AntiGPT

Results

We publish what we measure, when we measured it, and against which detector. Nothing on this site claims a pass rate that isn't on this page.

DetectorModePassed beforePassed afterHuman controls flaggedMeasured
Hello-SimpleAI RoBERTa detector (open model)balanced50% of 1688% of 1633% of 6

"Passed" = scored below the detector's likely-AI line (60% AI probability). Small sample; treat as an indication, not a guarantee.

Over time

Every run stays in the record. Losses are published with the wins; detectors update without notice, so expect movement.

0%25%50%75%100%21 Sept 2026: passed 63%, controls flagged 0%24 Sept 2026: passed 75%, controls flagged 33%28 Sept 2026: passed 88%, controls flagged 33%28 Sept 2026: passed 75%, controls flagged 33%5 Oct 2026: passed 88%, controls flagged 33%12 Oct 2026: passed 94%, controls flagged 33%19 Oct 2026: passed 88%, controls flagged 33%26 Oct 2026: passed 88%, controls flagged 33%2 Nov 2026: passed 88%, controls flagged 33%9 Nov 2026: passed 81%, controls flagged 33%16 Nov 2026: passed 88%, controls flagged 33%2026-09-212026-11-16
Hello-SimpleAI RoBERTa detector (open model) (balanced)Solid: AI samples that passed after one pass · dashed: human controls flagged

Raw data: download every per-sample score as CSV (11 runs). Licensed CC BY 4.0. Cite the run date.

How we test

Fixed sample set (eval/set.json, versioned; samples are added, never removed). Every AI sample was generated by a large language model (some prompted to sound human); human controls are public-domain passages by named authors, plus contributed modern writing where marked. Each AI sample is scored, humanized once (balanced strength, Standard readability), and scored again; controls are scored once. pass_rate = share of AI samples scoring below the detector's 'likely AI' threshold (0.60). control_flag_rate = share of human controls at or above it. Scores are estimates; raw rows are published as CSV.

The set is fixed so runs are comparable over time; we add samples but never remove them. Runs are dated, and older runs stay in the record. The detector we use is the same one behind the in-app checker, so the score you see on your own text is measured the same way.

Where the samples come from. Every AI sample was generated by a large language model; four of them were deliberately prompted to sound human, because that is what a worried student asks for. The human controls are public-domain passages by named authors (Thoreau, Twain, Douglass, Lincoln, Austen, Du Bois), chosen because their authorship is beyond dispute. They are older in style than a 2026 essay, which we note rather than hide; modern human writing is added as people contribute it, and is labelled as such in the CSV.

Two things matter about that design. First, human controls: if our own human samples get flagged, the detector is noisy and any improvement is suspect, which is why the flagged controls sit next to every pass rate. Second, dates and versions: detectors change without notice, so a result only means something attached to the day and the detector it was measured on.

What we will and won't publish. We publish measured pass rates (the share of the set that moved below the detector's "likely AI" threshold) with the date, the detector and the method, and we update them as the detector changes. We will not print a number we haven't measured, and we will never say "undetectable" or "guaranteed", because neither is a property any text can have. When a detector update reverses a result, this page will say so: the point of measuring is to tell the truth about a moving target, not to freeze a flattering snapshot.

What this does not show: results on detectors we don't run, results on your specific text, or results in a classroom or institutional setting. Detector scores are estimates from third-party classifiers and change as those tools update. No text can be guaranteed to pass any detector. Use this to improve your writing, and follow your institution's or employer's rules on AI use.