Don't let anyone detect you used AI.

Every number on this page is either a live detector response we recorded and can reproduce, or a cited independent study. No invented benchmarks, no marketing decimals.

1.The detectors you have to get past

AI detectors are everywhere now — coursework portals, hiring funnels, client review, editorial desks — and their verdict sticks. You rarely get to argue with it. So the question is never "is my draft mostly clean"; it is whether the detector in front of you finds anything at all.

They fire hard. On September 1, 2026 we ran official, pre-LLM regulatory text from the US Federal Register — rulemaking drafted by government attorneys between 2015 and 2018, years before ChatGPT existed — through ZeroGPT's live web engine:

• CFPB consumer disclosure rule (2015): 100.0% AI.
• Medicaid managed care standards (2016): 82.2% AI.
• Paperwork Reduction Act notice (2018): 87.4% AI.
• CMS innovation model rule (2017): 100.0% AI.

Published research says the same about the category. A peer-reviewed Stanford study (Liang et al., 2023, Patterns) found commercial detectors flagged 61.2% of real TOEFL essays by non-native students as AI. A multi-tool evaluation in the International Journal for Educational Integrity (Weber-Wulff et al., 2023) put commercial tool accuracy at only 38–68%.

Take that as the measure of what you are up against: a detector that lights up on prose with no machine anywhere near it will certainly not miss the fingerprints a model actually leaves in your draft.

ZeroGPT figures: live responses recorded 2026-09-01 from its public web engine, reproducible request-for-request. Study figures: peer-reviewed independent research as cited — not our measurements.

2.What remove(ai) does: every fingerprint, one pass

AI-assisted text carries marks you cannot see by reading it. remove(ai) reads your draft the way a detector does and strips all four kinds in a single pass:

• Hidden characters. Zero-width joiners, unusual spaces, homoglyph substitutions and the statistical watermark carriers published schemes embed.
• AI phrases. The stock turns of phrase a model reaches for and a human writer rarely does.
• Machine-typical passages. The stretches that read as generated — even rhythm, predictable word choice, uniform sentence shape.
• Disclosure boilerplate. "As an AI language model", generated-with notices and the rest of the labelling that hands a reviewer the answer for free.

You see every change in a side-by-side diff and the result stays editable. We take the machine out, not your voice — nothing is rewritten just to look different.

3.How it is proven: the detectors grade us

We do not grade ourselves on our own benchmark. Cleaned text goes back through the very detectors people actually face. If one of them still catches something, we have not done our job — that case becomes the next thing we fix.

One recorded pass, end to end: a machine-written sample scored 56.1% AI on ZeroGPT before our cleanup and 25.4% after it — same detector, same text, re-submitted by a second person who got the same two numbers.

That loop runs on every iteration of the engine, against fresh frontier-model generations, and each run is written to the repository's committed iteration log. The detectors everyone else is afraid of are our free, always-on grader.

The 56.1 → 25.4 pair is sample FW-ACT-001, recorded 2026-09-01 and committed with the rest of the minimal-pair set. It is one sample, and a deliberately tell-heavy one: proof that the mechanism works on text full of the marks we target, not an average over all writing.

4.Where we are, honestly

No text tool can promise to beat every detector forever — they change weekly, and we say so plainly rather than quoting an accuracy number nobody can stand behind.

What is working today: the visible tells — hidden characters, watermark carriers, stock phrasing, disclosure labels, machine formatting — are found and cleaned. The last mile is the subtle machine-typical prose the strongest detectors still read, and that is exactly what the self-improvement loop works on, every iteration.

5.See for yourself

Paste your AI-assisted draft on the front page, read the diff, then take the cleaned text to any detector you like — including the ones above. That is the whole pitch.