Deep checks, and what they send

A deep check also looks for references we could not find in Crossref or OpenAlex, unusual stock phrases, pairs of numbers, names or dates worth a second look, and up to 8 dates and numbers checked against Wikipedia. It shows the sentences it found as notes for a person to check. It never marks a text as AI-written.

A deep check sends up to 8 sentences from this text to Google to check dates and numbers against Wikipedia.

There is no separate consent step. Gemini, on Google Cloud in our own project, judges each sampled sentence against the passages found for it.

Search words drawn from those sentences go to Wikipedia, run by the Wikimedia Foundation. A sentence written in the first person is never looked up, and nothing containing a digit, an email address or a web address goes into a search.

The titles in a reference list are looked up in Crossref and OpenAlex, two public indexes of published work.

A deep check runs only on a text that is still a candidate for the mark. When the standard detector did not read the text as human, nothing is sent anywhere.

Only a date or a number is ever marked as different from a Wikipedia page, and only when the page really says something else. Names, places and titles are never marked.

A finding shows the sentence it came from and what was looked up. A reference we could not find is not proved to be fake: Crossref and OpenAlex can miss a real work. A pair of numbers, names or dates worth a second look is two mentions that differ, often because they describe different things. A fact that could not be checked is not a fact that was confirmed.

A lookup service can be unavailable. When it is, a reference reads as not checked, not as not found.

The checks mark things on honest writing too. When a reviewer looked at what they had marked in published papers, most of the marked references turned out to exist and none of the marked pairs was a real contradiction. The model card has the counts.

Today the findings are shown as notes. They do not change the mark, and none of them marks a text as AI-written.

Free: 1 a month. Education: 15 a month per teacher. Pro: 50 a month. Max: 250 a month.

A deep check uses one from the allowance when any part of it gives a result, even if a lookup service did not answer for the rest. One that is skipped, or that could check nothing at all, costs nothing.