Claude watermarks, detectors, and what they prove
Ten explainers on the mechanics: what a statistical watermark is, why character-level cleaners do not touch it, how a keyed detector differs from an AI-writing classifier, and what any of it actually establishes. Written to be accurate rather than reassuring.
Claude watermarking
-
Why Unicode Cleaners Do Not Remove Claude Watermarks
Stripping zero-width characters is real and useful work, but Anthropic states that Claude's watermark adds no hidden characters. Here is why the two operate on different layers of the text.
-
AI Detector vs Watermark Detector: Not the Same Test
An AI detector guesses from style. A watermark detector tests for a keyed signal the generator inserted. They answer different questions with very different reliability.
-
What a Claude Watermark Does and Does Not Prove
A watermark hit indicates possible Claude involvement, not authorship. A miss does not prove human writing. How to read the evidence honestly, in both directions.
-
Anthropic's Watermark Detection API: What Is Known So Far
Anthropic has announced Claude text watermarking and says a detection API is coming. What has actually been published, what has not, and what that means for third-party tools.
-
Claude Watermark Tools Compared: What Each Can Actually Verify
Unicode cleaners, AI humanizers and statistical estimators do different jobs. A capability-by-capability comparison, and the one claim none of them can currently make.
Fundamentals
-
What Is an AI Text Watermark?
An AI text watermark is a statistical bias built into a model's word choices. Here is how the green-list method works, what a detector actually measures, and why you cannot see it by reading.
-
How AI Text Detectors Actually Work
Watermark detectors, perplexity classifiers and stylometry work in completely different ways. Here is what each one measures, which are reliable, and why false positives fall hardest on non-native writers.
-
Does Paraphrasing Remove an AI Watermark?
Paraphrasing weakens statistical watermarks — the research says so plainly — but the effect depends on how much you change and how long the text is. Here is the honest answer.
-
AI Text Watermarking by Model: Who Actually Does It
Google's SynthID-Text runs in production and Anthropic has announced watermarking for Claude, with its detection API still to come. OpenAI built one and never shipped it. A provider-by-provider status check.
-
Writing With AI Without Misrepresenting Authorship
Where the real line sits between using a tool and claiming credit you have not earned, how to keep evidence of your own process, and what to do if you are wrongly accused.