Claude's new Scarlet Letter watermark is invisible-for now

Key Points:

• Anthropic will soon watermark all content processed by its AI models, not just newly generated content, to comply with the EU AI Act requiring watermarking of AI-generated or manipulated outputs across text, audio, image, and video.
• The watermarking applies globally from day one for new models and involves invisible embedded watermarks for text and digitally signed provenance metadata for other media, even marking minor edits despite EU exemptions for assistive or minimal changes.
• These watermarks are easy to remove or bypass, may degrade text quality slightly, and can persist through some editing, but they only signal possible AI processing without conclusively proving AI generation, risking user misunderstanding.
• Anthropic acknowledges it may over-label content, marking outputs even when Claude only performed light editing like proofreading or translation, potentially confusing users and complicating provenance interpretation.
• The EU AI Act’s transparency rules require labeling AI content under certain public interest conditions but allow exceptions if human editorial review occurs, creating inconsistencies between watermarking at the model level and public disclosure requirements; Anthropic is still developing detection tools amid potential heavy fines for noncompliance.