How Claude marks AI-generated content

Anthropic has signed the EU AI Act's Article 50(2) Code of Practice connected Transparency of AI-Generated Content, arsenic a supplier of some generative AI models and generative AI systems. This article describes really we’re readying to put those commitments into practice, really marking works, and what its limitations are. We’ll update this article and people much elaborate method guidance arsenic it becomes available.

Machine-readable marks successful Claude-generated content

As AI-generated contented becomes commonplace, greater transparency and signals astir wherever contented comes from tin springiness group useful discourse astir the accusation they consume. To support transparency and comply pinch our ineligible obligations, Anthropic is moving to see machine-readable marks successful contented that Claude generates.

What’s covered

How Claude marks content

Claude uses 2 complementary techniques to people contented generated and processed by Claude: (1) watermarks embedded successful text, and (2) signed provenance metadata attached to files.

1. Embedded watermarks successful matter

When a supported Claude exemplary generates text, it weaves an imperceptible watermark straight into the matter itself. You won’t spot it, and it doesn’t alteration the meaning, quality, aliases readability of Claude’s response.

Because the watermark is portion of the text, it will recreation pinch the matter erstwhile it’s copied and pasted elsewhere, and whitethorn persist done immoderate editing. Watermarking will beryllium applied astatine the exemplary level, which intends it will beryllium coming nary matter which Claude merchandise aliases aboveground the matter comes from.

2. Signed provenance metadata

When Claude generates a supported record type, specified arsenic a .svg, .png, aliases .jpg, it will connect signed provenance metadata. This metadata follows the Coalition for Content Provenance and Authenticity (C2PA) unfastened standard, which is utilized crossed the manufacture to grounds accusation astir contented provenance. If a signed metadata explanation is present, it signals that a record was processed by Claude and lets you observe whether the record has been tampered with.

Detecting Claude’s marks

We’re besides moving to alteration users and different 3rd parties to observe Claude’s embedded watermarks and provenance metadata. Detection checks whether a portion of matter aliases a record carries a supported Claude mark. If a supported people is found, it indicates that the contented whitethorn person been processed by Claude.

We’ll stock specifications connected discovery mechanisms successful forthcoming method documentation.

Limitations

Machine-readable marks supply important signals astir content, but it’s worthy knowing their limitations crossed each contented types.

If you build pinch Claude

If you deploy Claude successful your ain product, you should independently measure what Article 50 requires of your products and services. Consistent pinch our commitments nether the EU Code, our extremity is to support you successful gathering your ain transparency obligations, and we'll stock method guidance connected our marking and discovery attack arsenic it becomes available.