What is AI watermarking — and will ChatGPT’s new invisible watermark work outside the EU?

A statistical watermark nudges word choices so a detector can spot a pattern, but paraphrasing can weaken it and OpenAI’s default rollout is limited to the EU.

A statistical text watermark does not stamp a visible logo onto your words. Instead, the model quietly favors certain equally plausible words, creating a pattern that a detector can measure across enough text. OpenAI says ChatGPT will apply that signal by default in the EU under AI Act transparency rules; it has not announced the same default elsewhere.

The goal is to support academic-integrity reviews and misinformation forensics, not to deliver a perfect authorship test. Short text may not contain enough signal, while translation, heavy editing, or paraphrasing can erase it. If authorship matters, keep drafts and source notes, request the original chat or edit history, compare claims with cited evidence, and treat detector output as one clue—not a verdict.