GPT-5 detector & writing checker

GPT-5 detector

What the GPT-5 detector does in a working draft

Drop a draft in and the GPT-5 detector comes back with a list of the places it reads as machine-written, with a reason attached to each one. The reasons are the useful part, because a reason tells you what edit to make. A bare verdict tells you nothing you can act on at eleven at night before a deadline.

Treat it as the last pass before something leaves your hands: after the writing is done, before the send button. It is a reading tool for your own work, not a judgement on anybody else's.

Workflow

How to run the GPT-5 detector on a draft in three steps

The whole loop takes about a minute, and you can run it again after editing.

Step 1

Paste the draft, or the section you are least sure about. Up to 12,000 characters per run.

Step 2

Run the check. The text goes to our server, a model reads it, and the findings come back as a list.

Step 3

Work down the list. Fix the flags you agree with, ignore the ones that misread your draft, then re-run if you changed a lot.

Triage

What to do with each GPT-5 flag

Not every flag deserves an edit. Sorting them quickly is most of the value, and after a few drafts the sort becomes automatic.

Rhythm flags

Paragraphs all landing at the same length. Merge two, cut one to a single line, and the pattern breaks without losing a word of meaning.

Hedge flags

It is worth noting, it is important to remember. Delete the wrapper and keep the claim. The sentence gets shorter and firmer.

List-of-three flags

Ask whether the third item is load-bearing. If it was added for balance, cut it.

Closing flags

A final paragraph that restates the piece. Replace it with the thing you actually want the reader to do or think.

Flags you disagree with

Read the reason. If it does not describe your sentence, leave the sentence alone. The check has no standing to overrule you.

The model

How GPT-5 drafts differ from GPT-4 drafts

If your habit of spotting AI text was trained on GPT-4, it is partly out of date. Published comparisons describe GPT-5 writing in a looser, more conversational register, dropping much of the formal scaffolding that made earlier output obvious on sight. A GPT-5 draft can read fine and still be full of structural tells.

The tells that persist are habits rather than tone. Testing published by Originality.ai on GPT-5 rewrites found em dashes in most of the machine-written samples and none in the human control set, and found the model matching the readability of its input more closely than earlier versions did. Matching the input is exactly why a quick read-through stops being enough.

Straight answer

Can this tell GPT-5 apart from other models?

No, and it will not pretend otherwise. The model running the check is instructed never to assert which system wrote a piece of text. Product names across this category imply per-model identification, and nobody has shown it working.

In practice that rarely matters for the job you are doing. If you drafted with GPT-5, the findings will describe your draft. If you are checking someone else's work, the honest reading is that the text carries these patterns, not that a named model produced it.

Why flags happen

Why a human draft gets flagged as GPT-5

The check reads tidiness. Even paragraph lengths, level tone, no odd turns. Plenty of people write like that: anyone trained in technical writing, anyone working to a style guide, and very often anyone writing in English as a second or third language, where a safe register is the sensible one to hold.

So a flag is a description of the prose, never a claim about the person. If you are on the receiving end of one, the fix is not to argue about scores. It is to show the work: drafts, notes, edit history, the version of the paragraph you threw away.

Allowance

Checks per day and the 12,000 character limit

You get 5 checks per rolling 24 hours, tracked per browser, with no account and no e-mail. The window rolls rather than resetting at midnight, so the checks come back as the oldest ones age out.

Each run takes 1 to 12,000 characters. For a long piece, check the sections that matter rather than spending runs on the whole thing: the opening, the conclusion, and any section you know was drafted fastest. That is where the patterns cluster anyway.

Who it is for

Who runs the GPT-5 detector, and when

Students running a final pass on coursework. Freelancers checking a delivery before a client reads it. Content leads sanity-checking a batch of drafts. Job seekers who used an assistant on a cover letter and want it to sound like them again.

The timing is the same in every case: late, after the argument is settled and before anyone else sees it. Run it earlier and you will spend runs on sentences you were going to rewrite anyway.