Theo Browne examines the European Union's transparency requirements for generative AI and Anthropic's plan to add machine-readable marks to Claude outputs. Generated text will carry embedded watermarks at the model level, while supported files can include signed provenance metadata across Claude products and cloud platforms.
Media provides ample hidden data for imperceptible signals, but that also makes those signals vulnerable. Compression, resizing, re-encoding, filtering and format conversion can destroy pixel-level patterns without materially changing what a person sees. Signed provenance metadata can prove where an intact file came from, but it can also disappear when the file is resaved or converted.
Text has less room for invisible changes. Statistical schemes can bias token selection toward a detectable pattern, while Unicode substitutions can encode signals in apparently ordinary spaces or characters. Both approaches remain vulnerable to paraphrasing, manual edits, normalization or replacement, and short passages may not contain enough information for reliable detection.
Theo Browne concludes that watermarks may catch direct copy-and-paste spam but are unlikely to stop deliberate deception. He sees more promise in signing genuinely human-created media at capture time, then educating people that unsigned content is uncertain, rather than treating the absence or presence of an AI mark as definitive authorship evidence.
Watch on YouTube