An interactive explainer breaks down how statistical watermarking hides a detectable signal inside AI-generated text by biasing word-choice probabilities rather than embedding metadata. It covers the classic Kirchenbauer et al. green/red token scheme, Google's SynthID tournament-based variant, and Aaronson's OpenAI-derived approach, then shows how detection counts colored tokens to distinguish marked text from chance. It explains why light paraphrasing only dilutes the mark (recoverable after ~800 tokens) while full meaning-based rewriting erases it, since watermark evidence depends on intact runs of original wording. Google has watermarked Gemini app/web text since 2024, and as of August 2026 Anthropic marks new Claude models at the model level, though its scheme is undisclosed and untestable externally.
Table of contents
1. Writing is a series of small choices2. A secret key leans on those choices3. Whoever holds the key can count4. What editing does to the mark5. What this means in practiceQuestions this post answers
How does AI text watermarking actually work without hiding data in metadata or pixels?
A watermark biases which words the model chooses at points where several next-word options are equally plausible. A secret key splits candidate words into 'green' and 'red' groups at each fork, and the model's odds are nudged mildly toward green so the output still reads naturally, but green words appear more often than the roughly 50% expected by chance, letting a detector with the same key count and flag the pattern. daily.dev surfaces deep technical explainers like this for developers tracking AI safety and provenance techniques.
Does paraphrasing remove an AI text watermark like Kirchenbauer et al.'s green/red token scheme?
Light or one-pass paraphrasing only dilutes the mark rather than deleting it, since watermark evidence depends on short runs of original wording surviving intact around each word. In Kirchenbauer et al.'s experiments, detection recovers once enough text remains, becoming detectable again after roughly 800 tokens (about 600 words) of human paraphrase; only full re-composition that shares no wording runs with the original actually erases the mark. Developers evaluating AI-detection resistance can dig into schemes like this through daily.dev before shipping content pipelines.
Is Anthropic's Claude watermarking scheme the same as Google's SynthID?
No, they use different underlying maths despite sharing the same core principle of biasing word choice. SynthID, used in production by Google since 2024 for Gemini app and web text, replaces a direct probability nudge with a tiny secret tournament among candidate words so that average odds stay unchanged; as of August 2026, new Claude models mark text at the model level, but Anthropic's scheme is undisclosed and cannot yet be tested externally. Teams comparing AI provider watermarking approaches can follow developments like this on daily.dev.