In Part 1, we covered the EU AI Act's four transparency obligations and what they mean for businesses that publish AI-assisted content. One of those obligations, the machine-readable marking requirement in Article 50(2), requires AI providers to mark their outputs so the content can be identified as artificially generated.
On August 11, 2026, nine days after the law took effect, Anthropic became the first major AI provider to publicly commit to watermarking everything Claude generates. The announcement sounds like a transparency win. Here is what is actually happening, why Anthropic moved first, and why the detection gap matters for your business.
Why Anthropic moved first
The simplest explanation: legal compliance, not industry leadership.
Article 50(2) gives AI providers until December 2, 2026 to implement machine-readable marking if their systems were already on the market before August 2. That gives Anthropic roughly four months from their August 11 announcement to the deadline. OpenAI and Google face the same clock. They will follow.
Anthropic's watermark uses two mechanisms. The first is a statistical text watermark, an invisible pattern embedded in word choice and sentence structure that survives copy and paste. The second is C2PA metadata attached to files, the provenance standard backed by Adobe, Microsoft, and Sony.
Both are real technical achievements. Both satisfy the legal requirement. But there is a gap between marking content and being able to detect it.
Nobody outside Anthropic can read the watermark
The text watermark relies on a statistical pattern embedded in word choice, not invisible characters. The pattern is designed so that only Anthropic can currently verify it. No public detection tool exists today. Third-party AI detection services like GPTZero, Originality.ai, and Copyleaks cannot see it. Anthropic says it is "working to enable users and other third parties" to detect the watermark, with details coming in "forthcoming technical documentation." No timeline has been given.
That means a small business owner who suspects an AI search engine is feeding customers hallucinated information about their company has no way to check whether that text came from Claude. Not today.
This gap matters more than the watermark itself. A marking system only works if people can actually check for the mark.
Anthropic has committed to making detection available but has given no timeline. When it does arrive, detection will answer one question: was Claude involved in generating this text? It will not tell you where the text appeared, which AI search engine surfaced it, or how many customers saw it. And it will only work on Claude output, not ChatGPT or Gemini, which have not yet implemented their own watermarking.
Maria's problem
Maria owns a bakery in Tucson. Last week she searched for her business in three AI search engines and found the same wrong information repeated across all of them. Her shop closes at 9 PM when it actually closes at 7. She supposedly offers gluten-free custom cakes, which she never has.
If watermarking worked as advertised, Maria could run the AI-generated description through a detection tool, confirm it was AI-generated text, and flag the specific engine that produced it. She cannot do that today. She has no tool. Neither does any other small business owner.
Maria can still search for her business in every AI engine manually, screenshot the results, and trace wrong information back to its likely source. A Yelp listing with outdated hours. A blog post that confused her with a different bakery. But she cannot prove the text was AI-generated. The watermark exists in theory but is invisible in practice.
Some wrong descriptions have no fixable source at all. An AI model invented details from patterns it learned across thousands of similar businesses. There is no single listing to correct. The model just made it up.
The chain problem
If AI-generated content gets published on the web, other AI engines can crawl it. A blog post that quotes a Claude-generated summary. An article that republishes text from a ChatGPT answer. That content becomes part of the web that the next AI model crawls and learns from.
This creates a traceability problem for watermarking. Each AI provider would use its own watermark. If text passes through two different AI systems, it could carry watermarks from both, or from neither if it was paraphrased along the way.
Even Claude's text watermark has limits. Anthropic acknowledged it survives light editing but does not survive paraphrasing. A paraphrased rewrite of Claude's output strips the watermark. Research from independent AI safety groups has shown that watermark removal through paraphrasing costs roughly four cents per query using a separate model. A determined actor can erase the mark for pocket change.
What this means for your business
The watermark announcement does not change what you need to do today. The detection gap means you still have to check manually. Here is what that looks like in practice.
**Search for your business in every AI engine.** Perplexity, ChatGPT with search, Bing Copilot, Google AI Overviews, and any others you can find. Screenshot what each one says. Do this monthly. AI search engines pull from the live web, and web content changes constantly. A listing that was corrected last month may have been overwritten by an older cached version, or a new source may have introduced an error.
**Check your source listings.** Most wrong AI answers trace back to outdated listings on Yelp, Google Business Profile, TripAdvisor, or industry directories. Fix those first. When you correct a listing, the next time an AI search engine crawls it, the corrected information becomes available to answer questions about your business.
**Look for fabricated details.** If an AI engine describes a product or service you do not offer, that may be a model hallucination with no external source. Document it. There is no fix button for hallucinations yet, but tracking them builds a record you can use if the issue escalates.
**Run a systematic AI visibility audit.** Tools like Parceit crawl real AI search results for your brand and report what each engine is actually saying. This replaces manual searching with ongoing monitoring across engines and over time, so you catch problems before customers mention them.
Where this goes from here
Anthropic's watermarking is a genuine step toward AI transparency. The EU AI Act will push OpenAI and Google to follow. The C2PA file metadata standard is gaining traction across major technology companies. These are real developments that will mature over the next several months.
Transparency and detection are different problems. Anthropic has committed to the first. The second remains unsolved. Until small business owners can actually detect watermarked AI text in the wild, the watermark is a compliance checkbox for AI providers, not a tool for the businesses dealing with the consequences.
For now, the best defense against AI-generated misinformation about your business is the same as it was before August 11: search, screenshot, fix what you can, and track what you cannot.
*Read [Part 1: The EU AI Act's August 2 Deadline Just Passed](https://parceit.com/blog/eu-ai-act-article-50-business-content-compliance) for the full breakdown of the legislation, the human-review exception, and the compliance checklist.*