Tech giants must watermark AI-generated content or face $17M fines, Anthropic starts immediately
AI companies must follow EU’s rules.
Image by Getty/SOPA Images
- Anthropic has started adding machine-readable watermarks to Claude-generated content under the EU AI Act’s Article 50 rules.
- The rules apply to general-purpose AI models and can bring fines of up to €15 million ($17 million) or 3% of global annual turnover.
- Claude outputs may include invisible text watermarks and signed file metadata, but these marks can fail after edits or platform changes.
- Anthropic plans to roll out watermarks worldwide, showing growing pressure on AI companies to identify synthetic content clearly.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
Anything generated by Claude will now include an invisible, machine-readable watermark indicating that it’s AI-generated.
Anthropic, the company behind Claude, has recently signed Article 50 of the EU AI Act, which is designed to ensure AI companies act transparently.
As per Article 50, all AI companies must include machine-readable watermarks to show that they’ve been synthetically created or manipulated.
These new rules extend to general-purpose AI models like ChatGPT and Claude.
Anthropic has already started watermarking content created using Claude, but this only applies to “Claude models launched in the EU on or after August 2nd, 2026.”
Existing models will also have watermarks, but this won’t happen immediately.
The new regulations give AI companies like Anthropic room to program Claude models released before August 2nd, 2026, to watermark content.
Other AI companies like OpenAI are forced to comply with the new rules, but plan to adopt a layered approach to AI content identification.
“No single provenance or detection method is perfect, and signals can sometimes be removed or fail to survive edits or platform changes. That is why OpenAI supports a layered approach that combines standards, watermarking, and verification tools,” according to OpenAI.
If AI companies refuse to comply with Article 50, they may be fined up to $17 million (€15 million) and up to 3% of their total worldwide annual turnover for the preceding financial year.
How Anthropic watermarks AI content
Content generated by Claude, whether it be texts, PDFs, or images, will “carry embedded watermarks,” and files generated by Claude “will include digitally signed provenance metadata where supported.”
Embedded watermarks work by weaving an invisible watermark into the text. This won’t alter the text, meaning, quality, or readability of the content, says Anthropic.
As the watermark is part of the text, the label will follow it wherever it goes, meaning it will be present even when the content is copied and pasted.
“Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.”
Signed provenance metadata works by attaching a signature to files like SVG., PNG., and JPG.
“If a signed metadata label is present, it signals that a file was processed by Claude and lets you detect whether the file has been tampered with.”
Stay updated with our latest stories and follow us on social media
Be the first to discover new stories, ideas, and updates from our team.
Digital watermarks aren’t air-tight
While lecturers, teachers, and other professionals might see this as the end of unfair AI use, there are limitations to the tech, which Anthropic acknowledges.
If a watermark is detected, this doesn’t necessarily mean the entire text, image, or file is AI-generated – it may have just been processed by Claude.
This doesn’t automatically mean that a text was originally authored by Claude, as people often input their work for Claude to proofread, summarize, and translate into a different language.
If someone asks Claude to proofread a text or convert a video into a different format, that means it has been processed by Claude.
A text that has been proofread by Claude and a text that is entirely AI-generated will bear the same watermark.
By the same token, Anthropic also warns that users shouldn’t immediately accept that a text or file is human-created if it doesn’t have a watermark.
Pre-existing documents or files generated before the marking were supported may not show a Made by AI signature.
Anthropic rolls out watermarks worldwide
Given that AI-generated content has become common online and elsewhere, AI companies are increasingly under pressure to make it obvious.
“To support transparency and comply with our legal obligations, Anthropic is working to include machine-readable marks in content that Claude generates.”
While the rules apply within the EU, Anthropic will be watermarking AI-generated content for users worldwide.
The marks will apply to outputs from all supported Claude models “wherever Claude is offered worldwide.”
Cybernews has reached out to Anthropic for comment.
However, certain platforms or features might not support certain kinds of digital marking.
Similarly, if a text has been heavily edited, paraphrased, translated, or enmeshed with other texts, it's likely not to be marked.
The same goes for very short texts and content produced on a platform where marking isn’t supported.
