The detector does not change with the kind of text you paste: the same frozen V3 engine, English only, 100 words minimum. What changes with content type is the risk of a wrong signal and how to read the result. Short, templated, formulaic writing sits closest to the patterns the classifier associates with AI text, so honest human writing of that kind is flagged more often. Here is what to expect for twelve common types.
Guidance by content type
Three types have a full guide of their own. The other nine are covered here and nowhere else — this card is the whole treatment.
Essays
Short answers land in the 100–149-word band, where about 2.4 in every 100 human texts drew a wrong AI signal in our sealed test. Draft history is better evidence than any score. Full guide →
Research papers
Abstracts and methods sections are formulaic by convention, so they carry the highest false-flag risk in a paper; discussion sections much less. Non-native academic English is the main false-positive group. Full guide →
Business reports
Template-driven and recurring reports repeat their own structure, which reads as uniform prose. Strip tables and figures before pasting, and check whether the document is yours to submit. Full guide →
Product descriptions
Short, templated and specification-led: the highest false-flag risk of any type here. Paste 150 words or more so you leave the strictest length band, and read several items from the same catalogue before you conclude anything about one of them.
Marketing copy
Benefit statements, repeated value propositions and campaign phrasing are formulaic by design, and a good copywriter produces the same uniformity a model does. Expect Uncertain often; treat a strong signal on a long-standing brand line with real suspicion.
Website copy
Homepages, pricing and feature pages are written to a house pattern and edited repeatedly, which flattens exactly the variation the engine measures. Check the body prose of one page at a time, not headings, nav labels or button text.
Blog posts
Usually long enough for the most reliable length bands, which makes them one of the better cases. Paste continuous prose rather than a post full of subheadings and lists, and check a partly edited draft in the sections that were actually rewritten.
Articles
Editorially polished prose in a neutral register sits close to the patterns the classifier learned. A professional edit removes idiosyncrasy, so the score partly reflects the editing rather than the drafting.
News articles
Wire-style prose is deliberately standardised — attribution, inverted pyramid, neutral tone — and reads formal to the engine. Expect Uncertain on straight reporting, and do not read that as a finding.
Emails
One message is nearly always too short and too polite to say anything about. Check a whole thread if you check anything, and only your own drafts or messages shared with you for that purpose — never someone else’s correspondence.
Social media posts
Almost always under the 100-word floor, so the tool refuses them — and where a post is long enough, captions and hooks are formulaic. This is the wrong instrument for social content; use it on the long-form piece the post links to instead.
Resumes and cover letters
Formulaic by nature, short, and about a named person. Review your own application before you send it if you find that useful; do not run a detector on somebody else’s application, and never let a score influence a hiring decision.
How to read a result for any type
Every result carries the same five things: a 0-100 AI-writing signal score; a score stability range, showing how far that number moves when the source families behind the mapping are resampled; one of three bands — Likely human-written, Uncertain or Likely AI-written — decided by frozen thresholds, not by the displayed number; the length band; and the measured false-flag rate for that length (about 2.4 in every 100 at 100–149 words, about 2 in every 1,000 at 150 words or more). A score is a review signal, not proof, and never enough on its own for a decision about a person. What an AI detector score means →
Who this is for
Teachers deciding whether a submission is worth a conversation, editors triaging drafts they were asked to review, site owners auditing their own pages, and writers self-reviewing before submitting. In each case the check is a starting point for reading the text yourself. If your own writing was flagged, what to do if your text is flagged sets out the next steps.
Content-type FAQ
Does the detector work differently for different kinds of text?
No. The same frozen V3 engine scores every submission of 100 English words or more. What changes with content type is the risk of a wrong signal and how the result should be read.
Which content types are flagged wrongly most often?
Short, templated and formulaic writing: product descriptions, marketing copy, resumes and cover letters, and any text under 150 words. At 100 to 149 words about 2.4 in every 100 human texts drew a wrong AI signal in our sealed test, against about 2 in every 1,000 at 150 words or more.
Why can I not check a social media post?
Most posts are shorter than the 100-word minimum, so the tool returns “More text is needed for a meaningful analysis” rather than a result. Below that length there is not enough writing to carry a signal, and a result would be misleading rather than merely weak.
Should I check an email someone sent me?
No. This tool is for your own drafts, or text shared with you for review with the writer’s knowledge. One email is also nearly always too short to say anything about.
Can I use a result to decide about a job applicant or a student?
No. A score is a review signal, not proof, and never enough on its own for a decision about a person. Use it to decide whether a text is worth reading closely, then read it, ask about the drafting, and look at the revision history.
What should I do with an Uncertain result?
Treat it as “no useful answer” rather than a weak accusation. Human and AI writing overlap in that band, and edited or partly drafted text usually lands there. Formal, templated and second-language English land there often too.
Do the nine types without their own page get a shorter answer?
They get the whole answer. Those nine had separate pages that repeated the same generic material; the card here is the type-specific part, and what they shared is in “How to read a result for any type”.
Does it work on text that is not in English?
The form does not block it, but the result is unsupported: the current detector has been validated on English only and its behaviour on other languages is unknown. Route non-English material to human review.