# Gifi

Claude’s watermark is not hidden in your text. It is a trail of the words Claude chose.

Future Claude models will mark their output with SynthID-Text, to comply with the EU AI Act. Nothing is added. Light edits usually keep the pattern. A rewrite does not.

Paste the text, or drop an image or document. Continuing needs an account and a plan. [See how it works](https://gifi.ai/how-ai-text-watermarks-work).

## What a rewrite can change

SynthID-Text derives each random number from a secret key and the words so far. Choosing different words breaks that trail. We cannot see Anthropic’s key, so we cannot certify their detector.

- Light edits usually keep the signal. Replacing the wording does not — though the prose will change.
- Short text, code, and tightly factual passages have fewer forks, so the mark is weaker there to begin with.
- A hit only means Claude was likely involved. It does not prove human authorship, and it says nothing about another model.
- Images and documents are a different mark: Claude attaches a C2PA credential in metadata. Drop the file in the tool above to see what is riding along, then continue to strip it.

[See how file marks work](https://gifi.ai/how-file-watermarks-work).

Anthropic published the method on 14 August 2026. Other major providers who signed the same Code of Practice will ship their own versions. [Read Anthropic’s explanation](https://www.anthropic.com/news/claude-text-watermark).

## A plan covers the rewrite

Disturbing a statistical watermark means calling a model. One credit is one rewrite candidate, or one file cleaned of C2PA and other metadata.

See [pricing](https://gifi.ai/pricing.md).

## What this cannot promise

A watermark check is evidence that Claude was likely involved. Rewriting changes the words. Presenting either as proof of human authorship would be dishonest.

- Rewriting the prose does not mean the content was never AI-assisted.
- No tool can certify that rewritten text will fail Anthropic’s detector. SynthID-Text uses a private key; we cannot see it, and anyone claiming otherwise is guessing.
- C2PA soft binding survives a metadata strip — it can re-link to a remote manifest.
- Pixel-embedded image watermarks are out of scope. Audio and video are not supported at all.

Use it on content you own or are authorised to process, and keep any disclosure the law or a platform requires. [Terms](https://gifi.ai/terms.md).

## Start

- [Paste text or drop a file](https://gifi.ai/)
- [How Claude’s watermark works](https://gifi.ai/how-ai-text-watermarks-work.md)
- [Pricing](https://gifi.ai/pricing.md)
- [API reference](https://gifi.ai/docs.md)