ScribeLens checks text for the writing patterns common in AI-assisted drafting, including work drafted with Gemini inside everyday tools like Gmail and Google Docs, then shows the specific sentences and signals behind the result rather than a single opaque percentage.
Gemini is often not something people sit down to open in its own tab the way they might open a dedicated chat window. It shows up embedded inside tools people already write in — a Gmail draft, a suggestion inside Google Docs, a quick answer pulled up on a phone — which means a lot of Gemini-assisted writing gets accepted and turned in without the writer ever thinking of it as “using AI.”
That embedded, casual path changes what a submission actually looks like. Someone who accepts a few suggestions inside a longer essay or email isn't turning in a document generated end-to-end by an assistant — they're turning in their own draft with AI-assisted passages mixed in, which is exactly the kind of partial, uneven case a single score handles badly and sentence-level evidence handles well.
ScribeLens cannot tell you that a specific passage came out of Gemini rather than another assistant — no legitimate detector can. What can be measured is whether writing shows the statistical patterns common to AI-generated text generally: patterns that come from how these models generate language at all, not from one product's branding or where it happens to be embedded.
That distinction matters especially here, because so much of Gemini's use is embedded rather than standalone. A tool built only to recognize one assistant's known interface would miss most of that writing, since the same underlying patterns show up whether a passage came from a dedicated chat window or a one-line suggestion accepted inside Docs. Detection built around the shared mechanism catches both; detection built around one interface would not.
Writing that blends a person's own sentences with a few accepted AI suggestions produces a different pattern than a document drafted end-to-end by an assistant: flags concentrated in specific sentences or paragraphs, rather than spread evenly across the whole piece. That's a more informative signal on this kind of writing than a bare score, and it's exactly what sentence-level evidence is built to show.
Educators reading a Docs-drafted assignment, teams reviewing Workspace-assisted internal writing, and anyone who accepted a few suggestions and wants to see how their own draft reads before sending or submitting it are all looking at the same kind of case: a blend, not a binary. A single number collapses that blend into one figure; reason keywords attached to each flagged sentence do not.
Often, yes. A passage drafted with “Help me write” and accepted with only light edits tends to carry the same writing patterns ScribeLens looks for — even sentence rhythm, predictable word choice, formulaic structure — no matter which assistant produced it. Editing reduces those patterns; accepting a suggestion and moving on does not. The more reliable approach is to check your own draft before you submit it and revise any flagged sentences for genuine clarity, rather than assume a light edit is enough.
No. Detection reads for patterns common across generative models broadly — Gemini and other assistants tend to produce similar statistical tendencies at the sentence level, because those patterns come from how language models generate text at all, not from one product's design. A result tells you whether a passage shows AI-like patterns, not which product produced it.
Yes, as process evidence. Version history shows how a document was built up over time — drafts, edits, revisions — which is a useful companion to a ScribeLens report, not a replacement for it. The report shows what the current text looks like at the sentence level; version history shows how it got there. Bringing both to a conversation is stronger than either alone.
Yes, in the sense that matters. ScribeLens targets the statistical patterns produced by how language models generate text at all — not the specific quirks of one version of one assistant. An update changes fluency and phrasing; it doesn't change the underlying mechanism detection looks for, so a new release isn't a reset button for a result.
Possibly, and that's exactly the case sentence-level evidence is built for. If AI-assisted passages sit inside an otherwise human-written document, flags should concentrate around those specific sentences rather than appear evenly across the whole piece. Read the sentence-level view first, not just the overall verdict, to see where a signal is actually coming from.
Short passages give pattern analysis less to work with than a full essay or report, so treat a result on a very brief reply as a lighter signal. For anything you're planning to submit or send in a context where it matters, checking the longer draft or document it's part of produces a more reliable read.
Paste a draft — including passages drafted with Gemini inside Gmail or Google Docs — and see the sentence-level evidence behind the result. Free, no signup required.