Anthropic has announced that texts and images produced by its Claude artificial intelligence models will include invisible markings. The policy is partly aimed at complying with European Union rules: Claude-generated content will receive embedded, invisible watermarks that "travel with the text" when it is copied and pasted, while other AI-created files — including images — will carry digitally signed metadata.
Announcement timing and how it will work
The company published the new policy on Tuesday. Anthropic says it will provide more information later about how the watermarks can be detected, but it already noted important limitations: the textual watermark can disappear if the text is heavily edited, and file metadata can be lost through format conversion or when someone takes a screenshot.
Users cannot opt out of the new rules, and the change will apply globally to all Claude models and products, including Claude Code and Claude Cowork.
Public backlash and main criticisms
The announcement quickly drew criticism on X and Reddit. Some users called the measure "nonsense" or "extremely problematic," arguing that the watermark could make it appear that content they personally wrote or substantially edited was produced by Claude.
Many complaints came from people who say they use Claude mainly for proofreading their own writing. Programmers also raised concerns, warning that cryptographic signatures attached to code outputs could even degrade output quality.
Erick Erickson, a radio host and blogger, wrote on X that he had used Claude instead of Grammarly for proofreading because he preferred its results, and he feared the watermark would mark his own writing as if Claude had done the work.
Who owns AI-assisted work? — the debate
The announcement reignited a broader debate about who should receive credit for AI-assisted work: the machine that generated the content, or the person who issued prompts, provided context, made decisions and refined outputs. Some argue the human user deserves recognition because they supplied the context and iterative guidance; others say the watermark simply indicates that AI assisted in creating the content and is not a claim of authorship by the model.
Some commentators went further, calling it unethical to hide AI involvement and comparing the situation to a manager giving instructions to an employee and then claiming the employee's work entirely as their own.
Background: EU code and industry context
Anthropic’s move is intended to align with the EU’s practical code on transparency for AI-generated content. That code includes provisions for labeling and detecting AI-generated and manipulated content, and for tagging deepfakes and other artificially produced or altered materials.
By the end of July, nearly 200 companies had joined the code, including Anthropic, Meta, Microsoft and OpenAI.
Google has been applying watermarks to AI-generated images since 2023 and has extended the technology to text, audio and video. OpenAI also uses invisible watermarking for images. According to earlier reporting by The Wall Street Journal, OpenAI has long had the technical ability to watermark text produced by models like ChatGPT but has not implemented it. Internal debates at OpenAI have cited concerns about false positives, the possibility of the system being circumvented, and fears that watermarking might drive users to competitors — arguments that so far have prevented broad adoption.
Takeaway
Anthropic’s policy represents another step toward industry practices intended to increase transparency about AI-generated content, but practical details and limits of the approach remain unclear. The durability and detectability of the watermark, and how user concerns will be addressed, will depend on the technical disclosures the company provides in the coming weeks.
