Around October 5, OpenAI released a text-watermarking technique called textGrain and published the accompanying paper. In plain terms, it buries an invisible statistical signal into AI-generated text that human readers cannot see but detectors can recognize.
How it stays invisible
It does not change your words, add symbols, or touch punctuation. The mechanism nudges the model's word-choice probabilities at the moment it writes, giving the whole passage a rhythm a reader cannot feel but a detector can spot. Because the watermark lives inside the text itself, it travels with the words when you copy and paste.
The EU law is the trigger
The direct trigger is Article 50(2) of the EU AI Act. That transparency obligation took effect on August 2 this year and requires AI-generated content to be marked in a machine-readable way. OpenAI's move is a direct answer to regulation.
Rollout splits in two
Rollout splits in two. Inside the EU, ChatGPT and Codex text will carry the watermark by default over the coming weeks, and only in the EU. On the global API, developers can opt in per model, and the default is off. That means for most users outside Europe, ordinary text carries no signal for now.
Detection has limits
Detection has hard limits. OpenAI admits a 200-token passage is caught at roughly 80 percent confidence, while a 400-token passage reaches about 95 percent. Yet swapping just 10 percent of words for synonyms drops detection from 92 percent to 66 percent; replace a quarter and it collapses to 17 percent, effectively useless. Short text, math, code comments, and translations were always hard to tag.
The detector stays private
The detector is not public. OpenAI grants early access only to vetted researchers and expert organizations. It also will not claim that missing watermark proves human authorship: a passage can lose its signal by being too short or too edited, or it may come from another vendor's model.
Backdrop: Anthropic moved earlier
A useful backdrop: Anthropic added a SynthID-Text-based watermark to Claude two months earlier, and did so globally. OpenAI is catching up here, and says the algorithm will be open-sourced later.
An arms race behind the signal
Behind this signal is an arms race. Where there is a watermark, someone builds a remover. Soon after Anthropic shipped SynthID-Text, open-source tools appeared on GitHub that scramble the statistical signal through rewriting. OpenAI's move to keep the detector closed, granting access only to vetted institutions, is a defensive choice. But defense rarely wins outright. Given enough freedom to paraphrase, any statistical watermark gets diluted.
The real industry impact
The real impact on the content industry is not whether text can be detected, but whether platforms will start demanding self-disclosure. As watermarking spreads, schools, newsrooms, and e-commerce listings may ask you to declare AI use. What you submit becomes not just a piece of text but a source note. Building the labeling habit now beats scrambling later.
Do not treat it as the sole test
Do not treat a watermark as the sole test of truth. OpenAI stresses repeatedly that missing watermark does not prove human authorship, and a present watermark only means an OpenAI system touched that passage. A heavily rewritten or translated text may have lost its signal long ago and still be AI-generated. Use the watermark as a clue, never as a verdict.
C2PA runs alongside
Running alongside the watermark is a separate standard called C2PA content credentials. It is not the same thing. The watermark lives invisibly inside the text, while content credentials are source and history records written into file metadata, also used by camera makers and newsrooms. OpenAI says where technically possible it applies both, so provenance survives editing better. The catch is that everyday actions like screenshots and format conversion can strip metadata, and then only the watermark remains as a last line.
The writer's moat
For writers, the real moat shifts from 'sounds human' to 'can prove origin.' In the future, clients, platforms, and schools will less often ask the yes-or-no question of whether text is AI-written and more often ask the explanatory question of which steps you used AI for and where your own judgment sat. The earlier you build a habit of labeling sources, the less likely you are to get stuck when rules land. Treat AI as a collaborator, record the collaboration, and you stand on firmer ground than betting it stays undetectable.
Set rules early for enterprise
For enterprise content teams this line matters early. Set internal rules now for what AI-assisted material gets labeled, what stays unlabeled, and who verifies it. Do not wait until platforms start forcing fake-content detection before discovering you have no internal policy at all.
What this means for you?
If you use AI to write, do SEO, or publish at volume, watch this thread. In the short term, only EU users get watermarked text by default, and the global setting is off, so daily impact is small. The real lesson is to stop assuming AI writing is undetectable. Watermarking will spread, and platforms, schools, and newsrooms may treat it as a provenance clue. Using AI as a tool is fine, but label deliverables where labeling is due and verify by hand where it matters. Do not bet on nobody noticing.