EU AI Act requires AI makers to disclose training data summaries

In short: The European Union is setting new rules that make many AI developers explain what they trained their systems on and follow copyright rules, including creator opt-outs.

What happened

The European Union’s AI Act adds new requirements for companies that build general-purpose AI models, meaning AI systems that can be used for many tasks, like writing text or generating images.

Under the law, these providers must publish a “sufficiently detailed” summary of the content used for training. Training data is the large collection of text, images, and other material an AI system studies to learn patterns (like a student reading lots of books).

They also must put policies in place to comply with EU copyright law. That includes the EU’s text-and-data-mining exception, which can allow copying works for analysis in certain cases, and any valid opt-outs by rights holders. An opt-out is when a creator clearly marks, in a machine-readable way, that their work should not be used for commercial AI training (similar to a “do not use” label that computers can read).

In practice, if a creator has reserved their rights in the approved way, a company may need a license to use that material, or it must avoid using it.

Why it matters

These rules do not ban AI training on copyrighted material outright. Instead, they aim to force more transparency and set clearer boundaries, so creators have more control and the public can better understand what kinds of material helped shape an AI system. The topic remains disputed in other places, including the United States, where government guidance has leaned on existing copyright rules and licensing rather than broad new limits.

Source: NYTimes