What is the TDM exception and opt-out?
The EU’s text and data mining (TDM) exceptions sit in Articles 3 and 4 of the Digital Single Market Directive and allow the reproduction of lawfully accessed works for mining purposes. Article 3 covers research organisations broadly; Article 4 covers everyone else — including commercial AI training — unless the rights-holder has reserved its rights. That reservation is the opt-out: for content made available online it must be expressed in machine-readable form, which in practice means robots.txt-style signals, metadata or the terms of the site.
How the two articles divide the field
- Article 3 — scientific research by research organisations and cultural heritage institutions; rights-holders cannot contract out of it;
- Article 4 — everyone else, on condition of lawful access and no reservation; commercial model training on scraped web content relies on this provision, which makes opt-out compliance the legal boundary of data scraping for training;
- “It was publicly available” is not the test — lawful access plus no reservation is.
The legal dimension
The AI Act hardens this from private-law risk into regulatory duty: Article 53 requires GPAI providers to maintain a copyright policy honouring TDM reservations, so compliance becomes documented and auditable rather than a matter of litigation odds. The same provision drives the training content summary, and misalignment between the two documents is an obvious audit target. Providers of general-purpose AI models should treat opt-out handling as an engineering requirement in the crawler, not a paragraph in the policy.
The Türkiye contrast
FSEK (Law No. 5846) contains no TDM exception: training on protected works in Türkiye rests on licensing or narrower grounds such as quotation, none of which was designed for corpus-scale reproduction. That asymmetry is why we advise running EU-grade opt-out hygiene globally — a pipeline lawful under Article 4 in the EU can still infringe copyright under Turkish law where the works are protected and unlicensed.
Do: log for every source whether access was lawful and whether a reservation was present at crawl time. Don’t: treat public availability as permission, and don’t assume the EU exception travels — it stops at the EU’s borders.
Related guides: Who Owns AI Output? Training Data and Copyright Under FSE….