How will you be able to tell if text was written by ChatGPT?

OpenAI plans invisible text watermarking in the EU, using statistical signals that detectors can look for—but editing may weaken them.

OpenAI plans to watermark ChatGPT text in the EU by nudging word choices and sentence patterns in a way that looks normal but leaves a statistical signal. It may help investigate schoolwork, applications, or fake reviews, but cannot prove authorship every time. Short passages carry less signal, while paraphrasing, translation, and human edits can weaken detection.

The fair approach is to use a detector as one clue alongside drafts, sources, and conversation—not as an automatic guilty verdict.

Why watermarking AI text is harder than it sounds

A picture watermark sits on top of an image, but text has nothing extra to hide a mark in. Instead, the system quietly favours certain ordinary word choices, leaving a pattern only software can score. That makes detection a matter of probability, not certainty — and heavy editing, translation, or a very short passage can leave too little pattern to judge fairly.

What to do, step by step

1. Treat any detector result as a clue, never as proof on its own.
2. Ask for drafts, notes, or sources if authorship genuinely matters.
3. Be extra cautious with short texts, where the signal is weakest.
4. Remember that editing and translation can weaken or remove the pattern.
5. Talk to the writer first — conversation often reveals authorship better.