anyone later does with them. A general purpose model is one trained on a large amount of data that shows significant generality and can perform a wide range of distinct tasks, and that can be plugged into many different downstream systems. Large language models are the obvious example, but the definition is written around capability and generality rather than around any particular architecture.
The reason for the second half is a supply chain problem. A base model has no single purpose, so rules that start at the point of use would leave the model itself unexamined. The Act asks the provider to document what they built and pass it down the chain, so the company building a hiring tool on top has something to work with.
What a model provider has to do
The baseline duties for anyone placing a general purpose model on the EU market fall into four groups.
- Technical documentation about the model. The training and testing process, evaluation results, design choices, and the compute and energy used. This is kept and made available to the authorities on request rather than published.
- Information for downstream providers. A separate package aimed at the companies integrating your model, describing its capabilities and limitations well enough that they can meet their own obligations. This is the clause that decides whether a downstream developer can actually comply.
- A copyright policy. Providers must have a policy for complying with EU copyright law, including identifying and respecting reservations of rights made by rightsholders under the text and data mining rules. In plain terms: if a rightsholder has opted out in a machine readable way, the model provider is expected to have a process that honors it.
- A public summary of training content. A sufficiently detailed summary of the content used to train the model, published using a template provided by the authorities. It is not a file list, and it will not tell a photographer whether a specific image was used, but it is the first time the industry has been made to describe its inputs in public at all. What normally goes into that pool is covered in where AI training data comes from.
Models released under a genuinely free and open license, with weights and architecture published, get relief from parts of the documentation duties. That relief is narrower than the phrase suggests, and it is worth knowing what open actually means for an AI model before relying on it. The copyright policy and the training content summary still apply, and the relief falls away if the model is classed as carrying systemic risk.
Models with systemic risk
A subset of general purpose models is treated as carrying systemic risk, meaning risk at the scale of the whole EU market rather than to one user. A model lands in this category either by exceeding a compute threshold set in the law, which is a rough proxy for capability and can be adjusted as hardware improves, or by a decision of the authorities based on its reach and capability.
Providers of those models carry extra duties: evaluating the model against standardized protocols including adversarial testing, assessing and mitigating systemic risks at EU level, tracking and reporting serious incidents to the regulator, and maintaining adequate cybersecurity for the model and its physical infrastructure. Providers must also notify the regulator when their model meets the threshold.
Codes of practice, drawn up with industry and civil society, are the intended route for showing compliance before formal harmonized standards exist. Signing one is voluntary. A provider that does not sign has to demonstrate compliance some other way, which is more work, not less.
What changes if you only use these tools
Most businesses are deployers, and the general purpose model chapter is not written for them. Four things still reach you.
| Your role | What triggers duties | Practical effect |
|---|---|---|
| Deployer of a chatbot for internal work | Staff competence, plus company policy | Train people, set rules about what goes in |
| Deployer of a public facing bot | Transparency to the person on the other end | Make clear they are talking to a machine |
| Publisher of AI generated media | Synthetic content labeling duties | Disclose deepfakes and machine generated content |
| Builder of a tool on top of a model | You may be a provider of an AI system | Your duties depend on the use case tier |
| User in an area listed as high risk | Deployer duties under the high risk rules | Human oversight, logs, informing staff |
The labeling duty is the one most likely to catch a marketing team by surprise, since it covers generated images and video as well as text. The technical machinery for carrying that disclosure with the file is explained in how content credentials label a picture.
The Act also expects organizations using AI to ensure their staff have a sufficient level of AI literacy for their role and the context, which is a genuine obligation rather than a slogan. It is satisfied by proportionate, documented training, not by a certificate. This is the point where a written AI policy your team will follow stops being good practice and starts being evidence.
The rest of your exposure comes through your supplier. Ask what documentation they provide, what they will say about training data, and whether they commit to supporting your compliance. Those are contract questions, and they belong on the list of clauses to check before you sign an AI contract.
Who it applies to and when
Reach is not limited to European companies. The Act applies to providers who place a model or system on the EU market, wherever they are established, and to deployers established in the EU. It also catches providers and deployers outside the EU where the output is used inside it. A US company with European customers is in scope for those customers, and non EU providers generally have to appoint an authorized representative inside the bloc.
Timing is staggered rather than a single switch. The prohibitions on unacceptable practices apply first, the general purpose model duties follow, and the bulk of the high risk system obligations come later, with a longer runway for AI embedded in products that already go through safety certification. Models already on the market when the rules start get a transition period before they must be brought into line. Enforcement sits with a central AI Office for general purpose models and with national authorities for systems, and the penalty ceilings are set as a percentage of worldwide annual turnover, with the highest band reserved for the banned practices.
Copyright is the one area where the Act does not settle the underlying question. It requires a policy and a summary, but whether training on protected work was lawful in the first place is decided by copyright law and the courts, which is the subject of AI and copyright, from training to output.
What to do next
Work out which hat you wear for each tool. Write a one line answer for each AI system in the business: are we a provider, a deployer, or neither, and is the use in a listed high risk area. That single list resolves most of the confusion, because the duties follow the role, not the technology.
Then do three concrete things. Ask each supplier for their documentation and their statement on EU compliance, in writing. Run proportionate AI literacy training for the people who actually use the tools and keep a record of who attended. Check every public facing use for the disclosure question: would the person on the other end know they were dealing with a machine, or looking at generated media. Those three cover most of what an ordinary business is expected to have in place, and none of them require a lawyer to start.
Common questions
Does the EU AI Act apply to my company if we are not in Europe?
It can. The Act reaches providers who place an AI model or system on the EU market wherever they are based, deployers established in the EU, and providers and deployers outside the EU whose system output is used inside it. A company with European customers or European staff using its tool should assume it is in scope for that part of its activity and check, rather than assume distance protects it.
Is ChatGPT a high risk AI system under the AI Act?
No, not by itself. A general purpose chatbot is regulated as a general purpose model plus transparency duties, not as high risk. What can be high risk is a specific use you put it to. Using a model to screen job applicants or to score creditworthiness pulls that application into the high risk category, even though the underlying model was not built for it.
What is the training data summary actually going to tell me?
Less than rightsholders hoped and more than existed before. The summary describes the main data sources and categories of content at a level detailed enough to be meaningful, following an official template. It is not an itemized list, so you will not be able to confirm whether one particular photo or article was included. It is meant to support informed complaints and research, not individual lookups.
Do we need to do anything if we just use AI to write emails?
Very little under the model rules, but two things still apply. You are expected to make sure the staff using the tool have enough understanding of it for their role, and anything sent to the outside world that is generated has to respect the transparency rules where they bite. Beyond the law, your own confidentiality rules matter more here than the AI Act does.