Blog

Writing craft, how detectors work, measured results, and product notes. New posts on Tuesdays.

• DetectorsWhy "undetectable" is a marketing word, not a technical one

  No text can be guaranteed to pass an AI detector, because detectors are probabilistic classifiers that change without notice. Here is what the word actually hides, and what an honest tool promises instead.

  Sep 20, 2026 · 5 min

• WritingWriting in English as a second language with AI assistance

  AI assistants are genuinely useful if English is your second language: they fix grammar and suggest idiom fast. The catch is that 'polished' output can read as nobody's and trip detectors even when the ideas are entirely yours. The fix is to correct with an assistant, then restore your own voice. And keep your drafts.

  Sep 16, 2026 · 5 min

• WritingAcademic integrity and AI tools: a plain guide to staying within the rules

  The rules vary by institution, so the first step is always to read your own policy. Most allow some AI assistance with disclosure; some forbid it for assessed work. A humanizer changes how text reads, not whether help was used. It never removes an obligation to disclose, and it never makes work yours if it isn't.

  Sep 9, 2026 · 6 min

• WritingHumanizing vs paraphrasing vs rewriting: what changes and what doesn't

  Paraphrasing swaps words for synonyms and keeps the structure. Rewriting changes the structure but not necessarily the voice. Humanizing targets the specific signals (uniform rhythm, predictable phrasing) that make text read as machine-made, while keeping your meaning. They overlap, but they're not the same thing.

  Sep 2, 2026 · 5 min

• DetectorsFalse positives: what to do when your own writing is flagged

  If a detector flags text you wrote yourself, you're not alone and you're not necessarily doing anything wrong: short, formal, or non-native English writing is more predictable and gets flagged more often. Keep your drafts, vary your sentence rhythm, add specifics, and ask for a human review. A score is not proof.

  Aug 26, 2026 · 5 min

• ProductHow word counts work on antigpt (and why failed requests are free)

  We count the words in the text you submit, once, when you press Humanize: whitespace-separated tokens containing at least one letter or digit. Output length never counts. If a request fails, the words are released back to your balance. You're never charged for a rewrite you didn't get.

  Aug 19, 2026 · 4 min

• ResearchHow we test our output against detectors: our method, and what we'll publish

  We run a fixed set of texts through several detectors before and after humanizing, on a schedule, and record the change. This post explains the method. We publish measured pass rates with the date and the detector versions once we have them, and never a number we haven't measured.

  Aug 12, 2026 · 5 min

• DetectorsHow AI detectors actually work, and why they misfire

  AI detectors estimate two things: how predictable your text is to a language model (perplexity) and how much that predictability varies sentence to sentence (burstiness). Newer ones add trained classifiers. All of them output a probability, not a verdict, which is why they misfire on short, formulaic, or non-native writing.

  Aug 5, 2026 · 6 min

• DetectorsFree AI detectors compared: how they differ and how to read them

  Free AI detectors differ in what they measure (perplexity, classifiers, or both), how much text they need, and how they set the 'likely AI' line. This is a guide to those differences and how to interpret a free-tool score, not a ranking by pass rate, because that changes weekly and depends on the text.

  Jul 29, 2026 · 6 min

• ProductHow we price: the economics of a humanizer and what you're paying for

  A humanizer's real costs are the model calls behind each rewrite, the detector scans, and the infrastructure to run them reliably. We price by words because that's what those costs scale with, we don't charge for failed requests, and we don't roll words over. Here's the reasoning, in the open.

  Jul 22, 2026 · 5 min

• WritingTurning AI drafts into publishable blog posts: a workflow

  Draft with AI for speed, then do three things a model can't: add a specific detail only you know, cut the introduction in half, and make every subheading say something. Humanize the prose for rhythm along the way, check the score, and publish something a reader actually finishes.

  Jul 15, 2026 · 6 min

• DetectorsWhat changed in detectors this quarter

  A recurring, dated look at how the major AI detectors changed (model updates, threshold shifts, new false-positive reports) and what it means for writers. Detectors move constantly, so we track the movement here rather than pretending any single result is permanent.

  Jul 8, 2026 · 4 min