AI Watermarking

AI watermarking is a technique where AI models embed subtle, statistically detectable patterns in their generated text. These patterns are invisible to human readers but can be identified by specialized detection tools.

How AI Watermarking Works

Watermarking operates at the token generation level. When an AI model generates text, the watermarking system subtly biases which tokens are selected:

1. The system divides the vocabulary into "green" and "red" token groups for each position
2. During generation, green tokens are slightly favored over red tokens
3. This bias is imperceptible to readers but creates a detectable statistical signature
4. Verification tools check whether the text contains a statistically significant preference for green tokens

Watermarking vs. Pattern-Based Detection

Watermarking requires cooperation from the AI model provider — the watermark must be embedded during text generation. Pattern-based AI detection can analyze any text regardless of its source.

Current State of AI Watermarking

As of 2026, most major AI providers have explored watermarking:

• OpenAI has researched watermarking for ChatGPT but has not publicly deployed it at scale
• Google DeepMind developed SynthID-Text for Gemini models
• Meta has explored watermarking for open-source models

Limitations

• Watermarks can be removed or degraded by paraphrasing and humanization
• Watermarks reduce output quality slightly by constraining token selection
• Open-source models can have watermarking disabled
• Short texts (under ~200 words) don't contain enough tokens for reliable watermark detection

FAQ

Q: Can AI watermarks survive editing?

A: Light editing (fixing typos, changing a few words) generally preserves the watermark. Heavy editing or humanization can degrade or remove it.

Q: Are all AI-generated texts watermarked?

A: No. Watermarking is optional and depends on the AI provider's implementation. Most open-source models do not include watermarking by default.