Who Has Signed the EU Code of Practice on Transparency: What It Means for AI Text

Who Has Signed the EU Code of Practice on Transparency?

The EU Code of Practice on Transparency brings together the biggest players in artificial intelligence. Google and Anthropic have both signed on, affirming their commitment to label AI-generated content and disclose watermarking practices. Microsoft, riding high on its partnership with OpenAI, is also a signatory. The Code is not legally binding, but it signals a new standard for responsible AI use across the continent.

OpenAI, the company behind ChatGPT, is a notable absence. As of September 2026, OpenAI has not signed the Code, citing ongoing evaluation of watermarking technology and transparency frameworks. While OpenAI's competitors are moving quickly on visible and invisible labeling, OpenAI is taking a more cautious approach. This divide is fueling debates about what real transparency should look like in AI text, and what is actually possible to verify yourself.

The Code's main thrust is simple: users should know when content is AI-generated, and the companies should be transparent about how any watermarks work. But the implementation is anything but simple. Google and Anthropic use statistical watermarks like SynthID, which don't show up as special characters or Unicode quirks. If you're looking for a hidden character, you won't find the SynthID watermark that way. For more on what invisible characters actually look like, see our guide to visible watermarks.

What Does the Code Require from AI Companies?

Signatories to the EU Code of Practice on Transparency pledge to label AI-generated content, implement watermarking where feasible, and be open about their methods. The specifics are left to each company, so the watermarking methods vary. Google's Gemini and Anthropic's Claude models use SynthID, embedding subtle signals in word choice and phrasing patterns. Microsoft commits to visible labels and technical traceability for its Copilot outputs.

No company is forced to use a specific Unicode character or visible mark. The Code encourages transparency, not a recipe. Some companies label content with visible banners or disclaimers. Others embed statistical watermarks, which are invisible to the naked eye and undetectable by scanning for special characters. If you want to see what kinds of hidden Unicode might appear in AI text, our complete character list breaks it down.

The biggest caveat? The Code does not set technical standards. A statistical watermark like SynthID is not a Unicode character, and no browser or code editor will highlight it. The Code's power lies in public pressure and shared norms, not in enforceable, checkable codepoints.

Can You Detect SynthID Text Yourself?

SynthID is a statistical watermark, not a visible or character-based signal. You cannot detect SynthID text yourself by searching for hidden codepoints or unusual Unicode. Instead, SynthID works by subtly nudging the model's word choices so that, over hundreds of sentences, a proprietary detector can estimate if the content is AI-generated. There is no marker like U+200B or U+202F to find and remove.

You might hear claims that you can spot a watermark by looking for odd spaces, invisible punctuation, or strange formatting. That's sometimes true for legacy quirks or poor copy-paste, but it will not find a SynthID watermark. The only way to detect SynthID is with Google's or Anthropic's internal tools. No open-source software or Unicode scanner can spot it. If you need to check for visible Unicode quirks instead, try our free cleaner.

If your AI-generated text contains invisible codepoints like U+200B or U+202F, those are not SynthID. They're just artifacts, sometimes from copy-paste, sometimes from the model's tokenizer, or even from the training data. You can clean these out, but you won't touch the statistical watermark. That's the fundamental difference between cleaning characters and detecting statistical signals.

Which Invisible Characters Actually Show Up in AI Text?

In real AI-generated content, invisible Unicode codepoints like U+200B (zero width space), U+202F (narrow no-break space), and U+00A0 (non-breaking space) are common. These are not watermarks but can disrupt formatting, break code, or cause search mismatches. For example, a zero width space can turn two visually identical strings into different values in Python or JavaScript.

Here's a quick comparative table of the most common invisible characters and their symptoms:

If you're seeing invisible glitches in your AI text, you're probably dealing with codepoints like these, not a statistical watermark. The free cleaner removes these with a single click. For a deeper list, check our removal guide.

How Does the Free Cleaner Help with Invisible Unicode?

Our free tool at aitextwatermarkremoval.com finds and removes invisible Unicode, exotic spaces, and lookalike letterforms from AI-generated text. When you paste your content, the cleaner scans for codepoints like U+200B, U+202F, and U+00A0, stripping them out and leaving only ordinary ASCII characters and standard punctuation. This fixes the most common copy-paste headaches, especially for developers and editors working with AI output.

The cleaner does not touch statistical watermarks like SynthID, because those aren't characters, they're patterns in word choice. If your concern is strange formatting, invisible spaces, or text that breaks in a CMS, our cleaner solves that instantly. If you're curious about what we actually remove, see the full character list.

There's no download, no account, and no data stored. Just paste, clean, and copy. For teams or automation, we also offer an API for bulk cleaning. The free web tool is perfect for one-off fixes or content audits before publishing.

What Most People Get Wrong About AI Watermarks

Most people assume every watermark is a hidden character, something you can find and delete with a regular expression. That's not true for modern AI models. Google's and Anthropic's watermarking is embedded in word choice, not in Unicode. You cannot spot or remove a SynthID watermark by searching for codepoints. The only watermarks you can address with a character-level cleaner are visible or invisible Unicode marks, spaces, dashes, and styled characters.

Another common mistake: thinking that cleaning out invisible characters will affect AI detector scores or make content more likely to pass for human. AI detectors like GPTZero and Turnitin score statistical patterns, not hidden characters. Removing U+200B or U+202F will fix formatting, but it won't change a detector's verdict. Anyone promising detection "bypass" with a Unicode cleaner is selling hype, not results.

The only way to reliably remove invisible Unicode quirks is a character-level cleaner like ours. But you cannot "beat" a statistical watermark or detector this way. That's a crucial distinction most guides miss. For those who want to see exactly what gets cleaned, try our how-to guide.

How Do Unicode Artifacts Affect Code and Publishing?

Invisible Unicode characters aren't just a cosmetic issue. In code, a stray U+200B can break equality checks, fail imports, or create hard-to-find bugs. In publishing, U+202F can ruin search indexing, disrupt hyphenation, or block line wrapping. Copying from AI output into an IDE or CMS often brings these ghosts with you, especially if the model was trained on noisy web data.

A zero width space (U+200B) in a Python variable means two visually identical strings won't match. A no-break space (U+00A0) in HTML can force a line to overflow or break mobile layouts. Even ordinary punctuation, like em dashes (U+2014), can cause markdown renderers to fail. These are not watermarks, they're artifacts, but they matter for anyone shipping code or content.

If you move AI-generated text into production, always check for invisible characters. Our free cleaner is the fastest way to catch and fix them. For a more detailed workflow, see our removal guide. Don't wait for strange bugs to show up after deployment.

What Does the Future Hold for Watermarks and Transparency?

The EU Code of Practice on Transparency set a new tone for the industry, but the technical race is just starting. Statistical watermarks like SynthID are getting more robust, but they remain out of reach for public verification or removal. Companies are experimenting with visible banners, digital signatures, and even blockchain proofs, but no single method is dominant yet.

User demand for transparency is growing. Many want a way to see, trust, or erase AI fingerprints in their documents. But so far, only invisible Unicode and visible quirks can be found and cleaned by regular users. The deeper watermarks, those that actually matter for provenance, are locked behind proprietary detection tools. The best defense for now is to keep your text clean of artifacts and follow best practices for disclosure. For the full list of what we remove, check our complete character guide.

Expect the standards to keep evolving. When more firms sign the Code or new detection methods are released, we'll update our guides and tools. For now, focus on what you can see and control: Unicode artifacts, formatting quirks, and visible labels.

FAQ: The EU Code, SynthID and Invisible Unicode