AI detectors vs watermarks: the difference
An AI detector guesses. A watermark detector checks. Detectors like the ones schools use are statistical classifiers that read style and estimate how likely a human wrote it — they have no access to any watermark and produce false positives on ordinary writing. A watermark detector looks for a specific signal the model deliberately embedded. For Claude text a real watermark detector now exists, but Anthropic keeps it in private preview for organisations obliged to check under EU law[1], so anything you can put your text into today is still the guessing kind.
How do AI detectors and watermark detectors differ?
| AI detector (classifier) | Watermark detector | |
|---|---|---|
| What it reads | Style, perplexity, sentence patterns | A signal placed there at generation time |
| Needs cooperation from the AI vendor | No | Yes |
| Works on any model | Claims to | Only the scheme it was built for |
| False positives | Common, especially on non-native English | Tunable to a stated rate[4] |
| Available for Claude text today | Yes, and unreliable | Private preview[1] |
| Available for Gemini text today | Yes, and unreliable | Open library[3][4] |
Why does the distinction matter to you?
If you are accused of using AI, the question "what detected it?" has a very different weight depending on the answer. A classifier's output is an estimate about style, and non-native English speakers are disproportionately flagged by that kind of model. A watermark hit is a real signal — but even then, Anthropic says a mark means the text may have been processed by Claude, which includes proofreading and translation, and is not conclusive as to authorship[1].
Anthropic draws the same line itself, naming AI detection software such as Pangram: those services work differently because they do not hold its key, so they read the stylistic tells that tend to show up in AI phrasing instead. It gives its own examples — the "this isn't [X], it's [Y]" construction, and heavy use of the word "quietly"[10]. Picking up on habits of phrasing is a fundamentally different operation from checking for a signal placed there deliberately.
Google's published SynthID work is instructive on how a real watermark detector behaves: it scores a passage for how likely it is to carry the watermark, and leaves the acceptance threshold to you, to be set against the false-positive rate you are willing to accept[4]. The error rate is a dial someone has to choose, and a detector that is honest about itself says where it was set. A tool that hands you a confident percentage and no threshold behind it is not doing that.
See also what to do if you have been flagged.
Sources
- How Claude marks AI-generated content primary
- SynthID primary
- google-deepmind/synthid-text primary
- How Claude's text watermark works primary
Last verified against primary sources: