Anthropic shared its plans this week to mark text generated by Claude models in accordance with Article 50 of the EU AI Act. The reaction arrived immediately.
Writers who use Claude to edit their own work have objected to that work carrying a mark. At the same time, developers raised a different set of concerns, and both sides have spent the week arguing about where watermarking belongs.
What’s Being Said, And What The Documents Say
Forbes reported how people were objecting to their own work becoming detectable as AI-assisted. TechCrunch found similar feelings on Reddit, alongside users arguing back at the complainers. Meanwhile, Decrypt shared that open-source projects are emerging to bypass or disrupt these watermarks.
Some reports suggest that the marking system is already in use. Anthropic’s help page states that models launched in the EU from August 2 onward will support marking at launch, and that work on earlier models is in progress. The page names no model that currently carries a mark.
Forbes points out that users cannot opt out, and Anthropic’s help page doesn’t mention this option either. What the coverage mostly leaves out is that marking is uneven by design. Anthropic acknowledges that marked content may not carry a detectable mark, and the Code doesn’t require watermarking of free-form text shorter than 200 tokens.
What Article 50(2) Requires
Article 50(2) requires providers of generative AI systems to mark outputs such as audio, images, videos, and text in a machine-readable way to indicate that they are artificially generated or manipulated.
Providers have to make those marks effective, interoperable, robust, and reliable, as far as that is technically possible. What counts as possible depends on the type of content, the cost of the work, and where the technology generally stands.
The marking duty does not apply to the extent a system only assists with standard editing, or does not substantially alter the input data or its meaning. The Commission’s guidelines exclude certain outputs from this rule, including source code and short sequences of numbers, symbols, or letters.
The Code of Practice on Transparency of AI-generated Content offers a voluntary compliance framework. By the end of July, about 190 organizations had signed up. Signatories in Section 1 include Google, Meta, Microsoft, OpenAI, and Anthropic, while Section 2 includes Getty Images, Lenovo, and Lufthansa.
Article 50(2) puts the marking duty on the provider. Meanwhile, Article 50(4) imposes an additional obligation to label deepfakes and AI-generated texts published as public-interest information. The European Commission clarifies that deployers cannot rely solely on the provider’s machine-readable mark to fulfill their disclosure duties. Article 50 carries four exemptions in total, which Roger Montti covered on August 3.
Why Implementation Is Uneven
Google says SynthID marks text generated through the Gemini app and web experience. It signed the Code on July 24 and named Apple, ElevenLabs, Kakao, NVIDIA, and OpenAI as partners in watermarking that works across systems. In the same post, Google said it was concerned that adding more rules while the technology is still evolving could undermine Europe’s competitiveness goals.
OpenAI’s page on content provenance lists C2PA metadata and SynthID for supported images, and SynthID for supported audio, which it added on July 31. They intend to support additional content types over time, but currently do not list text as a supported format.
Anthropic names no model that currently carries a mark. It says marking will cover output from supported models worldwide, across its API, apps, and developer tools.
Meta and Microsoft are signatories of Section 1, though SEJ didn’t find any specific policies on text-marking in their official materials as of August 13.
What A Mark Does Not Establish
Claude might not be the original author of watermarked text because users often proofread, translate, summarize, or convert files, which can introduce a mark even if the ideas or words originate elsewhere. Additionally, the content may have changed after Claude processed it.
Anthropic lists five reasons why marked content might not show a detectable mark:
- The model predates support for marking.
- The text was heavily edited, paraphrased, translated, or integrated into other writing.
- The passage is too brief to produce a reliable signal.
- File metadata was removed through format conversion, re-saving, or screenshots.
- The surface did not support that specific marking type.
The Code applies watermarking to free-form text longer than 200 tokens. Its glossary refers to anything shorter as very short text, expecting this cutoff to decrease as methods improve.
For audio, images, video, and text in files circulated online, the Code generally requires two separate marks because no single technique meets all four requirements. Since free-form text cannot carry metadata, the Code accepts one layer of watermarking for this format, noting that watermarking in this format may be less reliable than for longer passages.
The Commission’s final Guidelines from July 20 list AI-generated translations among examples covered by the Article 50(2) exception, alongside grammar correction and spellchecking. Anthropic states that translated output can still bear a Claude mark. Therefore, a detected mark does not necessarily mean that Article 50(2) required marking that output.
Researchers Scrubbed And Spoofed The Watermarks They Tested
At ICML 2024, researchers from ETH Zurich’s SRI Lab showed that querying a watermarked model via its public API allows an attacker to infer enough about the scheme to remove or spoof the marks the paper previously considered safe. The cost was under $50, at an average success rate above 80%. A separate experiment covered existing text, in which at least 74% of good paraphrases of non-watermarked material were detected as watermarked, with an expected false-positive rate of 1 in 1,000.
At ICML 2025, another team reported nearly complete success against seven recent watermarking methods, at $0.88 per million tokens. Their paraphrasing attack targets watermark tokens without needing access to the watermarking algorithm or model.
The Code asks the companies that sign to test how well their marking holds up against deliberate attempts to copy, remove, regenerate, or alter it, and lists paraphrasing and translation among the everyday handling a mark should survive. Neither paper tested Anthropic’s implementation, which the company has not described in technical detail.
Who Can Check A Mark
The Code mandates that companies watermarking output must offer a way for people to verify it
Anthropic will assist users and external parties in identifying its marks and plans to release technical details later. Google’s SynthID page offers guidance on verifying images, videos, and audio in Gemini and says its SynthID Detector accepts image, video, and audio uploads, which are being tested with journalists and media professionals. Google has also open-sourced SynthID’s text watermarking. SEJ covered that verification reaching Search in May, where it applies to images.
OpenAI’s verification tool supports images and audio. Currently, none of these tools allow the public to verify text. OpenAI stated that their image and audio verification tools do not confirm content was not generated by OpenAI when no signals are detected.
Why This Matters For Search Professionals
Content teams ship AI-assisted copy to client sites every day. A growing share of that copy now leaves Claude carrying a machine-readable mark, which travels with the text when it’s pasted into a CMS and published.
Google signed the same Code, along with Microsoft and Meta. These watermarks exist to be read by machines, and the companies that decide which ranks are among those that can read them. Whether marked content is treated any differently when it’s crawled, indexed, or surfaced has not been stated by anyone.
Then there’s the everyday problem. Using Claude to clean up your own draft can put a mark in your own writing. A detection says AI may have processed the text, not that it wrote it. A miss says almost nothing, because older models, short passages, and heavily edited text all come back clean. But people may still see a detection hit as proof anyway.
Looking Ahead
Watermarking is happening faster than we can read it. Anthropic hasn’t shared its detector yet, while Google’s covers images, videos, and audio, and OpenAI’s covers images and audio.
This gap creates some concern. Clients, universities, and marketplaces will soon start asking for proof that content is human-made, even before we have a clear way to provide that answer. The likely scenario is a market filled with AI-detection claims based on signals that weren’t meant to answer this question in the first place.
Interoperability is the key factor here in determining how widely this idea spreads. If checking can be done easily by anyone in a single step, rather than having to query each provider separately, then marks will move from mere compliance tools to meaningful signals that platforms, publishers, and search engines can use at scale.
More Resources
- Google Ads Now Requires Disclosure Labels On AI-Generated Content
- Google Claims AI Training Is Fair Use In New Governance Paper
- Google May Be Penalizing AI-Generated Content As Thin Content
Featured Image: Cast Of Thousands/Shutterstock
