Two questions arrive together in every AI product review: can we train on this, and who owns what comes out? Türkiye’s FSEK and EU copyright law answer them differently, and the gap between the two answers is where Turkish AI companies serving Europe live.
Training: the input question
In the EU, text-and-data-mining exceptions (DSM Directive Arts. 3–4) permit training on lawfully accessed works, but commercial TDM is subject to rights-holder opt-outs, and the AI Act’s Article 53 obliges GPAI providers to maintain a copyright policy that respects those reservations, plus publish a training-content summary. The practical consequence: “we scraped it, it was public” has stopped being an answer in Europe; opt-out compliance is now documented, auditable behaviour.
FSEK has no TDM exception. Training on protected works in Türkiye rests on narrower grounds, and the safest structures we see are licensing, use of genuinely licensed datasets, or training pipelines whose EU-compliant opt-out hygiene is applied globally. Running one permissive pipeline for Türkiye and one strict one for the EU is technically possible and commercially unwise: your enterprise customers will ask for the strict one everywhere.
Output: the ownership question
FSEK protects works bearing the hususiyet (individual character) of an author; a human concept. Purely machine-generated output with no meaningful human creative contribution is, on the prevailing view, not a protected work; it enters the world ownerless. EU jurisprudence points the same way. Three product consequences:
- You cannot promise customers “you own the output” in the property sense; draft usage rights and exclusivity contractually instead, which your terms can do regardless of copyright status.
- Human-in-the-loop is an IP strategy, not just a quality step: documented human selection, arrangement and modification is what pulls output back into protectable territory.
- Your competitors can reuse unprotected output; if generated assets matter to your moat, layer trade-dress, trademark and contract protection over them.
Infringing output: the liability question
When a model reproduces a protected work (a recognisable character, a melody, paragraphs of a novel) liability analysis runs through reproduction and adaptation rights, and neither FSEK nor EU law cares that the copying was stochastic. Product mitigations that also read well in court: similarity filters on output, indemnity-backed enterprise models from your upstream provider (check your vendor contract clauses), and takedown responsiveness.
Update: the first major precedent, from Munich
In late July 2026 the Munich Regional Court held, in a case brought by the collecting society GEMA, that the AI music platform Suno trained on protected works without authorisation, ordering the company to disclose its revenues and compensate the resulting loss. It is one of the first major European rulings to move the training-data debate from theory to damages, and it shows the risk described in this piece is now priceable: a model developer unable to document the provenance of its training set can face revenue disclosure and compensation claims.
The contract layer nobody drafts until it hurts
| Relationship | Clause that matters |
|---|---|
| You ↔ model provider | Copyright policy warranty, output indemnity scope, training-use restrictions on your data |
| You ↔ customer | Output usage rights, no-ownership disclosure, IP-infringement responsibility split |
| You ↔ employees/freelancers | Whether AI-assisted work product is assignable “work” at all; update your IP assignment language for the AI era |
Our IP practice increasingly spends its time exactly on that third row. If your team ships AI-assisted code, designs or content daily, your 2019-vintage assignment clauses were not written for it.
This article is for general information only and does not constitute legal advice. It reflects the position as of July 2026.
Author
If this is on your desk
Templates and checklists are free in the Founder Academy; for a specific situation, book a 30-minute intro call.
Founder AcademyBook an intro call