What happened
On 11 August, Anthropic announced that all future Claude models will generate text with a watermark identifying it as AI-generated. Google already uses a similar watermark on Gemini outputs, and OpenAI plans to introduce one.
The spread of watermarking is partly a response to the EU AI Act, which requires watermarks for AI models released after 2 August 2026, and other regulations aimed at curbing deceptive AI content.
Text watermarking works by subtly altering word selection in AI responses, making it detectable only with a specific key, as described in a 2023 paper by John Kirchenbauer and colleagues.
Why it matters
Watermarking could help identify AI-generated text, but it may change the output quality, raising concerns about utility for users who want the best results.
The debate centers on whether the slight changes from watermarking meaningfully degrade the user experience, with experts disagreeing on the impact.
As regulations like the EU AI Act push for watermarking, the technology's effectiveness and trade-offs will become increasingly important.
Key facts
Anthropic announced on 11 August that all future Claude models will include text watermarks.
Google's watermark for Gemini models is the basis for Anthropic's approach.
OpenAI has not yet introduced a text watermark but plans to do so.
The EU AI Act mandates watermarks for AI models released after 2 August 2026.
Text watermarks are not metadata but subtle changes in word selection, detectable only with a specific key.
What to watch next
Whether OpenAI will implement a text watermark and how it might affect its models' performance.
If text watermarking can achieve high detection rates without compromising text quality, as seen with image and video watermarks.
How the EU AI Act's requirements will shape the development and adoption of watermarking technologies.
