GPT-5.1 detector
What the GPT-5.1 detector does in a review
A passage arrives that does not sound like the person who sent it. The GPT-5.1 detector is the step between that feeling and doing something about it: you paste the sample, our server reads it with one hosted model, and you get back written findings describing the signals present and why each one was noted.
It is a server route, so the text does leave your device. Five runs are free in any rolling 24 hours, the input window is 1 to 12,000 characters, and there is no account to create before the first run.
The output is a paragraph of reasoning, not a dashboard. That matters for a workflow, because a sentence you can quote in a note to a colleague is more useful than a number you would have to explain.
The three steps
How to run a GPT-5.1 check in three steps
The run takes seconds. Picking the right sample is the part that needs thought.
- Step 1
- Choose your sample and paste it. Pick a stretch that is representative rather than the one paragraph that annoyed you, because one paragraph is a thin basis for anything.
- Step 2
- Run the check. The text is posted to our server once and read by a single hosted model, and one of your five daily runs is spent.
- Step 3
- Read the findings, then compare them against the writer's earlier work and the draft history before you act on anything.
After the run
What to record after a GPT-5.1 check
Published guidance for reviewers is consistent on this point: a detector result is evidence to be documented alongside other evidence, not a finding on its own. If a check ever becomes part of a conversation with a writer or a manager, these are the things you will wish you had written down.
- The sample
- Which passage you actually ran, and how long it was. A result on 200 words and a result on 2,000 words are not comparable.
- The date
- When you ran it. Models and checkers both change, and a result is a reading taken on a day.
- The findings verbatim
- Copy the wording the check returned rather than your summary of it, so the reasoning survives being retold.
- The context you already had
- Earlier drafts, the writer's previous work, the brief they were given. This is what turns a signal into or out of a concern.
- What you did not do
- Note that no version or authorship claim was made, because the check does not make one.
Before you run anything
The AI writing signals you can spot yourself
Wikipedia's editors publish a guideline on signs of AI writing, assembled from cleaning up submitted drafts at scale. Half the value of running any check is knowing what it is looking at, so this list is worth a slow read once.
- House vocabulary
- Delve, underscore, foster, crucial, intricate, pivotal, tapestry, showcase, and paragraphs that open with additionally.
- Anything but the word is
- Serves as, stands as, represents, functions as, boasts. A plain is would have been shorter every time.
- The not-just construction
- Not just X, but also Y. Not X, but Y. The shape does a lot of work in generated prose.
- Attribution with nobody in it
- Industry reports, experts argue, observers have cited, none of them ever named.
- The challenges pivot
- Despite its strengths, it faces several challenges, followed by speculation about the future. It is close to a reflex.
- Threes
- Three-item lists where the subject does not have three parts, alongside heavy bold, curly quotes and a high em dash count.
The limit that shapes the workflow
Why the GPT-5.1 detector cannot name a version
Plan your process around this one, because it will come up the first time someone pushes back. The check cannot say that GPT-5.1 wrote a passage rather than GPT-5, another model, or a careful writer with a house style. Naming a version would require that version to leave a fingerprint in ordinary prose, and no such fingerprint has been published.
The refusal is built in. The instruction behind this family of routes tells the model never to claim certainty about authorship or which AI produced a text, and to describe observable signals and their limitations instead. You cannot prompt your way past it and neither can we.
In practice this makes the workflow simpler, not harder. You are gathering signals to discuss with a person, and a tool that overclaims would only make that conversation worse.
Know the failure mode
The submissions that flag for the wrong reason
Formulaic human writing is the classic false positive, and one checker's own documentation names the genres: academic definitions, legal copy and structured templates. A methods section or a policy summary is repetitive because the form requires it.
Non-native English writing is the second known population. Steadier construction and a narrower working vocabulary produce the same surface pattern that a checker reads as machine output.
Build both into the process: if a flagged sample belongs to either category, that alone is reason to set the result aside rather than weight it.
Working inside the limits
Five runs and 12,000 characters a day
The allowance is 5 runs per rolling 24 hours, counted from the moment each one is used rather than reset overnight. If you have a stack of submissions, that budget is the real constraint on the workflow, so spend runs on the ones your own reading has already flagged.
Each run takes 1 to 12,000 characters, about two thousand words. A long dissertation goes through as sections, and in practice a representative section answers the question the whole document would.
There is no batch mode, no upload, no API and no queue. One passage, one run, one written answer.
In use
Where a GPT-5.1 detector fits, and where it does not
It fits at the point where you need language for a concern you already have: a marker preparing to ask a student about their process, an editor deciding whether to send a commission back, a hiring manager comparing a cover letter against an interview transcript.
It does not fit as a gate. No result from this or any checker is proof of authorship, we have no affiliation with OpenAI, and nothing here predicts what a different checker would say about the same words. Use it to decide what to ask, not what to conclude.