neeporaneepora
Start free
Results log

Every test, failures included.

GPTZero Model 4.9b, paid account, sentence view, 2026-09-13 to 15. Not the best runs picked out. What did not work matters as much as what did.

What did not work

Every common way of "removing AI tells" left the verdict unchanged.

AttemptWordsVerdict
Delete AI vocabulary (delve, crucial, moreover…), vary sentence length, add concrete numbers1,900AI 100%
Same text in eleven registers: essay, academic, Korean business English, plain, rare vocabulary, citation-dense…600–1,300 eachAI 100%, all eleven
Keep the original sentences and edit them lightly2,400AI 100%
Borrow a human skeleton but keep only function words, replace every content word2,079AI 100%
Wrong genre of skeleton (theory paragraphs carrying an argument)1,100AI 100%
Mix an essay voice and an official voice in one document4,134Mixed 88%
Append the reference list to the scan4,688Mixed 89%
One sentence accidentally duplicated4,292Mixed 94%

What worked

Borrow the skeleton of pre-2022 human writing, keep at least 80% of its words, change only what the subject forces.

DocumentDonorWordsDraftAfter
Thesis §1.2 (research questions)Korean FSC press release, 2023614AI 100%Human 100%, 0/18
Thesis §1.1 (background)Wikipedia, 2021700AI 100%Human 98%
Thesis §6.3 (interpretation)FSC 2023, 80% of tokens kept605AI 100%Human 89%
Thesis ch. 1–4, wholeWikipedia 2021 + FSC 2020–22, three quotations9,017AI 100%Human 90%, highly confident
Thesis ch. 5–7, wholeFSC 2020–22 surveys and briefings4,263AI 100%Human 89% (rescan 88–89)
Language-learning essay, one pagePaul Graham essay, 2013622AI 100%, 21/23 flaggedHuman 90%, 0/35
Stablecoin policy brief, one pageFSC 2020-11, 2021-05471AI 100%, 16/16Human 90%, 0/16
Tea article, one pageWikipedia "Coffee", 2021774AI 100%, 23/23Human 90%, 0/36

"Flagged" is the count of sentences marked AI in sentence view. Discussion and conclusion chapters are the floor of this method, moving between 80 and 89%.

Ten rules we learned

  1. No sentence of the original survives. One lightly edited sentence sends the whole chapter to AI 100%.
  2. Keep at least 80% of the donor's words. Borrow only the skeleton and replace every word: AI 100%.
  3. Pick a donor whose subject is close to yours. Where a donor sentence is simply true, quote it with a citation.
  4. Match the genre. Narrative from encyclopedias, rules and data from official documents, argument from briefings or essays.
  5. One fact per frame, at the donor sentence's length.
  6. Build and scan in 600–1,000-word sections, then join. The document verdict is roughly a word-weighted average.
  7. One donor voice per document. Mixing voices trips the mixed-authorship verdict.
  8. Keep reference lists and number tables out of the scan; keep tables as table objects.
  9. Number-heavy quotations and duplicated sentences trip the mixed-authorship head.
  10. Rescan noise is ±3. 87 and 93 are the same result.