ChatGPT Now Watermarks Its Own Text in the EU — Here's What It Means for Your Work

For years, AI companies treated text watermarking like a toxic idea — technically possible, commercially radioactive. Last week, OpenAI did it anyway.

On October 5, the company announced that text generated by ChatGPT and Codex for users in the European Union will carry an invisible watermark, rolling out over the coming weeks across all plans. The system is called textGrain. The reason isn’t product strategy. It’s law: Article 50 of the EU AI Act, whose transparency rules for AI-generated content have been in force since August 2, 2026, requires providers to mark synthetic text in a machine-readable way. OpenAI signed the EU’s Code of Practice on Transparency of AI-Generated Content alongside Anthropic, Google, Meta, and Microsoft — roughly 190 signatories in total — and this is the first concrete move any of them has made.

That’s the news. Here’s what it actually means for people who work with these tools every day.

It’s not a stamp. It’s a statistical fingerprint

There is no hidden character to find, no invisible ink, no suspicious spacing. textGrain works by nudging the model’s choice of the next word — thousands of tiny preferences across a long passage that, added together, form a pattern a detector can recognize. Because the pattern lives in the wording itself, copying text into Word or Google Docs doesn’t strip it. That’s the part most coverage glosses over.

The scheme goes beyond text. Images generated through ChatGPT, Codex, or the API carry C2PA Content Credentials plus Google’s SynthID watermark; audio carries an inaudible SynthID signal; video carries Content Credentials. API developers anywhere in the world can switch text watermarking on per project or across an organization — it’s off by default, and OpenAI says it won’t become a global default at launch. The watermark doesn’t identify who generated the text, and OpenAI says it saw no meaningful change in output quality with the watermark switched on.

Edit it lightly and the signal dies

Here is the uncomfortable part — for everyone. OpenAI’s own testing shows how fragile this is. Replace just 10% of the words with synonyms and detection drops from around 92% to 66%. Replace 25% and it collapses to 17%. Detection climbs with length: at a 1% false-positive rate, OpenAI caught the watermark in roughly 80% of 200-token responses, rising to about 95% at 400 tokens. Math answers, translated text, and code are harder to catch — EU guidance even excludes code snippets and sub-200-token responses (around 150 English words) from watermarking expectations.

So yes, anyone who lightly edits AI text defeats the system. And the detector is only available to approved researchers and expert organizations for now — you can’t check your own text even if you wanted to. OpenAI’s own warning says it best: the absence of a detected watermark does not prove human authorship.

That sentence cuts both ways. Remember it.

If you ship AI-written work to clients, this is your problem now

The mistake we see most often isn’t people getting “caught” using AI. It’s agencies and freelancers delivering raw, unedited ChatGPT drafts and calling it a deliverable. That was already a quality problem — flat, generic copy that reads like everyone else’s. Now it’s a provable one. In the EU, at least, a client or platform with detector access can machine-check whether the text you handed over was AI-written.

We don’t think the answer is hiding it better. We’ve watched freelancers obsess over “humanizer” tools built to dodge detectors, and it’s the wrong game — OpenAI’s own numbers show any detector is one light edit away from useless. The honest move is simpler: edit AI output properly until it’s genuinely yours, or disclose it upfront. One practical step is adding a short AI-use line to your freelance contract — which tools you use, what the client can expect — so the conversation happens before delivery, not after an accusation.

Store owners: the watermark changes nothing about the real argument

If you use ChatGPT to draft product descriptions, the watermark doesn’t make a bad description worse or a good one better. A thin, generic description still converts nobody, watermark or not. The real risk was never “Google will punish my AI text” — Google’s position has consistently been about content quality, not origin. The risk was always shipping copy you didn’t bother to make specific.

Still, there’s a new wrinkle worth knowing. A description you paste in unedited is now, in principle, machine-identifiable as AI output. Anyone reviewing your listings — a marketplace, a picky wholesale partner — has one more reason to ask whether the copy is actually yours. The fix is the same process we’ve always recommended: feed the tool real context, generate angles, then edit hard. Our product description workflow walks through exactly that, and it’s worth more now than it was a month ago.

What we still don’t know

A few honest unknowns. First, enforcement. The EU’s grace period for systems placed on the market before August 2 runs to December 2, 2026, so this rollout is OpenAI getting ahead of a deadline, not the end state. Second, the detector is researchers-only, which means nobody outside OpenAI can independently stress-test these claims yet — including the “no meaningful quality change” one. Third, the US and the rest of the world are untouched for now, and API watermarking is opt-in, so most readers’ day-to-day won’t change this month.

But the direction is set. Every major lab signed the same transparency code. More watermarking is coming — from more providers, in more regions. Treat this as the first domino, not a one-off EU quirk.

The takeaway

Watermarks don’t make AI text bad, and their absence doesn’t make text good. What matters — for client work, for product pages, for anything an AI search engine might quote — is still whether the words are specific, accurate, and edited by someone who gives a damn. If your robots.txt lets AI crawlers read your site, assistants will quote your pages regardless of who drafted them first. Invest in the editing, not in beating the detector. That’s the part no watermark can fake.