“Give the detector more text and it will get it right” is one of the most repeated pieces of advice in this field. We tested it on the 1,180 human-written papers from our false-positive study. Across that range, the risk of being wrongly flagged is flat. It does not fall as the text gets longer.
Within the range we can actually speak to — 150 to 392 words of published academic English — text length does not predict whether our detector wrongly flags human writing as AI. The false-positive rate is about 5.8% at the short end and about 5.9% at the long end. A formal test for trend across five length bands returns p = 0.777, which is as close to “no pattern” as this kind of test gets.
This is a negative result and we are publishing it because the opposite is widely believed, including by people who have been wrongly accused and told to submit more of their work as proof. Within this range, submitting more words does not lower the risk.
Read the limits section before quoting this. It says nothing about a 40-word forum comment or a 3,000-word dissertation chapter, and it was measured on one detector version against one corpus. It is a narrow finding, stated narrowly on purpose.
| Words | Documents | Flagged AI | False‑positive rate | 95% CI | Uncertain |
|---|---|---|---|---|---|
| 150–199 | 347 | 18 | 5.19% | [3.3–8.0] | 17.3% |
| 200–249 | 447 | 31 | 6.94% | [4.9–9.7] | 21.0% |
| 250–299 | 271 | 14 | 5.17% | [3.1–8.5] | 21.8% |
| 300–349 | 86 | 1 | 1.16% | [0.2–6.3] | 22.1% |
| 350–392 | 29 | 5 | 17.24% | [7.6–34.5] | 13.8% |
Every confidence interval overlaps every other one. The two bands that look like a story — 1.16% at 300–349 words and 17.24% at 350–392 — are the two smallest samples in the table, at 86 and 29 documents. The 350–392 interval runs from 7.6% to 34.5%. An interval that wide is not a finding, it is a shrug.
We are showing you the 17.24% rather than collapsing the top two bands to hide it. Merging bands after seeing the results is how a study gets flattering, and the whole point of publishing the per-document data is that you do not have to take our word for where the cut-offs went.
| Test | What it asks | Result | Reading |
|---|---|---|---|
| Cochran‑Armitage | Does the flag rate trend up or down across the five bands? | z = +0.284, p = 0.777 | No trend |
| Pearson r | Does the AI score itself rise with word count? | r = +0.023 (t = 0.78, n = 1,180) | No relationship |
| Median split | Is the shorter half flagged more than the longer half? | 5.79% vs 5.90% | Indistinguishable |
The median-split comparison is the plainest version. Split the corpus at its median length of 228 words and you get 587 shorter documents flagged at 5.79% [4.2–8.0] against 593 longer documents flagged at 5.90% [4.3–8.1]. Documents that were flagged averaged 234.0 words; documents that were not averaged 231.9. A two-word difference.
The obvious objection is that a length effect might exist in one group and be washed out by averaging. The people most often harmed by false positives are second-language English writers, so that is the split worth checking.
| Group | Half | Documents | Flagged AI | 95% CI |
|---|---|---|---|---|
| Second‑language English | Shorter | 394 | 4.31% | [2.7–6.8] |
| Longer | 365 | 4.93% | [3.1–7.7] | |
| Native English | Shorter | 193 | 8.81% | [5.6–13.7] |
| Longer | 228 | 7.46% | [4.7–11.6] |
Neither arm shows a length effect. Within each group the two intervals overlap almost completely. Second-language writing moves slightly up with length and native writing slightly down, and neither movement is distinguishable from noise.
The native-versus-second-language gap visible in this table is the subject of the false-positive study, which found the difference is not statistically significant either. It is repeated here only to show the length result is not hiding inside one group.
There is a second, quieter result in the table above. The rightmost column is the share of documents the detector declined to call confidently either way — the “Mixed / Uncertain” verdict. If more text gave the model more signal, that share should shrink as documents get longer. It does not: 17.3% at 150–199 words, then 21.0%, 21.8% and 22.1% as length increases.
That is the opposite of the expected direction, though the differences are small and we are not claiming it as a finding. What we can say is that within this range, adding words did not move a document out of the uncertain band, which is the band people actually get stuck in when they are asked to prove authorship.
Read this before citing the result. The finding is narrow and the honest version of it is narrower than the headline.
Everything above comes from two columns of a file we already published: word_count and flagged_ai, in the per-document CSV from the false-positive study. No new scoring was run, so there is no private step between the published data and these tables.
word_count at 150/200/250/300/350 and take the mean of flagged_ai per band.verdict column carries the shipped three-way label, so the uncertain share is read directly rather than derived from a threshold we chose afterwards.If you reach a different number we would rather hear about it than not: tell us.