It’s a Tuesday morning in Lisbon. You’ve got a client deadline at noon, ChatGPT open in one tab, a half-finished product description in the other. You paste a paragraph, tidy it up, ship it. Nothing on your screen looks different. But starting in the coming weeks, if you’re sitting in the EU, that text may carry something you can’t see: an invisible watermark, baked in by OpenAI.
OpenAI announced on October 5, 2026 that it’s rolling out text watermarking for eligible output from ChatGPT and Codex in the European Union. The reason is regulatory, not philosophical. The EU AI Act’s transparency requirements under Article 50 call for machine-readable marking of AI-generated content, and those rules have applied since August 2, 2026, with systems already on the market given until December 2 to fall in line.
So this is compliance engineering. Which, as someone who spends most of his week poking at AI tools to see where they crack, I find more interesting than another model release.
What’s actually changing
Three things worth separating, because the coverage has been blurring them:
- ChatGPT and Codex text output in the EU gets an invisible watermark on eligible text. Not optional for those users, as far as the announcement describes.
- In the API, watermarking stays off by default — everywhere, including the EU.
- API customers anywhere in the world can opt in to text watermarking for select models.
That split tells you who OpenAI thinks the regulation is aimed at. Consumer-facing chat gets the marking automatically. Developers building on top of the API keep the switch in their own hands, and if they want it on, they can turn it on from Ohio or Osaka just as easily as from Oslo.
The rollout is described as happening “over the coming weeks.” No firm start date, no completion date. If you’re planning a content workflow around this, plan for ambiguity.
Why “eligible text” is the phrase to watch
OpenAI’s framing refers to eligible text outputs, and “select models” on the API side. Neither term is defined in what’s been made public so far. For anyone who reviews tools for a living, that’s the part I’d want pinned down before making promises to a client.
Does a three-word completion get marked? A code block? A translation? Text you’ve heavily rewritten? Watermarking schemes for text generally depend on having enough tokens to encode a detectable signal, which implies short outputs are harder to mark reliably. I can’t tell you where OpenAI drew that line, because OpenAI hasn’t said. I’d rather admit that than guess.
Same goes for durability. If you paste watermarked text into a doc, cut half of it, reorder the rest, and run it through a human edit pass, does the signal survive? Unknown from the announcement. Anyone selling you a confident answer right now is extrapolating.
What this means if you build with these tools
For most people reading agntbox, the practical takeaway is narrow but real. If you’re an EU-based solo operator using ChatGPT directly, assume your output may be markable from here on. That’s not a scandal — it’s a label, and it’s invisible to readers. But it does change the risk calculus for anyone who’s been quietly passing model output off as fully hand-written. Not a judgment, just arithmetic.
If you’re building on the API, you now have a decision to make rather than a default to accept. Turning watermarking on is a trust signal you can show clients, publishers, or platforms. Leaving it off keeps your output indistinguishable. Both are defensible. What’s not defensible is not knowing which setting you’re running.
And if you run a detection or provenance product, this is the first time a major vendor has shipped text marking at this scale with a regulatory deadline behind it. Whether detection tooling can read OpenAI’s mark, and who gets access to that capability, is the question I’d be chasing. Nothing in the announcement answers it.
My read
This is a modest, honest piece of plumbing shipped under deadline pressure, and I’d rather have it than not. The uncomfortable part isn’t the watermark — it’s the geography. Europe gets transparency by default because a law required it. Everyone else gets a checkbox. One market’s regulator is effectively setting the baseline for what AI provenance looks like, and the rest of us are opted out until someone legislates otherwise.
I’ll be testing what actually survives a normal editing workflow once the rollout lands. Until then, treat the details as provisional and the direction as settled.
🕒 Published: