| Test | Text length | Found |
|---|---|---|
| Normal topics, like psychology | 200 tokens | About 80% |
| Normal topics, like psychology | 400 tokens | About 95% |
| Maths | 200 to 400 tokens | Much lower (no number given) |
| No changes (edit test) | 400 tokens | About 92% |
| 10% of words changed to synonyms | 400 tokens | About 66% |
| 25% of words changed to synonyms | 400 tokens | About 17% |
textGrain detectability calculator
textGrain is a pattern spread over many words. So how easy it is to detect depends mostly on two things: how long the text is, and how much it was changed. Move the sliders to see how this works, using OpenAI’s own numbers.
Detectability calculator
Kind of text
Chance that OpenAI’s detector finds it
Unedited10% edited25% editedYour settingOpenAI’s published results
Our estimate, based on the numbers OpenAI shared on 5 October 2026. It does not read your text and cannot tell if it is AI.
OpenAI tuned its detector to flag only 1% of normal text. At that setting it found about 80% of 200-token passages and about 95% of 400-token passages. When 10% of the words were changed to synonyms, detection on 400-token passages fell from about 92% to 66%. When 25% were changed, it fell to 17%.
The numbers OpenAI published
From OpenAI’s announcement on 5 October 2026, as reported by the press. In every test the detector flagged only 1% of text without a watermark.
How the estimate works
A word watermark gets easier to find as the text gets longer, because there are more words to check. We drew one smooth curve through OpenAI’s two points for normal text: 80% at 200 tokens and 95% at 400 tokens. Then we lowered it for edits, so it passes through 66% at 10% changes and 17% at 25% changes.
The formula is: detection = Φ(0.564 × tokens0.326 × edit factor − 2.326). Φ is the normal curve from statistics. 2.326 is the cut-off that flags only 1% of normal text.
Points between and beyond OpenAI’s tests are our estimate, not OpenAI’s. OpenAI gave no number for maths or code, so for those the calculator shows the normal-text number as the highest possible value.
What makes detection harder
- Short text
Fewer words means fewer chances for the pattern to show. A tweet or a short email reply is almost impossible to detect.
- Fixed wording
Maths, code, lists, quotes, names and formulas leave the model few word choices. That leaves less room for a watermark.
- Editing
Every changed word breaks the pattern around it. OpenAI’s numbers show a big drop: with a quarter of the words changed, only 17% were found.
- Translation
OpenAI is still testing how watermarks hold up after translation. Translation replaces most word choices, so expect a big drop.
- Mixing with your own writing
A few AI sentences inside a long text you wrote yourself make the signal weaker. A detector that checks the whole text may miss them.
Questions people ask
Does this calculator check if my text is AI?
No. It only uses the length you set. It answers a different question: if a text this long had textGrain and was changed this much, how often would OpenAI’s detector find it?
What does a 1% false positive rate mean?
The detector is set so that only about 1 in 100 texts without a watermark gets flagged by mistake. That sounds small. But across thousands of essays, it means real people flagged for nothing.
How many words is 400 tokens?
About 300 English words. Other languages often need more tokens for each word.
Why is maths harder to detect?
In maths there is usually one right way to write each step. So the model rarely has a choice of words, and with no choice there is no room for a watermark.