AI Act starts requiring transparency from major AI models

On August 2, 2025, AI Act obligations began applying to new general-purpose models placed on the European Union market. The original source supports the documentary core of the event; the specific duty depends on provider role, date and model classification.

It is the first time a market the size of the EU has imposed binding rules of this scope on providers of generative AI. The measure affects companies such as OpenAI, Google, Anthropic, Meta and Mistral, but also any company that develops and markets a model that third parties can reuse.

Which models fall under the new rules

The Artificial Intelligence Act, known as the AI Act, defines general-purpose AI models as models that can perform many different tasks and be integrated into other products. A language model capable of writing, summarizing or programming is the most familiar example, although the category is not limited to text.

The European Commission’s guidelines establish a practical presumption: models trained with more than 10^23 FLOP — a measure of computing capacity — and capable of generating language are considered general-purpose models. This is not a quality benchmark or a parameter count; it is intended to identify models with enough capacity to support a wide range of downstream uses. Document supporting the figure.

The obligation falls on the entity that develops or places the model on the market, not on every company that uses an AI tool. A company that integrates a provider’s API into its service will have other responsibilities under the AI Act, depending on the circumstances, but it will not have to disclose how someone else’s model was trained.

Documentation, training data and copyright

Providers will have to supply technical documentation to companies that incorporate their models into applications. The aim is to ensure that those companies understand the capabilities, limitations and terms of use of the system they are integrating.

They must also draw up and maintain a policy for complying with European copyright law. This includes respecting rights reservations that authors, publishers and other rights holders can assert against text and data mining—the legal mechanism at the center of the debate over training generative models.

The third visible obligation will be to publish a sufficiently detailed summary of the content used to train the model. The Commission has released a common template for this purpose. It does not require providers to disclose every file, link or dataset, or to reveal trade secrets, but they must describe the main categories of sources and content used. For researchers, creators and rights holders, that information could make a process that has usually remained opaque more auditable.

Additional requirements for models posing systemic risks

The AI Act establishes a stricter regime for the most powerful models—those that may pose systemic risks. The law presumes that models exceed this threshold once their training involves 10^25 FLOP of computation, although the Commission can designate models based on their capabilities or impact. Document supporting the figure.

Their providers must notify the Commission, assess and mitigate systemic risks, conduct adversarial testing—deliberate attempts to find flaws and dangerous uses—and strengthen cybersecurity. These obligations aim to reduce risks such as enabling cyberattacks, creating harmful content at scale or losing control over especially advanced capabilities.

The Commission and the Member States have also recognized the General-Purpose AI Code of Practice as a voluntary way to demonstrate compliance. Drawn up by independent experts, the code sets out measures on transparency, copyright and safety. Signing up does not eliminate the legal obligations, but it gives companies a common framework and greater legal certainty.

Two timelines for providers

The rules apply from today to models now entering the European market. General-purpose models already on the market before August 2, 2025, have a transitional period and must comply by August 2, 2027.

The date marks a significant shift in global AI competition. The European Union is not deciding how every model in the world is trained, but its market is too important for major providers to ignore. As happened with data protection after the GDPR, European requirements could ultimately influence the documentation and policies companies adopt outside Europe as well.

The system’s effectiveness will now depend on two less visible questions: whether the European AI Office can oversee complex models and whether training-data summaries provide useful information rather than vague formulas. Transparency is no longer a voluntary promise from AI labs; it is becoming one of the conditions for selling generative AI in the EU.

Turning the headline into a check

Policy reading begins with authority. Separate statute, regulation, guidance, plan, contract and a platform's private setting. Then identify who is bound, from which date and before which body. the specific duty depends on provider role, date and model classification. Without that classification, a policy priority can be misreported as an enforceable right or prohibition.

The text also needs an implementation chain: published rule, technical specification, budget, owner and evidence of compliance. The existence of a duty does not show that every tool can perform it or that enforcement detects every breach. To assess how to identify the regulated party, timeline, required document and exception before claiming compliance, look for the points where the chain can break.

Exceptions and transitions are part of the rule. Older products, smaller actors, research, security or market dates may receive different treatment. A practical check preserves the applicable article or section and explains why the case belongs there. A list of duties without subject or timeline can prompt the wrong action.

What the record must preserve

The final indicator is observable conduct: a published document, enabled option, submitted report, awarded contract or appealable sanction. Intermediate announcements are dated, but their effect is not brought forward. This lets the reader follow change without confusing intent, technical capacity and effective compliance.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.

The skill that outlasts the announcement

A valid comparison preserves denominator and axis. It does not pit a point figure against an average, future capacity against installed capacity or a forecast against an observation. When two sources use similar language, reconstruct what they counted and over what period. If those differ, publish them as different measures instead of inventing a ranking.

The record should survive a version change. Keep URL, consultation date, document, configuration and decision. When new evidence appears, add it with its date and explain what it changes. That traceability prevents opposite errors: keeping an expired conclusion or pretending later information was known on the event date.

The transferable skill in this story is how to identify the regulated party, timeline, required document and exception before claiming compliance. The procedure is short: name the document, preserve the date, fix the axis, find the condition and design a check that can fail. With those steps, a reader need not accept or reject the announcement by intuition; the decision follows a visible chain of evidence.

Before closing, another person should be able to reconstruct the conclusion without knowing the headline. Give them the sources, conditions and negative case, then ask what they would accept and reject. If they need an assumed intent, a figure without a denominator or an undated later fact, the chain still has a gap. That short review catches errors that fluent prose can conceal.