Anthropic rolls out global watermarking for Claude

Anthropic rolls out global watermarking for Claude

Tech firms face growing pressure to make AI-generated material easier to identify

Anthropic has begun adding machine-readable marks to text and images generated by its Claude AI models.

The AI company confirmed the move in an updated support document after signing the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-generated Content.

The rules, which took effect on 2 August, require providers of generative AI systems to make their output detectable as artificially generated or manipulated where technically feasible.

Anthropic said it would apply its marking system to Claude users globally, rather than limiting it to Europe.

For text, Anthropic says Claude embeds an imperceptible watermark designed not to affect the meaning, quality or readability of its responses.

The company says the mark can remain present when text is copied or lightly edited, although it acknowledges that heavily altered or reworked material may no longer contain the signal.

For images and other supported files, Anthropic uses signed provenance metadata based on the open C2PA standard, which is designed to record information about where digital content originated and whether it has been altered.

The standard is already used by a number of technology and media companies, including Adobe and the BBC, as well as camera manufacturers.

Anthropic said Claude models launched in the EU from 2 August would include the marking system at launch, while older models would be updated to support it.

The technology will cover its developer API, Claude applications, Claude Code and enterprise deployments.

The company stressed, however, that the presence of a mark does not prove conclusively that content was generated by Claude. It indicates only that the material "may have been processed by Claude".

Likewise, content without a mark should not automatically be assumed to be human-generated, particularly if it was produced using an older model or substantially edited.

Anthropic said it plans to publish more technical information on how third parties will be able to detect its marks.

The European Commission published its voluntary Code of Practice on 31 July, shortly before the transparency provisions became applicable.

Companies that sign and follow the code receive a presumption that they are complying with the relevant requirements.

Around 190 organisations have signed parts of the code, including major technology companies such as Google, Meta, Microsoft and OpenAI.

Anthropic's move comes as technology companies face growing pressure to make AI-generated material easier to identify.

OpenAI joined the C2PA coalition in May and has worked with Google on using its SynthID watermarking technology for images. Google DeepMind's system is designed to keep an invisible signal intact through changes such as resizing and compression.

Anthropic's decision also comes as online platforms face pressure from users and regulators over the spread of AI-generated material.

AI music platform Suno recently said it would mark tracks created on its service following legal challenges, while newsletter platform Substack has partnered with Pangram to identify AI-generated content.

The financial consequences of failing to comply with the EU's transparency requirements could also be significant. Companies can eventually face fines of up to €15m or 3% of their total worldwide annual turnover, whichever is higher, for certain breaches.