AI‑Generated Content Must Be Watermarked—OpenAI, Anthropic, Google Sign EU Pact

The European Union’s AI‑Generated Content Transparency Code of Conduct, which came into force on 2 August, requires that any AI‑generated text, image, audio, or video be marked in a machine‑readable way and be detectable.

OpenAI, Anthropic, Google, Meta, Microsoft and other leading AI companies have signed the agreement, committing to extend source‑signal tagging to all modalities.

Claude’s watermark implementation

Anthropic announced that new Claude models released in the EU will automatically embed an “AI watermark” at launch. The watermark is invisible to human readers but can be detected by machines; it survives copy‑paste and many minor edits. For image outputs (SVG, PNG, JPG), Claude adds digitally signed source metadata using the open C2PA standard, indicating the file was processed by Claude and helping detect later tampering. The mechanism applies to Claude, its API, and third‑party cloud services, and will cover all models and channels globally by 2 December.

Earlier watermark technologies

DeepMind disclosed a “SynthID Text” system in 2024, used in Gemini and Gemini Advanced, which subtly adjusts token‑generation probabilities to embed a statistical pattern that serves as a watermark. Detection reverses the process, checking whether the token choices match the expected pattern.

Google’s internal tests on nearly 20 million real Gemini responses showed that the watermark does not noticeably affect user‑perceived text quality. The watermark remains detectable after partial text removal, minor word changes, or light rewriting, but detection confidence drops sharply after full rewrites or translation.

Critiques and implications

Critics argue that a watermark does not guarantee that content is AI‑generated, as the signal can be weakened by extensive editing or mixing with human‑written text. The added attribution shifts some verification responsibility to readers, raising questions about whether watermarked output truly counts as “AI‑written.”