Claude vs ChatGPT vs Gemini: Which AI Has Watermarks and How to Remove Them

Overview

Despite the common assumption, not every AI provider watermarks text. As of 2026, only Google Gemini watermarks text in production. Anthropic has announced watermarking for Claude's newest models, while OpenAI's ChatGPT and open-weight models like DeepSeek and Llama do not watermark text at all. The EU AI Act is driving adoption, but rollout is uneven, and knowing who actually marks text is the first step to removing it.

Claude (Anthropic)

Claude uses SynthID-Text, a technology developed by Google DeepMind. According to Anthropic's research, the watermark works by biasing token choices during generation.

• Text watermark: SynthID-Text (statistical pattern)
• Image metadata: C2PA "Made with Claude"
• Survives: Copy-paste, light editing
• Detection: API coming soon

ChatGPT (OpenAI)

OpenAI has discussed adding provenance signals to text but has not shipped a text watermark. ChatGPT output today carries no statistical text watermark. Its image generator does attach C2PA metadata to images, which is separate from a text watermark.

• Text watermark: None shipped (discussed, not deployed)
• Image metadata: C2PA on generated images
• Detection: No text detector, because there is no text watermark

Gemini (Google)

Google uses SynthID, the same technology behind Claude's watermark. Gemini also adds metadata to generated images.

• Text watermark: SynthID
• Image metadata: SynthID + C2PA
• Survives: Common transformations
• Detection: Google has detection tools

How to Remove Watermarks from All Providers

ClaudeRewrite handles Claude output. The watermark itself is broken by the rewrite; the other two steps clean separate artifacts. The same approach works for any watermarking provider:

1. Rewrite the text through a different, non-watermarking model — this removes the actual statistical watermark.
2. Strip stray invisible characters left over from a chat interface (not the watermark, but a common tell).
3. Strip C2PA and EXIF metadata from images and files.

Which Provider's Watermark is Hardest to Remove?

All major providers use similar technologies (SynthID, statistical patterns). The difficulty of removal depends on:

• Model strength: Larger models create stronger patterns
• Watermark robustness: Some survive editing better than others
• Metadata persistence: C2PA is easy to strip; text watermarks require rewriting

References

• Anthropic: Claude text watermark announcement
• Google DeepMind: SynthID
• C2PA: Content Credentials
• OpenAI: New AI Defense Tools