OpenAI, Anthropic, Google Sign EU AI Content Transparency Code

In August 2026, a large group of companies, including OpenAI, Anthropic, Google, Meta, and Microsoft, signed the EU's AI-Generated Content Transparency Code of Practice, committing to advance the labeling and detection of AI-generated content. This code comes against the backdrop of the EU Act's content transparency requirements taking effect on August 2: AI-generated or manipulated text, images, audio, and video must be labeled in a machine-readable way and be detectable.

OpenAI stated that, in line with its commitments under the code, it aims to extend source signals to all modalities, including text. This means that AI content may commonly carry an 'attribution' in the future.

Anthropic is the first signatory to announce a concrete plan. According to its support documentation, starting August 2, new Claude models launched in the EU will have the 'AI watermark' enabled upon release; older models, per the EU transition period, should have watermarks added by December 2.

The watermark is embedded directly into the generated text, invisible to the human eye but machine-detectable, and persists after copy-pasting, and may even survive some editing. For files such as SVG, PNG, and JPG, Claude will attach digitally signed source metadata using the C2PA open standard, indicating that the file was processed by Claude and helping to determine if it has been tampered with.

This mechanism is not limited to EU users, and the watermark is added at the model level, so whether accessed via Claude, API, or third-party cloud services, generated content will carry the watermark. At the current pace, after December 2, Claude's watermark will cover globally, across all models and channels.

Text watermarking is not a new concept. As early as 2024, DeepMind publicly released SynthID Text, which has been used in Gemini and Gemini Advanced. The principle is to make very slight adjustments to the token generation probabilities during text generation, leaving specific statistical patterns in word selection, which accumulate to form a watermark over the entire text; detection works by checking whether word choices match the statistical patterns of watermarked text.

Google's public tests show that signals can still be detected when parts of the text are cut, a few words are changed, or light rewriting occurs; however, complete rewriting or translation into another language may significantly reduce detection confidence. Google also tested on nearly 20 million real Gemini responses, showing that user evaluations of text quality remained essentially unchanged after adding watermarks.

The credibility of watermarks has been questioned. Anthropic itself notes that a watermark does not necessarily mean AI-generated, and extensive rewriting or mixing in human content can weaken or even make the watermark undetectable. This implies that watermarks are not an absolutely reliable indicator of AI content.

Two voices emerged in the comments: one calling for 'cancellation,' and another arguing that the tool used doesn't matter; final quality is what counts. Some point out that emphasizing 'this part used AI' has the effect of reminding readers to verify facts, but it may also shift the burden of verification from content producers or platforms to readers and recipients, often without their awareness.