Anthropic Faces Backlash Over Invisible Claude AI Watermarks

Anthropic is facing growing criticism over its decision to embed invisible watermarks in text and images produced by its Claude artificial intelligence models, with users arguing that the technology could make AI assistance detectable even when they use Claude for tasks such as proofreading or coding.

The policy, announced Tuesday, is intended to bring Anthropic into compliance with European Union rules governing the identification of AI-generated and manipulated content. Under the new system, text created by Claude will contain an invisible watermark designed to remain with the writing when it is copied and pasted, while images and other files generated by the AI will contain digitally signed metadata.

Anthropic has not yet provided all the technical details needed to identify the marks, saying it plans to publish additional information. The company has also acknowledged that the system will not be foolproof. Text watermarks could be removed through extensive editing, while metadata attached to files could be lost when a file is converted into another format, screenshotted or otherwise altered.

The policy will apply worldwide to every Claude model and product, including Claude Code and Claude Cowork. Users will not have an option to disable the watermarking system.

The announcement quickly triggered criticism across X and Reddit. Some users described the policy as “bulls—” and “hugely problematic,” while others argued that it signals the end of “the era of unprovable AI writing.” A significant portion of the criticism came from users who said they rely on Claude to proofread material they originally wrote themselves. Software developers also raised concerns that adding cryptographic signatures to code could affect the quality of generated output.

Radio host and blogger Erick Erickson said he had replaced Grammarly with Claude because he found it more effective for proofreading. “But now the stuff I’ve written will be watermarked that Claude did the work. This is ridiculous,” Erickson posted on X.

The policy has also prompted debate over who deserves recognition when AI is used to create content. One Reddit user argued that human users should receive credit because they provide the instructions, context, decisions and revisions, describing Claude as the tool used to produce the final work. Other commenters disagreed, saying the watermark does not amount to Claude taking credit but instead identifies the material as AI-generated. Some users took the opposing view, arguing that failing to credit the AI would itself be unethical because the model performs the substantive work after receiving the user’s instructions.

Anthropic’s policy follows the European Union’s “Code of Practice on Transparency of AI-Generated Content,” which establishes measures for marking and identifying AI-generated or manipulated material, including deepfakes and altered text. By the end of July, nearly 200 companies had signed onto the code, including Anthropic, Meta, Microsoft and OpenAI.

Google has been watermarking AI-generated images since 2023 and has expanded the practice to text, audio and video. OpenAI also uses invisible watermarks for images.

OpenAI has technology capable of watermarking text generated by models such as ChatGPT but has not released such a system. The Wall Street Journal has reported that the company has debated deploying text watermarking internally for years, with concerns about false positives, the ability of users to circumvent the technology and the possibility that watermarking could push users toward competing AI services contributing to the decision not to release it so far.