GPT-5.2 detector, signals only

One request, one completion. The sample is posted to our server, read by a hosted model under a fixed system instruction, and returned as prose describing the writing properties present and the reasoning behind each. The analysis branch runs at a low temperature, so repeat runs converge rather than wander.

No weights are trained here, no perplexity is computed, no token-level probabilities are assigned. Pages that rank alongside this one describe exactly those mechanisms in their own products, and borrowing the vocabulary to describe ours would be a misrepresentation.

What this buys is inspectable reasoning. A score cannot be argued with. A described reading can be tested against the text and rejected on grounds you can state.

Three steps, and only the last one requires judgement.

Step 1

Submit the sample within the 1 to 12,000 character band the server enforces after trimming.

Step 2

Run the analysis: one call, one fixed instruction, one of five runs in the rolling 24 hour window.

Step 3

Weigh the findings against the provenance you already have. The reading contains no information about who wrote the text or under what conditions.

Heuristics decay, and this one decayed publicly. Bulleted structure and bolded subheads were the fastest visual cue for generated text. A published eight-task writing comparison of the two most recent versions found the reverse behaviour: asked to summarise dense material, the earlier version broke the copy into bolded subheads and bullet points, while GPT-5.2 returned what the testers described as much less digestible, solid blocks of text.

The aggregate scores moved the same way, 60 out of 80 for the earlier version against 52 for GPT-5.2, which makes this a measured difference in one published test rather than a general law. It is enough to retire the heuristic and not enough to invert it.

The correct conclusion is narrow. Formatting reflects a release's defaults, defaults change between releases, and any signal that tracks a default has a short life. Signals that track how sentences are built survive longer, which is why the reading this route returns is weighted toward structure rather than layout.

These are drawn from published writing tests and editing guidance rather than from any detector's marketing, and every one of them can be checked by eye before a run is spent.

Title case in headings

Docked a point explicitly in the published test and named there as one of the most obvious signs of AI-generated copy.

Register failure under social pressure

On a subscription price increase email, neither recent version attempted warmth, both defaulting to flat announcement. Where a human would soften, the output states.

Uniform sentence length

A passage with no short sentence anywhere in it, which is a distributional property and cannot be seen in a single line.

Empty transitions

Furthermore, moreover and in conclusion where the logical relation they assert is not present in the argument.

Stakeless examples

Illustrations general enough to attach to any subject, which survive rewriting because nothing in them is specific enough to break.

Attribution requires a marker present in one version's output and absent from its neighbours. The result set contains no such marker. Tool pages list model names without describing per-model differences, and the one genuine comparison in the set measures quality rather than identifying signatures.

There is a tempting inference available here and it should be refused explicitly. OpenAI's chief executive publicly conceded the writing weakness, telling a developer town hall in January 2026: I think we just screwed that up, and adding that most of the effort in 5.2 went into intelligence, reasoning, coding and engineering. It does not follow that weak prose identifies GPT-5.2. Humans produce blocky, flat, title-cased writing in quantity, and a trait common in a population is not evidence about a member of it.

The refusal is also enforced rather than professed. The system instruction for this family directs the model never to claim certainty about authorship or which AI wrote a text.

A detector vendor advertising a 97 per cent detection rate and a sub-3 per cent false positive rate publishes the qualification itself: short, generic and essay-like text is the most likely to be flagged wrongly. That is a useful admission, because it names the population rather than gesturing at uncertainty.

Two more follow from what the markers measure. Formulaic genres, methods sections, policy summaries, legal boilerplate, produce low lexical variety by construction. Competent non-native English produces steadier syntax and a narrower active vocabulary. Neither is evidence of anything except the constraints the writer was working under.

The same vendor states the operating rule directly: no current AI content detector should be used as a standalone check. We agree, and this page is written on that basis rather than around it.

Most markers above are distributional. Sentence-length uniformity, structural symmetry and transition frequency are properties of a passage, so a sample too short to contain a distribution cannot exhibit them. The server floor of one character is a validation boundary, not a recommendation.

At the top end, 12,000 characters is accepted but the response carries its own length cap, so a maximal submission returns a broader and shallower reading than a chosen excerpt of a few hundred words.

The output contains no probability, no confidence interval, no highlighted spans and no three-way classification. Competing tools offer all four. Adding a number to a prose reading would dress it in precision it does not have.

No file ingestion, no batch queue, no API, no account, no stored copy. Usage is counted against a hash of the connection over a rolling 24 hour window, and five runs is the whole allowance.

No affiliation with OpenAI is held or implied, and no output here forecasts what a third-party checker will report about the same text. What remains is a described reading of a passage, with its reasoning exposed and its limits stated, for a reader who intends to judge for themselves.