GPT-5 detector

The GPT-5 detector takes a block of text, sends it to a model on our server, and returns a description of the writing patterns present in it: distribution of sentence length, density of hedging constructions, structural symmetry, the shape of the closing paragraph. Every observation carries its justification.

The output is an analysis of prose, not an identification of an author. That distinction is the whole design. A tool that reports what it can observe remains useful when it is wrong, because you can see where it went wrong. A tool that reports a verdict does not.

The procedure is short, and the interesting part is what you do with the result.

Step 1

Supply the sample. The endpoint accepts 1 to 12,000 characters of text in a single run.

Step 2

Run the analysis. The text is posted to our server, read by a hosted model, and the findings are returned.

Step 3

Interpret the findings against what you already know about the text's origin. A signal confirms or contradicts a hypothesis; on its own it establishes nothing.

Several established detectors describe their method in statistical terms: perplexity, meaning how predictable each token is under a reference model, and burstiness, meaning how much sentence-level variance the text shows. Those are trained classifiers producing a scalar, and the scalar is then presented as a probability of machine authorship.

This route works differently and it is worth being precise about it. There is no trained detection classifier and no perplexity computation behind it. A language model reads the sample under a system instruction that requires it to describe observable signals and forbids it from claiming certainty about authorship, plagiarism or which model produced the text. The result is qualitative by construction: it explains why a passage reads as machine-written rather than scoring it.

Two findings from published GPT-5 testing are worth carrying here because they are specific enough to check. Originality.ai, testing 100 GPT-5 rewrites, reported em dashes present in most of the machine-written set and absent from the human control set. The same work reported GPT-5 tracking the readability of its source text more closely than earlier models, with human inputs averaging a Flesch-Kincaid figure in the high forties and the GPT-5 rewrites landing a few points below.

The second of those matters more than the first. A model that mirrors its input's readability defeats register-based intuition: the output does not feel stiff, because it was never aiming at stiff. What survives is habit at the structural level, and structure is what this check reads.

Register

Looser and more conversational than GPT-4, per side-by-side comparisons. Less useful as a tell than it was.

Punctuation

The em dash habit persists across the version change.

Readability matching

Output tracks the input's reading level, which suppresses the obvious formality signal.

Structure

Symmetrical construction and balanced enumeration remain, and remain readable in aggregate.

No published work demonstrates reliable attribution of text to a named model, and the research that touches the question points the other way: detectors trained on one model's output generalise poorly to another's, which is the opposite of what per-model attribution would require.

The category nonetheless markets itself on attribution, because the model name is the search term. We name GPT-5 in this page's title for the same reason, then say plainly that the check reports patterns common in GPT-5 output rather than identifying its author. The instruction given to the model enforces that, so it is not merely a disclaimer at the bottom of the page.

Competing pages publish figures: 96.5 per cent, 99 per cent, 99.8 per cent. Each is measured on a sample set the vendor assembled, so none is comparable to another or verifiable by a reader.

We have run no such evaluation, so there is no figure we could print that would mean anything. Printing one anyway would be the single easiest way to make this page rank and the single fastest way to make it dishonest. What the page offers instead is a description of the method, the instruction the model runs under, and the conditions in which it is known to be wrong.

The failure modes are structural rather than random, which means they can be listed. Prose edited to a house style loses the variance that reads as human. Second-language English writing frequently holds a level, cautious register that produces the same profile, a bias significant enough that at least one competitor publishes a target false-positive rate specifically for it. Short samples give the check too little to read. Heavily edited machine text passes, because the text genuinely has been rewritten by a person.

None of that is repairable by tuning, because the evidence is genuinely ambiguous at the level where these tools operate. It is repairable at the level of use: a signal is a reason to look at process, drafts and history, and it is not admissible as a finding about a person.

Each run accepts 1 to 12,000 characters, enforced at the endpoint, and the allowance is 5 runs per rolling 24 hours per browser with no account. The ceiling is roughly 1,800 to 2,000 words of English prose. Competitors publish larger ceilings, 20,000 characters in one case and 10,000 before an account is required in another, so the comparison is easy to make and we would rather state ours than bury it.

The allowance exists because each run is a model call with a real cost. Rolling means the window moves continuously, so runs return as the oldest expire. There is no upgrade path being withheld here: this is what the tool is.