01Partition
Group candidate tokens
A secret key and preceding context divide the vocabulary into blocks. A token is a word or part of one.
AI TEXT PROVENANCE
Text GrainA watermark in
the words themselves.
OpenAI’s textGrain embeds an invisible statistical signal in word choices during generation. A detector checks for that signal to identify an OpenAI watermark.
Embedded at generationDetected statisticallyNot a certainty
◎ An independent guide to the technology
TEXTGRAIN / SIGNAL FLOW ILLUSTRATION
01 / GENERATE
Context ＋ Secret key
Token groupsBlock ABlock BBlock C
Sample within an entropy budgetEmbed a statistical signal in token choices
Readable text. Invisible watermark.The signal lives in the choice of tokens.
02 / DETECT
Text ＋ keyStatistical test
Is an OpenAI watermark detected?
A closer look at text provenance
◌ Invisible watermark▧ Statistical signal↗ Source transparency
TEXT / TRUST / TRANSPARENCY01 / THE CORE IDEA
What is text grain?
A signal added at the source.
To understand the technology, start with two separate actions: embedding and detection.
textGrain checks for a deliberately embedded watermark, rather than judging writing style. Its scope is an OpenAI watermark, not every kind of AI-generated text.
Follow the signal02 / HOW IT WORKS
From token choices
to statistical evidence.
Three steps to follow.
One technical report to go deeper.
02Sample
Balance signal and randomness
Optimal transport selects blocks under an entropy-loss budget. Within each block, relative token probabilities stay unchanged.
03Detect
Reconstruct and test
Text and the key reconstruct the statistical signal. Detection does not require the generation-time entropy budget.
OFFICIAL TECHNICAL FIGUREFIGURE 1 / BLOCK OPTIMAL TRANSPORT
Source: OpenAI textGrain technical report, page 4, Figure 1. Diagram excerpt.View original ↗
Technical reference: textGrain: Entropy-Calibrated Watermarking for Language Model Text ↗
03 / MEASURED PERFORMANCE
How reliable is detection?
Length and editing matter.
Longer passages, stronger detection
Psychology-like content · target false-positive rate: 1%
200 tokens
≈80%
400 tokens
≈95%
Math content had substantially lower detection rates.
Editing weakens the signal
Separate evaluation · 400-token English passages
92%
Unedited
66%
10% replaced
17%
25% replaced
Different experiments: the baseline rates are not interchangeable.
Charts redrawn from OpenAI’s October 5, 2026 evaluations. These detection rates are not a guarantee of real-world accuracy.
No verdict on content or authorship
A watermark does not establish truth, human contribution, ownership, or responsibility.
False positives and misses are possible
No detected watermark does not prove human authorship.
No user identification
The signal does not identify accounts, prompts, or conversations.
05 / AVAILABILITY
Where is textGrain available?
Launch plans as of October 5, 2026.
Check the official source for updates.
Global opt-in
Select models; watermarking is off by default.
Phased EU rollout
Eligible text outputs over the coming weeks.
Approved access
Initially researchers and expert organizations, not the general public.
Open sourcing is planned, not confirmed as released. Read the announcement ↗
06 / QUESTIONS, ANSWERED
A little more clarity
on text grain.
Start with the question that brought you here.
Is text grain the same as textGrain?
textGrain is OpenAI’s technology name. This independent guide uses “text grain” as its title to help readers find and understand it.
Does watermarking affect output quality?
OpenAI reports no meaningful differences in its published Astra benchmarks. That does not guarantee identical quality in every setting.
Where can I find the original research?
Use the official article and technical report links on this page. Our guide was reviewed on October 6, 2026; consult the original sources for later updates.
Can I check my text on this website?
This site is an illustrated guide, not a detector. It does not collect or upload your text. See the official article for detector access information.
START WITH THE SOURCE
Understand the signal.
Read the research.
Take the next step with OpenAI’s original article and technical report.