OpenAI just made a move that will quietly reshape how AI-generated text is tracked — and most people won’t even notice it happening.
On October 5, 2026, the company announced textGrain, its text watermarking system, which embeds an invisible statistical signal directly into the words its models choose. The watermark rolls out to eligible ChatGPT and Codex users across all plans in the European Union over the coming weeks. API customers worldwide can opt in starting today, though it remains off by default.
This isn’t a visible stamp or a hidden character. It’s something far more subtle — a pattern woven into the model’s word choices that readers can’t see but a detector can pick up. And it’s all happening because of a European Union law that took effect on August 2, 2026.
The EU AI Act Changed the Game
Article 50 of the EU AI Act introduced transparency rules requiring generative AI providers to mark their output in a machine-readable way. The goal is straightforward: help people recognize when they’re interacting with AI-generated content. The law applies from August 2, 2026, with a grace period until December 2 for systems already on the market before that date.
OpenAI’s response is textGrain. The system adds an invisible statistical signal to the model’s word choices — not by inserting hidden characters or unusual punctuation, but by subtly adjusting how the model randomly selects between possible words or word pieces as it writes. Across a passage, these choices create a pattern that a detector can look for, even after some edits.
The company was careful to set expectations. In its announcement, OpenAI stressed that text watermarking and detection remain early technologies with significant limitations. A missing watermark doesn’t prove human authorship — the text could be too short, too heavily edited, or generated by another company’s AI.
How textGrain Actually Works
The mechanism is elegant in its simplicity. When a language model generates text, it typically picks the next word from a probability distribution. textGrain tweaks that selection process using a secret key, creating a statistical drift in the chosen tokens that’s detectable over a long passage but invisible to any reader.
Think of it like a chess engine that doesn’t just pick the best move — it picks moves that follow a specific, verifiable pattern only if you know the key. The moves are still good moves, but they carry a signature.
OpenAI’s own evaluations show the system works, but with important caveats. At a target false-positive rate of 1 percent, the detector identified watermarks in around 80 percent of 200-token passages, rising to about 95 percent for 400-token passages. Shorter texts are harder to verify.
Editing is the Achilles’ heel. In tests with 400-token passages, replacing 10 percent of words with synonyms reduced detection from about 92 percent to 66 percent. Replacing 25 percent of words brought it down to 17 percent. A complete rewrite where every word is replaced will defeat the watermark entirely.
What the Watermark Can and Cannot Tell You
OpenAI was unusually transparent about the limitations. The watermark does not identify the user. It does not measure human contribution. It does not establish ownership or responsibility. And it does not verify accuracy.
“Watermarks can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it,” the company wrote.
This is a crucial distinction. A watermarked text might be 100 percent AI-generated, or it might be a human draft that ChatGPT lightly edited. The watermark can’t tell the difference. And conversely, the absence of a watermark doesn’t prove a person wrote the text — it might just mean the text was too short, too heavily edited, or produced by a different AI system.
The Competitive Landscape
OpenAI isn’t alone in this. Anthropic has been watermarking Claude’s text since August 2, 2026, using a version of Google’s SynthID-Text approach published in Nature in 2024. Google has deployed SynthID-Text in Gemini since 2024. Both companies, along with Meta, Microsoft, and around 190 other signatories, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026.
The approaches differ in implementation but converge on the same principle: invisible, model-level watermarking that travels with the text when it’s copied and pasted. Anthropic’s watermark is enabled by default with no user opt-out. OpenAI’s is opt-in for API customers and limited to the EU for ChatGPT and Codex users.
This regional approach gives OpenAI room to learn from real-world use and feedback before considering a global rollout. It also reflects the reality that the EU is currently the only jurisdiction with enforceable AI transparency requirements.
Why This Matters Beyond Compliance
On the surface, textGrain is a compliance play — a company responding to a legal requirement. But the implications run deeper.
First, it normalizes the idea that AI-generated text should be identifiable. Once users in the EU expect watermarks, that expectation will spread. The technology becomes a feature, not a bug.
Second, it creates a detection infrastructure. OpenAI is opening applications for access to its text watermark detector, initially limited to approved researchers and expert organizations — a move that echoes Google’s own struggles with AI-generated submissions. This is the beginning of an ecosystem where AI-generated content can be verified at scale — and where AI-generated phishing emails might get caught earlier.
Third, it sets a precedent for how AI companies handle transparency. OpenAI’s decision to be upfront about the limitations — rather than overselling the technology — is a sign of maturity. It’s also a recognition that watermarking is not a silver bullet.
The Arms Race Is Already Here
Where there’s a watermark, there’s a watermark remover. The GitHub project watermarks-remover already has over 4,500 stars. Tools like StealthGPT and Human Writes promise to strip AI signals from text. Most of these work by replacing metadata or requiring full rewrites — exactly the kind of heavy editing that degrades watermark detection.
This is the fundamental tension: watermarking is probabilistic, not absolute. It can indicate likelihood, not certainty. And as detection improves, so does evasion.
I’ve written before about the broader landscape of AI text detectors and what they change. The OpenAI announcement is the next chapter in that story — the moment when the largest AI company in the world decided that invisible provenance is worth building, even with all its imperfections.
The Bottom Line
textGrain is not a lie detector. It’s not a copyright registry. It’s not a measure of human creativity. It’s a statistical signal that says, with varying degrees of confidence, “an OpenAI system was involved in producing this text.”
That’s both less and more than it sounds like. Less, because it can’t tell you how much the AI contributed or whether the result is accurate. More, because it creates a verifiable trail that didn’t exist before — a way to ask, and sometimes answer, the question of where text came from.
The EU AI Act forced this moment. But the technology it’s producing will outlast the regulation that spawned it. Watermarks are becoming part of the fabric of AI-generated text, whether we notice them or not.
I’d rather know. Even if the answer is probabilistic. Even if it’s imperfect. Even if a determined editor can strip it away. The alternative — a world where AI-generated text is completely indistinguishable from human writing, with no way to ask — is worse.
The watermark is invisible. The choice to build it was not.