AI Transparency · 4-minute read
A new generation of Claude models will mark AI-generated content even after the words have been copied somewhere else. That sounds technical, but it changes the politics of authorship.
Anthropic says supported Claude models launched on or after 2 August 2026 will embed an imperceptible watermark in generated text and attach signed provenance information to supported files. The move responds to the EU AI Act’s transparency framework and has triggered debate about detection, false attribution and human–AI co-creation.
What happened
Anthropic has described two approaches to marking Claude output. Generated text can contain a statistical signal woven into word choices at the model level. Supported image files can carry signed provenance metadata that records their origin and whether the file has been altered.
The watermark is intended to survive ordinary copying and pasting. It is not a visible badge and does not identify every possible piece of AI-assisted writing. Short text, heavy rewriting, translation or multiple rounds of editing can reduce the reliability of detection.
The change follows the 2 August application of European transparency requirements for generative AI. Other companies are also investing in C2PA metadata, SynthID and verification tools, making provenance an industry-wide infrastructure problem.
Why it matters now
The first era of AI detection tried to infer authorship from how writing sounded. Those classifiers produced false positives and often confused formal or second-language writing with machine output. Provenance takes a different approach: look for a signal intentionally added by the generating system.
That can help platforms and investigators, but it also raises questions. Does a watermark prove that a model authored the final work, or only that Claude participated somewhere in the process? What happens when a person substantially edits the output? Co-creation does not fit neatly into a binary label.
What changes
- For creators: AI assistance may remain detectable beyond the original application.
- For publishers: provenance checks could become part of editorial and rights workflows.
- For platforms: “AI-generated” labels will need context, appeals and uncertainty—not automatic punishment.
The tension
People deserve to know when synthetic media is being used deceptively. Yet invisible marking can stigmatise legitimate assistance if platforms treat detection as proof of dishonesty. A machine-readable signal must not become a machine-made verdict.
The Agentica IX view: Provenance should clarify the creative process, not erase the human contribution. The useful question is not “Was AI present?” but “Who directed, shaped and stands behind the final work?”
What to watch next
Watch how Anthropic reports accuracy, how watermarks survive editing and whether detection tools expose confidence levels. The decisive issue will be how schools, employers, publishers and social platforms interpret a positive signal.
Primary reading: European Commission: AI transparency rules and TechRadar’s report on Claude marking.