Sources

Article 50(2) requires providers to mark AI-generated or manipulated output in a machine-readable form where technically feasible.

It does not require a particular method. It also excludes assistive standard editing, and cases where the output does not substantially alter the input or its meaning.

Anthropic says models launched in the EU from 2 August 2026 support marking at launch, and that the marks apply worldwide rather than only in the EU. The consolidated Act gives some existing systems until 2 December 2026.

Text detectors are not watermark detectors, but they show why a score must not become a judgement about a person.

A 2023 study found severe errors against non-native English writers in seven tested detectors. A 2026 Czech study did not find the same pattern across three detector families. The results depend on the language, population and detector.

C2PA can attach signed claims about sources, actions and history to a file. Validation can show that the signed record still matches the file.

It cannot show that the assertions inside a claim are true, and it does not identify the responsible person on its own: that needs a separate identity assertion or a controlled workflow. Its guidance recommends creating records at meaningful points in a workflow, not after every action.

The evaluation uses public-domain works published long before modern language models. Authors are separated between the calibration and evaluation sets.