Anthropic explains how Claude’s invisible text watermarks will work

Anthropic has sparked controversy by introducing watermarks to the text produced by its Claude chatbot, as reported by Futurism. The watermark subtly alters the AI's word choices in generated text, creating a pattern detectable with a specialized tool. This move is in response to the EU's AI Act, which mandates that AI companies label content generated or edited by their systems. Major tech companies such as Google, Meta, Microsoft, OpenAI, and Synthesia have pledged to comply with the EU's guidelines.

However, the introduction of watermarks has sparked widespread concern on X, with users expressing alarm about the permanent "scarlet letter" that will now be associated with AI-assisted writing. Some worry that media companies are normalizing AI while simultaneously stigmatizing its output. Anthropic clarified that the watermark signifies that Claude has processed the text, whether it was originally authored by Claude or not.

The company emphasized that the watermark remains even after Claude proofreads, translates, summarizes, or edits content.

The decision to watermark Claude's output has led some users to cancel their subscriptions. Vladislav Rajtmajer, a freelance developer from the Czech Republic, cited the watermark as the primary reason for canceling his Claude Max subscription, expressing concerns about potential issues with authorship and contract penalties. Similarly, Richard Echols, an AI consultant in Georgia, also canceled his subscription, worried that the watermark might appear even when he wrote the underlying material himself and only used Claude for edits.

Even if an individual edits their own work, Anthropic's watermark would still be applied, according to Echols. Despite the controversy, Anthropic is not alone in implementing watermarks. Google employs its SynthID technology for watermarking AI-generated content, and OpenAI uses SynthID for images and audio. Elon Musk's social media platform, X, also adds a "Made with AI" tag to content it identifies as AI-generated or manipulated.

Additionally, tech consultant kmleon reported that Claude has occasionally taken credit for work it was only involved with, sometimes labeling the work as "100%" or "Co-Authored/Created by Claude.Ai / Claude Code." This issue has been observed since Sonnet 4.6, though it seems to have become more frequent in recent weeks. Anthropic acknowledged the problem and highlighted the possibility that watermarks could potentially help prevent AI from training on AI-generated content.

However, Forbes reported that Google has announced plans to allow users to remove visible watermarks from AI-generated images, videos, and music.