# How AI text watermarks and provenance differ

Hidden Unicode artifacts, statistical generation-time signals, and signed provenance are different mechanisms with different evidence and limits.

Canonical: https://watermarkslayer.com/research/how-ai-text-watermarks-work
Markdown alternate: https://watermarkslayer.com/research/how-ai-text-watermarks-work.md
Last reviewed: 2026-08-12

## Unicode artifacts

Format and control characters are literal code points in the text. They can be inspected deterministically, but their presence is never inherently malicious and does not identify an author or generator.

## Generation-time statistical watermarking

Google describes SynthID Text as changing token probabilities during generation and later looking for a statistical signal. That is not a Unicode-character scheme, and it is not a universal authorship verdict.

## Signed provenance

C2PA defines provenance structures, including wrapping for unstructured text. Recognizing candidate framing is not the same as validating signatures or trust chains; launch inspection reports structural candidates as detected but unverified.

## Treat classifiers as limited evidence

Text classifiers can be unreliable and context-dependent. OpenAI retired its earlier classifier because of its low rate of accuracy. Watermark Slayer does not provide a detector score or classify a draft's author.

## Evidence

- [Google DeepMind on SynthID Text](https://deepmind.google/blog/watermarking-ai-generated-text-and-video-with-synthid/)
- [C2PA Content Credentials 2.4 specification](https://spec.c2pa.org/specifications/specifications/2.4/specs/ContentCredentials.html)
- [OpenAI on the retired AI text classifier](https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/)
- [Unicode 17.0 Core Specification, Chapter 23](https://www.unicode.org/versions/Unicode17.0.0/core-spec/chapter-23/)