Home › Research

We measure our own detector in public, including when it looks bad.

AI detection is full of confident percentages with no corpus behind them. We publish the dataset, the confidence intervals, the results that did not go our way, and a plain list of what we have not measured. Everything here is downloadable and reusable under CC BY 4.0.

Studies 2 Documents tested 1,180 Datasets 1 (CC BY 4.0) Updated 23 August 2026

What we hold ourselves to.

These are commitments about how we publish, not claims about how good the product is.

  • The data ships with the claim. Every study publishes per-document results, not just a summary. If we say 5.85%, you can count the rows.
  • Intervals, not point estimates. A rate without a confidence interval hides its own sample size. Ours are Wilson intervals at 95%.
  • Negative results get published. Both studies so far report a finding we expected to see and did not. That is the whole reason to run the test.
  • Thresholds are fixed before the analysis, not after. Where a cut-off could have been chosen to flatter us, we show every cut-off instead.
  • Limitations are on the page, not in a footnote. Each study carries a section saying what it cannot support.
  • No single headline accuracy number. One figure across every length, genre and language would be misleading, and it is exactly the move overclaiming detectors make. Our reasoning is in the methodology.

Published studies.

Study 01 · 17 Aug 2026

False positives on 1,180 human papers

Every document was published in 2018, years before ChatGPT, so every AI flag is a false positive by construction. Our detector flagged 5.85% of them. We also went looking for the well-known bias against second-language English writers and did not find it.

1,180 documents5.85% false positivesCC BY 4.0 Study 02 · 23 Aug 2026

Text length and false positives

“Submit more text and the detector will get it right” is standard advice, including to people who have been wrongly accused. Between 150 and 400 words we found no relationship at all: 5.79% in the shorter half, 5.90% in the longer half, trend test p = 0.777.

5 length bandsp = 0.777 no trendCC BY 4.0

The data.

One dataset currently backs both studies. It is the per-document output of the false-positive run: one row per paper, with word count, first-author affiliation group, the score the detector returned and the verdict it shipped.

Licensed CC BY 4.0. You may republish, redistribute and build on it commercially, including to argue against our conclusions, provided you credit the source.

TextSight (2026). TextSight false-positive rate on human-written academic English. Lacewing Technologies. CC BY 4.0. https://www.textsight.ai/benchmark/

What we have not measured.

This list exists so nobody has to guess what our silence covers. These are real gaps, stated plainly.

  • No head-to-head benchmark against other detectors. We have not run GPTZero, Copyleaks, Originality.ai, ZeroGPT or Turnitin against a shared corpus, so we publish no comparative accuracy figures anywhere on this site. Where our comparison pages discuss other tools, they compare features, pricing, formats and privacy — not measured accuracy.
  • No student-essay corpus. Both studies use published academic abstracts. Student work is the case people actually worry about and it is the obvious next study.
  • No non-English evaluation. The detector is calibrated on English and we do not publish reliability figures for other languages, because we have not measured them.
  • No independent third-party audit. Everything here is our own testing of our own detector, which is a real limitation however carefully it is done. We say so on every study.
  • No measurement of AI text that has been edited or paraphrased. The corpora here are entirely human-written, by construction. False negatives are therefore out of scope.

For researchers and journalists.

The data is open and the method is written to be rebuilt. If you are running a study in this area, three things may be useful:

  • You can replicate without asking us. The per-document CSV plus the manifest is everything a re-analysis needs. Study 02 is itself a re-analysis of Study 01's data and no new scoring was run for it.
  • We would rather be corrected than quoted uncritically. If your numbers disagree with ours, tell us and we will publish the correction.
  • The limitations are the interesting part. The gaps listed above are open questions, and several of them are more answerable by a university than by us.

Written by Dipak Bhosale, who runs the detection models these studies test.