ChatGPT textGrain Watermark: What Actually Works and Fails

OpenAI has begun deploying its textGrain watermarking system for ChatGPT and Codex users, initially targeting the European Union to comply with the EU AI Act. According to reports from The Verge and The Decoder, the technology embeds an invisible statistical signature into the model's word choices, matching or exceeding approaches like Google DeepMind's SynthID.

How reliable is the textGrain detection?

Detection rates vary dramatically depending on text length and subject matter. Internal data cited by The Decoder shows that for 400-token passages about psychology, the detector spots the watermark about 95% of the time. However, that figure drops to roughly 60% for math content, where the model has fewer alternative word choices.

Furthermore, human editing severely degrades the system's effectiveness. Replacing just 10% of a text's words with synonyms reduces detection from 92% to 66%. If a user changes 25% of the words, the detection rate plunges to 17%, making the watermark trivial to evade during standard editing.

What changes for professionals using ChatGPT on Monday morning?

For API customers worldwide, the watermarking is strictly opt-in, unlike Anthropic's approach with Claude. This gives developers control over their transparency obligations. However, OpenAI explicitly notes that textGrain does not prove human authorship, measure human contribution, or determine text accuracy.

Because basic paraphrasing or replacing a quarter of the terms breaks the statistical signature, professionals cannot rely on this tool as a definitive audit trail for AI usage. Detector access is also heavily restricted; only approved researchers can apply through case-by-case evaluations, preventing public misuse or false positives from triggering faulty automated punishments.