# AI text watermarks vs AI detectors

Originally published at https://gifi.ai/guides/ai-watermark-vs-detector.

A text watermark is a signal the generating model put in the words, using a key the provider holds. An “AI detector” or “humanizer” score is a separate guess about how typical the prose looks. Those tools do not hold ChatGPT, Claude or Gemini keys. Gifi inspects hidden characters, rewrites a statistical trail, or strips file metadata. It is not a detector API, and it will not sell a bypass of GPTZero or Originality.

Reviewed by [Can Balkaya](https://gifi.ai/authors/can-balkaya). Published 2026-08-17; updated 2026-08-17.

## How to verify the job

1. **Ask which question you have.** “Did this lab mark the words?” is a watermark question. “Does this website think the prose looks like a model?” is a detector question. They are not the same job.
2. **Inspect what can be named.** Paste the text. Invisible Unicode is real and deletable. A detector score will not appear there, and a statistical watermark will not either.
3. **Refuse the wrong product.** If you need a lab’s detector, use that lab. If you need different words, that is a rewrite. Gifi will not certify either outcome against a private key.

## Two different objects

SynthID-Text and Claude’s published mark are generative watermarks. The model biased its next-word choice. A later reader who holds the same key can test for that bias. No key, no test. The sentence does not need hidden characters.

Classifier “AI scores” look at style: burstiness, favourite phrases, how predictable the next word seems. They can fire on human text and miss marked text. They are not reading Anthropic’s key or Google’s key.

## Why humanizers talk past the mark

A page that promises to “remove the ChatGPT watermark” and then sells a style rewrite is usually selling a detector score. ChatGPT does not watermark text today. Claude and Gemini do. Mixing those sentences is how the category stays confused.

Gifi will change the words when the concern is a statistical trail. It will not promise that GPTZero, Originality, or any other classifier will move. That would be a different, false product.

## What Gifi is, and is not

The API can inspect, clean and rewrite. Inspect reports named characters and a style heuristic. The heuristic is not a watermark detector. The rewrite is not a certificate.

If a vendor later publishes a detector, we may attach a best-effort score to a rewrite. That score will still not be a certificate. The key remains theirs.

## Questions

### Is Gifi an AI detector?

No. The paste box names hidden characters and a writing-style heuristic. It does not hold a lab key.

### Will a rewrite make GPTZero say the text is human?

We do not know, and we will not sell that outcome. A classifier is not a watermark.

### Then what is the point of a rewrite?

To choose different words, which is the only honest lever against a keyed word-choice mark. The prose will change. That is all we can say.

## Primary sources

- [Google DeepMind on SynthID](https://deepmind.google/models/synthid/)
- [Anthropic on Claude text watermarking](https://www.anthropic.com/news/claude-text-watermark)