How do AI text watermarks affect readability?

Summary

IEEE Spectrum examines how AI text watermarks—mandated by the EU AI Act and adopted by Anthropic and Google, with OpenAI set to implement them soon—function. These systems steer the token probabilities of large language models (LLMs) through subtle adjustments. The article also questions whether this steering significantly degrades the quality of the generated text.

Why it matters

The watermarking requirement under the EU AI Act, set to take effect in August 2026, transforms this feature from a voluntary option into a regulatory reality. Consequently, every end user consuming AI-generated text will encounter watermarked outputs. The fundamental tension between watermark detection reliability and writing quality remains unresolved. The article highlights that this debate is still active even among the researchers designing these systems.

Highlights

• Anthropic’s Claude watermark is based on Google’s Gemini watermark. OpenAI has not yet offered a watermark but has it in its plans.
• Kirchenbauer and team’s 2023 red-green list method reports a 98.4% detection rate with zero false positives. However, critics like John Gruber argue that texts contain so few tokens that they can absorb such changes without detectable degradation.
• Text watermarking is not hidden metadata. It is a statistical bias in token selection. Without the detection key, it is practically invisible to humans.

Critical take

The article relies too heavily on the 98.4% detection rate from a single 2023 study. It cuts off before addressing real-world robustness. Yet, paraphrasing, translation, or simple editing can remove the watermark. The article mentions this security gap but provides no quantitative data. Framing the debate between Gruber and Kirchenbauer as a balanced dialogue exaggerates the divide. Kirchenbauer’s emphasis on 'utility' assumes the user never edits the text. In reality, this is the most common usage scenario.