On October 5, OpenAI announced its plan to introduce textGrain,
which embeds an invisible watermark in generated text.
API customers worldwide can opt in for select models,
and over the coming weeks the watermark will be rolled out
to eligible ChatGPT and Codex text output in the EU.
The company is also accepting applications
for access to its watermark detection tool.
What Changes: Adding an Identifying Signal at the Generation Stage
OpenAI describes the move as a response to the EU AI Act.
According to the company’s announcement, the law requires
generative AI providers to make generated text
identifiable in a machine-readable way.
textGrain adds a statistical pattern to how the model chooses
words and word pieces as it generates text.
According to OpenAI’s help article, it does not insert hidden characters,
invisible spaces, or unusual punctuation into the text.
The detection signal is carried in the word choices
of the visible text itself.
Unlike detectors that estimate whether text looks AI-generated
by analyzing its characteristics after generation,
this approach looks for a signal embedded at generation time.
The technical report describes a mechanism that links
word-piece selection to randomness based on a secret key,
allowing a detector to use the text and the key
to examine statistical evidence.
Off by Default in the API, Phased Rollout for EU Products
In the API, watermarking is off by default,
and customers enable it from their organization or project settings.
They can choose which models to apply it to,
and the list of supported models is available in the settings.
When enabled, the watermark is added during text generation
for supported models.
For ChatGPT and Codex, the company plans to roll out watermarking
to eligible users and outputs in the EU across all plans
over the coming weeks.
It is not becoming a global default at launch.
Enabling watermarking in the API also does not grant access
to the text detection tool.
The detector will initially be available only
to approved researchers and expert organizations.
Detection Is Harder for Short or Rewritten Text
OpenAI explains that watermark detection can produce
both false positives and false negatives.
In the company’s evaluations, at a target false positive rate of 1%,
the detector identified watermarks in about 80% of 200-token passages
and about 95% of 400-token passages for content such as psychology.
Detection rates were substantially lower for content such as mathematics,
where there is less flexibility in word choice.
According to OpenAI’s help article, the EU Code of Practice
on the Transparency of AI-Generated Content does not require watermarks
in outputs shorter than 200 tokens (about 150 words in English)
or in code snippets.
The company explains that short passages usually do not contain
enough text for reliable detection,
and that code is harder to watermark because it offers
fewer plausible word choices than ordinary prose.
Rewriting also affects detection.
In a separate evaluation of 400-token passages,
replacing 10% of words with synonyms reduced detection
from about 92% to 66%, and replacing 25% reduced it to 17%.
This evaluation used English text, and the figures do not indicate
that the same results apply to all languages or editing methods.
LC Signal’s Perspective: Separating the Generation Signal from Human Involvement
OpenAI states that detecting a watermark does not reveal
how much human judgment, editing, or creativity went into a passage.
It does not identify the user or the conversation,
nor does it verify accuracy or establish ownership.
Conversely, the absence of a detected watermark
does not prove human authorship.
The company is starting with opt-in use in the API
while limiting who can access the detector,
so that it can evaluate and improve the technology.
A mechanism that leaves clues about a text’s origin
adds another way to check for AI use.
Still, the distinction that such results cannot determine human contribution
seems likely to matter in workplaces that use AI to produce writing.
Key Sources
-
OpenAI official announcement, October 5, 2026:
Our approach to EU text provenance rules
-
OpenAI official help article (scope, settings, and how it works):
Provenance signals in OpenAI-generated content
-
OpenAI technical report, October 5, 2026:
textGrain: Entropy-Calibrated Watermarking for Language Model Text
This article was prepared with AI assistance using publicly available information
as of October 6, 2026.
Announcement dates follow the dates shown on the official pages.
Descriptions of the EU AI Act and the Code of Practice
reflect explanations in OpenAI’s published materials.
Detection rates are evaluation results under specific conditions presented
by the company and do not guarantee detection for any individual passage.
Features already available are distinguished from planned rollouts.
“LC Signal’s Perspective” reflects this publication’s interpretation
of the available information.