Skip to content

Detector tell · 4 patterns · weight 4

When the assistant announces its own output

Four rows at weight 4 matching meta-commentary that leaked past an instruction to output only the text: an announcement, a claim of having done the work, or a bold label above it.

Our detector calls this category “GPT-4 labeled output”. The patterns below are read out of functions/api/tools/detect.js when this page is built, and the examples are produced by running the shipped scorer over the 48 licensed samples in our benchmark. If a claim here stops matching the engine, the build fails.

What our own benchmark saw

Did not appear anywhere in the benchmark — no human sample and no AI sample. That is a fact about the benchmark. This label fires on text matching its own pattern; the probe below is run on every build and fails it if it stops.

We are not going to invent an example to fill the gap. What we can show instead is a sentence we wrote ourselves, purely to check that the rule is live.

A constructed example — written by us, not evidence

This sentence was written for this check. It is not from the benchmark, nobody wrote it as real prose, and it says nothing about how anyone writes. Its only job is to prove the rule still fires, and the generator fails the build if it stops.

“Here's the rewritten version of your opening paragraph.”

  • /^here(?:'s| is)\s+(?:the|your|a)\s+(?:rewritten|humanized|revised|updated|edited|polished|corrected)/gim → “Here's the rewritten”

What it means when a person writes this way

A person does not usually preface their own document by announcing that they rewrote it. But a person absolutely writes "## Results" and "**Output**" — and until 2026-08-27 two rows in this label matched exactly those, at the table's top weight. Measured then: a 66-word human engineering note carrying "## Results" and "## Output" scored 62 — "Reads like AI in places" — with three of its seven sentences flagged red. The rows were narrowed to "rewritten" and "humanized" only, the two words in this family nobody uses as an ordinary section heading. The rest of the note is still in the engine source.

The patterns, as the engine holds them

4 of the 70 rows in our phrase table carry this label. They are printed here as they are written, because a paraphrase of a regular expression is a different regular expression.

/^here(?:'s| is)\s+(?:the|your|a)\s+(?:rewritten|humanized|revised|updated|edited|polished|corrected)/gim weight 4

Matches “here's the rewritten”, “here's the humanized”, “here's the revised”, “here's the updated”, “here's the edited”, “here's the polished”, “here's the corrected”, “here's your rewritten”, “here's your humanized”, “here's your revised”, “here's your updated”, “here's your edited”, “here's your polished”, “here's your corrected”, “here's a rewritten”, “here's a humanized” and 26 more combinations , in any capitalisation , and only at the start of a line.

/^i(?:'ve| have)\s+(?:rewritten|humanized|revised|updated|edited|polished|corrected)/gim weight 4

Matches “i've rewritten”, “i've humanized”, “i've revised”, “i've updated”, “i've edited”, “i've polished”, “i've corrected”, “i have rewritten”, “i have humanized”, “i have revised”, “i have updated”, “i have edited”, “i have polished”, “i have corrected” , in any capitalisation , and only at the start of a line.

/^\*\*(?:rewritten|humanized)[^*]*\*\*/gim weight 4

This row uses wildcards, so there is no finite list of phrases to print. Read the expression.

/^#{1,3}\s*(?:rewritten|humanized)\b/gim weight 4

This row uses wildcards, so there is no finite list of phrases to print. Read the expression.

What a hit does to your score

The phrase model is one of four in our ensemble, and its formula is patternScore = min(98, max(5, 10 + 7 x total weight)). One occurrence of the heaviest row in this category takes that model from 10 to 38 out of 100 — and the pattern model carries 40% of the ensemble when the em-dash signal fires and 55% when it does not.

This is the clearest example on the site of what these patterns cost when they are too broad. We accept .docx and .pdf upload, which is to say we accept precisely the documents that carry headings.

Check your own text

Our detector is free and shows you which sentences it flagged and why, including this category. It also gets things wrong, and we publish how often.

Run the detector Our measured error rates If you have been accused Which registers score highest

Other things the detector looks for