Anthropic’s SynthID Watermark Is Catching Its Own Users Off Guard

Anthropic has deployed an invisible watermark across its Claude platforms using a version of Google DeepMind's SynthID-Text technology, a move tied to compliance with the EU AI Act and other emerging transparency rules. The problem: the watermark cannot tell the difference between text Claude generated on its own and text a human wrote that Claude then heavily edited. That means professional writers and coders paying for Pro or Enterprise subscriptions may unknowingly be sending watermarked work to clients, even when the ideas and most of the words are their own.

What happened

Detail	Fact
Technology used	Google DeepMind’s SynthID-Text (invisible watermarking)
Who deployed it	Anthropic, across Claude platforms
Compliance driver	EU AI Act and other transparency requirements
When SynthID was first rolled out (Google)	2023 and 2024
LinkedIn AI-slop flags in first few weeks	More than 1 million
Visibility drop for suspected AI slop on LinkedIn	About 40%

SynthID was originally built by Google DeepMind to embed imperceptible markers into AI-generated content. For images and video, those markers live in pixel data. For audio, they sit in frequency patterns. For text, they are woven into the statistical patterns of token generation, the process by which a language model picks each word. Google began applying it to Gemini outputs in 2023 and 2024.

Anthropic adopted the text variant for Claude and extended it across its platforms. The timing landed in the middle of a heated public conversation about AI detection tools and what it means professionally to use AI assistance.

Why the watermark is catching writers off guard

The core technical problem is a blunt one. The watermark is applied whenever Claude is substantially involved in producing or editing text. It cannot distinguish between a passage Claude wrote from scratch and a passage a human wrote that Claude then reorganised, polished, or expanded. Anthropic has acknowledged this directly.

That gap matters most to Claude’s core user base. Claude has developed a strong reputation for natural, nuanced long-form writing and complex coding assistance. Many users switched from other AI tools specifically for those strengths. For a freelance writer or a software developer submitting work to a client, a watermark on a heavily edited document raises real questions about authorship and professional trust.

Pro and enterprise subscribers paying significant monthly fees now face the possibility that work they spent hours crafting could be flagged as AI-generated by a detection tool on the receiving end, even if the ideas, structure, and substance came from them.

Is this about stopping AI slop?

Watermarks do not stop low-quality AI content on their own, but the concern about that content is real and growing. LinkedIn is a useful data point here. The platform added a “Seems like AI slop” option for users to flag posts, where AI slop refers to polished-sounding content that carries little original perspective or substance. According to the source, more than one million people used the option within the first few weeks of its launch. LinkedIn says it combines that feedback with other ranking signals, and suspected slop is now receiving about 40% less visibility than before.

Crucially, clicking the flag does not automatically penalise the author. LinkedIn feeds the signal into a broader detection and ranking system rather than applying an instant penalty.

The watermark debate and the AI slop debate are related but not the same. Not all watermarked content is slop, and not all slop carries a watermark. The deeper issue is that platforms and regulators are trying to impose accountability on AI-assisted content before anyone has agreed on what “AI-assisted” actually means at each point on a spectrum from light editing to full generation.

Our take

From where we sit, shipping a watermark that cannot separate “Claude wrote this” from “the user wrote this and Claude touched it up” is a meaningful implementation gap, not a minor footnote. Anthropic’s transparency goal is reasonable. The EU AI Act is real, and some form of AI disclosure is coming whether users like it or not. But a signal that misfires on heavily human-edited work undermines the credibility of the whole system. It creates false positives for the exact users who are engaging with AI most thoughtfully.

For businesses that use Claude as part of a content or AI integration workflow, this is worth stress-testing now rather than after a client flags a deliverable. Run your typical Claude-assisted documents through a SynthID-aware detection tool and see what comes back. If the watermark persists on work that is primarily yours, you need a documented workflow that shows your process, not just a denial.

There is also a broader signal here for anyone tracking where Anthropic and the wider AI industry are heading: transparency requirements are accelerating. Companies that build on these tools should assume more disclosure mechanisms are coming, not fewer. Plan for it now.

What to do about it

1. Test your current workflow: submit a representative Claude-assisted document to a watermark detection tool and note the result.
2. Document your process for client work, recording how much of the drafting, structure, and editing you contributed versus Claude.
3. Check whether your contracts or client agreements address AI-assisted work, and update language where needed.
4. Monitor Anthropic’s product updates for any refinements to how SynthID-Text is applied, particularly around heavy user-editing scenarios.
5. If the watermark is a hard blocker for your business, evaluate whether a different model or a self-hosted approach fits your transparency requirements better.

The bottom line: the watermark is not going away, so the practical move is to build a paper trail that shows where your work ends and Claude’s begins.

Source: Bing News · Anthropic