Read Page 145 of the EU Draft AI Training Rules

Xataka reported on 28 October that the Council of the EU had agreed a position that opens the door to training AI models on personal data without asking each person for permission. Read the document and the claim gets more specific, and more interesting, than the headline.

On 21 October, Coreper, the committee that gathers the member states' ambassadors, signed off the Council's negotiating stance on the Digital Omnibus. The published text runs 179 pages and touches the GDPR, the ePrivacy directive, the Data Act and the NIS2 cybersecurity directive. One stated aim is to unblock AI training: the document argues that AI has stopped being a purely technological matter and become a geopolitical one the EU cannot afford to fall behind on.

The line worth reading is on page 145.

What actually changes

Current GDPR never names AI. It does not need to. Legitimate interest is already a legal basis for processing personal data without consent, provided the controller's interest is not outweighed by the rights of the person, and that balance is assessed case by case. That case-by-case weighing is the friction point, and it is where data protection authorities have been slow and inconsistent.

The Council's text writes the AI case in explicitly, as a new article 88 bis: a company may develop and operate AI models using personal data on the basis of its own legitimate interest, or a third party's, without consent. That is not a new power so much as a named one, but naming it removes most of the argument about whether training is covered at all.

Then the trimming. On page 145 the Council deletes safeguards the Commission had put into its own draft: an unconditional right to object, an express carve-out for minors, and a carve-out for national laws that require consent. What remains is the ordinary balancing test. An objection still exists under current GDPR, and a controller still has to stop unless it can show compelling grounds. But it stops being a switch and becomes an argument.

Two smaller changes matter for anyone whose work ends up in a corpus. First, a new exception lets a company process sensitive data that was incidentally and residually exposed during development or operation, provided it did not intend to collect it. The conditions are real: avoid collecting it, delete it on detection, protect it. And the exception explicitly does not cover what a person types into a chatbot, or the deliberate use of their voice or image. Second, pseudonymised data stops counting as personal data for a recipient who genuinely cannot re-identify anyone. If they can, the GDPR still applies in full.

Thomas Byrne, Ireland's Minister of State for European Affairs and Defence, framed the package in the Council's own release, as Xataka quotes it: "These measures will reduce compliance costs and boost competitiveness and innovation."

What did not happen

It is not law. It is the Council's mandate for a negotiation with the Commission and the European Parliament, and the Parliament has not yet adopted a position at all. The pending omnibus files are meant to close before the end of the year, which is a schedule pressure rather than a vote, and there is not much of the year left.

It is also not the AI Act. That regulation travels on a separate track, and confusing the two is how a draft negotiating position turns into a viral claim that Brussels legalised scraping overnight.

So the accurate version of the story is narrower and duller: the Council has proposed to make an existing legal basis explicit for AI, and to remove three safeguards from the Commission's version of the same proposal. Whether that survives trilogue is unknown.

Why a creator should care anyway

Because the mechanism you would actually use to refuse is the one being narrowed. If your portfolio, your face or your illustration style sits in a scraped image set, your practical levers today are a platform opt-out, a licence, and a request. The statutory objection is a balancing exercise, and the Commission's draft would have made it unconditional in this specific context. The Council's text removes that. Nothing about the removal is retroactive, and nothing about it stops you objecting; it changes how much weight the objection carries when a model developer answers it.

That matters more for images than for text, because the individual work is traceable. A style, a face and a signature are identifiable in a way that a paragraph usually is not, which is exactly why licensing fights over training sets have been fought on image corpora, and why the argument about cloned performers keeps returning to consent rather than to copyright.

What to do about it this month

Keep a dated record of what you publish and where. Not because a registry will save you, but because a specific, timestamped, machine-readable trail is the difference between an assertion and evidence when you write to a developer or a platform. Put licence terms in the page metadata as well as the visible footer. Register with the opt-out mechanisms that exist for the datasets you can name, and note the date you did it. Do not treat a robots.txt line as a legal shield; it is a request, and its standing is contractual at best.

Then read the next draft yourself instead of the coverage of it. Page 145 of a 179-page PDF is public, and it is the difference between "the EU legalised training on your data" and "the Council proposed to narrow one safeguard in a proposal that may not survive the Parliament."

If you handle client work, the practical consequence is contracts rather than politics. The offline and confidential part of your stack becomes easier to justify to a client who asks where their unreleased product shots went, and an indemnity clause about training data is worth more than a promise about model behaviour. Regulation moves in three-year arcs. A statement of work moves in weeks.