GPT-5.2 detector
What the GPT-5.2 detector does
Paste a passage and the GPT-5.2 detector sends it to our server, where one hosted model reads it and describes the writing signals it finds, with a reason attached to each. You get sentences back, not a dial or a percentage.
Five free checks in any rolling 24 hours. The box takes 1 to 12,000 characters, and anything outside that is refused before a check is counted. No account, no card, no file upload.
It does not tell you a passage came from GPT-5.2 rather than GPT-5.1 or a person having a bad writing day. The instruction the check runs under says to describe observable signals and their limits and never to claim which model wrote a text.
How it works
How to run the GPT-5.2 detector in three steps
Nothing to configure. The judgement stays with you.
- Step 1
- Paste the passage you want read. A few paragraphs works better than a single line, because most signals only exist across sentences.
- Step 2
- Press the check button. The text goes to our server once and is read by one model, spending one of your five daily checks.
- Step 3
- Read the findings with the reasoning attached, then weigh them against what you already know about the text.
The tell that stopped working
Why bullet points no longer mean AI
For two years the quickest way to spot generated text was the formatting: bolded subheads, bullet lists, everything chunked. Anyone checking GPT-5.2 output for that is checking for the wrong thing.
A published side-by-side writing test of the two most recent versions found the opposite behaviour. Asked to summarise dense material, the older model broke the copy up with bolded subheads and bullet points, while GPT-5.2 returned what the testers called much less digestible, solid blocks of text. Across eight tasks the older model scored 60 out of 80 and GPT-5.2 scored 52.
So a wall of unbroken prose is no longer evidence of a human hand, and a tidy bulleted summary is no longer evidence of a machine. Both are formatting habits, and the habits moved.
Read it yourself
Signals worth reading by eye
These come from published writing tests and editing guidance rather than from any tool, which means you can check every one of them without spending a run.
- Title Case In Headings
- The testers docked a point for it outright and called it one of the most obvious signs of AI-generated copy.
- Solid blocks
- Long undivided paragraphs where the material clearly has parts. Newer output tends this way rather than toward lists.
- Flat register where warmth belongs
- On a price increase email, both versions produced unsoftened formal phrasing. A message that should apologise and simply announces is worth a second look.
- Even sentence rhythm
- A passage where every sentence arrives at the same length, with no short one to break the pattern.
- Empty transitions
- Furthermore, moreover and in conclusion carrying no argumentative weight where they appear.
The honest answer
Can a GPT-5.2 detector identify the version?
No. Naming a version would need that version to leave a mark in ordinary prose that its neighbours do not leave, and nobody has published one. Our check refuses it at the instruction level, not out of politeness.
There is a related trap worth naming. The coverage of GPT-5.2 says its writing got worse, and OpenAI's own chief executive said so publicly, telling a developer town hall in January 2026: I think we just screwed that up. He explained the trade, saying most of the effort in 5.2 went into intelligence, reasoning, coding and engineering.
That is useful context, and it is still not a detection method. Blocky, flat, title-cased prose is also produced by tired people writing to a deadline. A trait being common in a model's output does not make it evidence about a particular passage.
Getting it wrong
Which writing gets flagged wrongly
Short, generic and essay-shaped text is the most likely to be misread, and that is not our observation, it is the caveat published by a detector vendor claiming a 97 per cent detection rate on its own page.
The same vendor states the rule this page follows: no current AI content detector should be used as a standalone check. Formulaic human writing, a methods section, a policy summary, a set of product descriptions written to one template, will read as low variety because the form demands it.
Writing by non-native English speakers runs into the same trap. Steadier construction and a narrower working vocabulary produce the pattern a checker reads as machine output.
Limits
What the free GPT-5.2 checker will not do
No percentage, no confidence score, no per-sentence highlighting. One run is one model call and one written answer. Other tools present a probability and colour the sentences; this one does not, and putting a number on a reading would be borrowing precision we do not have.
No file upload, no batch mode, no API, no account. Five runs per rolling 24 hours, counted from when each was used rather than reset at midnight, and 1 to 12,000 characters per run.
No affiliation with OpenAI, and no prediction of what any other checker will say about the same words.
In use
Who runs a GPT-5.2 check, and when
Usually someone holding a piece of writing that does not sound like the person who sent it: a marker with a submission, an editor with a commission, a manager with a report that reads like a press release.
The check gives you language for the thing you already noticed. It does not give you proof, and a finding should open a conversation with the writer rather than settle one.