ChatGPT Will Now Watermark Your Words

OpenAI is rolling out invisible text watermarks for eligible ChatGPT and Codex output in the EU.

The default ChatGPT rollout is EU-only, which makes Anthropic's earlier claim that it needed a global rollout look very odd. OpenAI has just shown regional deployment is doable.

A couple of months ago, people told me they were leaving Claude after Anthropic announced its own text watermark. I said ChatGPT would follow before the EU rules kicked in.

And here we are!

OpenAI calls its version TextGrain. Just like Google’s SynthID and Anthropic’s system it adds a statistical signal through the words the model chooses.

Copy the text into Word, paste it as plain text or retype the whole thing and the wording stays the same. So does the signal. Doesn’t matter - the signal IS the text.

Your document can be yours and still flag

Importantly the ChatGPT watermark applies as soon as you generate significant text with it.

Imagine you spend a week writing a report. The argument, examples and experience are yours. Then you ask ChatGPT to tidy it up.

If there’s significant reworking then it’ll flag as AI.

A positive result is evidence that an OpenAI system may have generated or processed part of the passage. False positives are (very!) possible. Even if you put a lot of “human” work into it.

However this does not mean OpenAI own the document.

This is the same misunderstanding I saw after the Claude rollout. People assumed a mark proved cheating, transferred ownership or identified the user. Those claims were nonsense then. They're still nonsense now. It’s nothing so nefarious.

OpenAI's own numbers are messy

Like with SynthID and Anthropic’s own system rewriting dilutes the watermark.

This time we actually have an idea of how aggressively the dilution happens.

In OpenAI's test of 400-token English passages, detection fell from about 92% to 66% after 10% of the words were replaced with synonyms. Replace 25% and it fell to 17%. Pretty bad!

BUT that’s a 400 token passage - around 350 words or so - so there’s not much for the watermark to “glom” onto! Longer passages will be more immune to dilution via rewrites.

Also: this is the justification OpenAI have given to justify limiting detector access to approved researchers and expert organisations. Same as Anthropic they are keeping the detector itself very limited.

To be very clear: your boss, client or university tutor cannot open a public OpenAI checker today.

So when a tool gives you a neat score saying your document is "80% AI", ask which detector it used. It is VERY unlikely to be the official detectors therefore guess what: they don’t know squat! As before, the people selling guaranteed watermark removers are shifting snake oil.

Keep the trail

Keep using AI, but save your original notes and pre-AI drafts when authorship matters. Turn on version history. Track the edits. Agree what a client allows before you send the work.

FYI I've updated my free AI Writing Toolkit with a guide, checklist, nine prompts and an AI Tells skill for ChatGPT and Claude. It helps you keep your voice, review suggested edits and save a record of what changed. Nobody can promise it will beat a watermark detector. I certainly won't! You know me better than that ha!

To the Task,

Kyle