What is AI watermarking?
AI watermarking is the practice of embedding signals into AI-generated content so that machines can later detect it as synthetic: imperceptible pixel or audio patterns, statistical token biases in text, or robust identifiers designed to survive compression and cropping. The watermark is aimed at software, not at the human eye; its counterpart for people is the visible label. Under Article 50 of the EU AI Act, providers of generative systems must mark outputs in a machine-readable format from 2 August 2026 (systems already on the market before that date have until 2 December 2026), and watermarking is one accepted technique alongside provenance metadata such as Content Credentials (C2PA), fingerprinting and cryptographic methods.
How the techniques differ
- Pixel and audio watermarks alter the signal itself; well-designed ones survive resizing, recompression and format changes;
- Statistical text watermarks bias token selection during generation; they remain technically fragile and degrade under paraphrasing or translation;
- Metadata approaches travel with the file rather than inside it, which makes them interoperable but strippable.
The legal dimension
The Act does not prescribe a technology. It requires marking solutions to be effective, interoperable, robust and reliable as far as technically feasible: a standard to be evidenced, not a box to be ticked. Because text watermarking is fragile, a layered approach combining watermark, metadata and a detection interface is the defensible position for a provider of generative AI. Note the separate duty downstream: deployers of deepfake tools owe a visible disclosure to the audience, and a machine-readable watermark alone does not satisfy it. See the deepfake regime for that layer.
Turkish context
Türkiye has no AI-specific statute in force; KVKK and general provisions apply. For Turkey-connected products serving EU users, however, machine-readable marking is a launch-blocking product requirement, not a policy footnote, because the AI Act reaches outputs used in the Union. Türkiye’s November 2025 criminal-law draft on AI-generated content points in the same direction domestically, though it remains at draft stage. Building watermark support into the generation pipeline early is cheaper than retrofitting it after an EU launch, and the same infrastructure also marks synthetic content for platform policies.
Do: layer watermarking with provenance metadata and keep evidence that your marking survives common transformations. Don’t: treat a watermark as a substitute for the visible deepfake label, or assume text watermarks alone will hold.
Sources. Regulation (EU) 2024/1689 (AI Act).