Texts generated by Claude in models launched in the European Union from August 2 onwards will carry a hidden watermark that will allow them to be identified. Anthropic has announced this measure to comply with the EU Transparency Code, which came into effect on August 2 and requires companies to identify all content they generate. OpenAI, creator of ChatGPT, announced last week that it would also include the text when applying the new EU Code.
Read more Residents evacuated due to wildfires in Segovia return to their homes after the situation improves
The measure has the potential to disrupt the work of millions of workers who used AI to write with the certainty that no one would find out. That certainty is now about to disappear: “The watermark is part of the text, so it will travel with the text when copied and pasted and can persist even with some editing,” the company writes in its presentation of the measure. That mark is “imperceptible” because it is in the text itself, although it will not affect or be seen in the generated content.
The mark will be present even if the generated text is a translation or an edit of a fragment written by a human: “Claude may not be the original author. People often use Claude to edit, translate, summarize, or convert files. The result may carry a Claude mark even if the ideas, text, or underlying data originate from another source,” the company writes.
Anthropic says it is developing who and how will be able to find that watermark, but everything seems to be relatively simple: “We are working to enable users and other entities to detect the included marks and the provenance of the metadata.” Models launched from August 2 will already include these marks. For earlier models, the European Union gives a deadline until December 2, which Anthropic plans to meet.
Read more A court in Syria sentences dictator Bashar al-Assad to death in absentia
In some cases, the mark may not work well in various scenarios: if the text has been heavily edited, retouched, translated, or mixed within a larger fragment; if it is a very short piece; if the metadata is removed through file conversion, screenshots, or other means. Other ways to remove these marks are likely to emerge, but it will be difficult to have absolute certainty that a text has not been tampered with by Claude. If the check is available to all users, it may be possible to verify if the mark persists before sending or publishing it.
In addition to Anthropic and OpenAI, Google, Meta, Microsoft, and Mistral have also signed the EU Transparency Code. Neither xAI nor Chinese companies have adhered so far.
Read more The Government expects up to 1.5 million additional road trips due to the eclipse
