OpenAI Integrates Invisible textGrain Watermarks into ChatGPT

OpenAI has integrated invisible textGrain watermarks into ChatGPT to label AI-generated content. The mechanism is embedded in word generation statistics and operates in compliance with the EU AI Act.

How Watermarking Works

The signal is embedded directly into the LLM's response text by altering word selection probabilities. This allows detectors to identify characteristic patterns even after content is copied using Ctrl+C.

• Watermarks are not metadata
• Operates via token statistics alteration
• Mandatory in the EU under the EU AI Act
• Not available as a public tool

Technical Limitations and Future Prospects

The main issue is watermark resilience against text rewriting by other models. When ChatGPT output is fully rewritten using third-party LLMs, the original token sequence disappears.

This leads to degradation of the statistical watermark. There is currently insufficient independent testing on textGrain’s resistance to modern paraphrasers, so OpenAI warns that the signal does not constitute absolute proof of text origin.

An increase in services removing such marks from student papers is expected, creating new niches in the AI services market.