What does the EU's code of practice for general-purpose AI mean?

The EU AI Act is the European Union's law on artificial intelligence. Most of it deals with how AI is used, for example in education or employment. One part deals with the models themselves: the large, flexible systems behind chatbots and many other tools, which the law calls general-purpose AI models. That part takes effect on August 2, 2025, and companies have been asking a practical question: what exactly do we have to do?

The code of practice is the answer the Commission has put on the table. It is not a law and not a regulation. It is a detailed, voluntary document that turns the AI Act's broad obligations for model makers into concrete steps. A company that signs it and follows it can point to it as evidence that it complies.

Who wrote it and what it contains

The Commission received the final text on July 10, 2025. Thirteen independent experts led the drafting, through a process that began in July 2024 and drew on three rounds of consultation, more than 1,600 written submissions and 40 workshops, according to the Commission. Model providers, small businesses, researchers, rightsholders and civil society groups all took part.

The code has three chapters:

Systemic risk is the AI Act's term for models with high-impact capabilities, roughly the ones at the frontier. The law presumes a model falls in this group if training it took more than 10^25 floating-point operations, a measure of total computing work.

In its coverage, TechCrunch summed up the practical commitments this way: keep model documentation current, do not train on pirated content, and respect requests from content owners to leave their works out of training data.

What signing does, and does not, do

The Commission's Q&A is explicit that the code creates no new obligations. The duties come from the AI Act itself; the code is one recognized way of meeting them. A provider can also prove compliance some other way, but the Commission says signatories will face less paperwork and more legal certainty, because enforcement will focus on checking that they keep their commitments.

Two steps remain. EU member states and the Commission still have to assess whether the code is adequate, and the Commission has published separate guidelines explaining who counts as a provider and which models are in scope. TechCrunch reported those guidelines on July 18.

Not everyone is on board. On July 18, Meta said it will not sign. Its chief global affairs officer, Joel Kaplan, argued on LinkedIn that the code brings legal uncertainty and goes beyond what the AI Act requires. Choosing not to sign does not exempt a company from the law; it only means proving compliance without the code.

What it means if you use AI tools

If you use ChatGPT, Claude, Gemini or a similar assistant, the code places no duties on you. Its effects reach you indirectly: through documentation that providers must prepare, through clearer copyright policies, and through safety work on the largest models.

If you build products on top of someone else's model, it is worth knowing where your provider stands. A short worked example of the questions to ask:

1. Is the model you use a general-purpose AI model offered in the EU? The Commission's guidelines explain how to tell.
2. Has the provider said whether it signs the code? Look for a public statement.
3. What documentation will it share with you about the model's capabilities and limits?
4. What is its copyright policy, and how can rightsholders opt out?

This is a summary of the official texts as published, not legal advice. Anyone with a specific compliance question should speak to a lawyer who works on EU law.

What changes on August 2, 2025?

From that day, any company that offers a general-purpose AI model in the EU must meet their AI Act obligations. Providers of models with systemic risk must notify the Commission's AI Office without delay.

Enforcement is staged. During the first year, the AI Office says it will work closely with providers that sign the code, and will treat those that have not yet put every commitment in place as acting in good faith rather than as violating the law. From August 2, 2026, the Commission can enforce the rules with fines.

Models that were already on the market before August 2, 2025 get more time: they have until August 2, 2027 to comply. The code itself is due for review by the AI Office no less than once every two years, so the details are expected to change as the technology does.