GPT-5.1 detector, signals only

One request. The sample is posted to our server, passed to a single hosted model under a fixed system instruction, and returned as one block of prose describing the writing properties observed and the reasoning behind each. The analysis branch runs at a low temperature, so repeat runs converge, and the answer is length-capped, which is why a 12,000 character input yields a summary rather than a line-by-line commentary.

There is no classifier weight file here, no perplexity calculation, no corpus of our own. Calling it a statistical detector would be a category error. It is a reading, produced by a model instructed to report signals and limitations and forbidden from asserting authorship.

Knowing that is what makes the output usable. You are reading an argument about a text, and an argument can be evaluated and set aside on grounds you can articulate.

Three steps, of which only the third is difficult.

Step 1

Supply the sample. The server enforces 1 to 12,000 characters after trimming and rejects anything outside that band before a run is spent.

Step 2

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

Step 3

Weigh the findings against your priors: who wrote it, under what brief, in what genre. The findings hold none of that and cannot substitute for it.

Every documented marker of machine prose is a frequency claim, not an identity claim. The word pivotal is not a machine word. It appears more often per thousand words in generated text than in human text, which is a much weaker statement, and it is the strongest one the evidence supports.

Frequency claims compound weakly. A sample carrying six markers is more interesting than one carrying one, but the six do not multiply into certainty, because they are correlated: they arise from the same drafting habits and genre conventions. A house style guide can produce most of them with no model involved.

They are also the easiest thing in writing to remove. Anyone who has read a list of tells can strike them out in ten minutes without touching a sentence of substance, which makes their absence close to uninformative even where their presence is suggestive.

The most specific public catalogue is not a vendor page. Wikipedia's editors maintain a guideline on signs of AI writing, compiled from triaging submitted drafts, and it sorts its markers by kind rather than listing adjectives.

Lexical

Delve, underscore, foster, crucial, intricate, pivotal, tapestry, showcase, and sentence-initial additionally, all overrepresented relative to ordinary prose.

Syntactic

Copula avoidance, meaning serves as, stands as, represents, functions as and boasts in place of is, plus negative parallelism in the form not just X but also Y.

Evidential

Attribution without an attributable party: industry reports, experts argue, observers have cited, none of them named or citable.

Rhetorical

Inflated significance, as in marking a pivotal moment or leaving an indelible mark, and the despite-its-strengths pivot into challenges and future prospects.

Structural

The rule of three, meaning three-item lists imposed on subjects that do not have three parts, and outline-shaped sections with predictable headings.

Typographic

Elevated em dash density, curly quotation marks where straight ones belong, unnecessary title case, boldface scattered through body text, and emoji used as separators.

Version attribution requires a marker present in GPT-5.1 output and absent from its neighbours. Nobody publishes one. Tool pages competing for this term list model names without describing any per-model difference, and the review guides state the limit outright: no detector can prove model origin with certainty.

The refusal is enforced rather than merely professed. The system instruction for this family of routes directs the model never to claim certainty about authorship, plagiarism, or which AI wrote a text, and to describe observable signals and limitations instead. The page targets the term people search; the tool underneath declines the claim the term implies.

The same argument disposes of cross-model attribution. If GPT-5.1 cannot be separated from GPT-5, it cannot be separated from another vendor's model trained on overlapping data and tuned toward the same register.

Two populations absorb most of the error, both predictable from what the markers measure. The first is formulaic human prose. One checker's documentation names the genres directly: academic definitions, legal copy and structured templates. Low lexical variety there is a product of the form, and a marker set that reads variety as a signal will read genre as evidence.

The second is competent non-native English. A narrower active vocabulary and steadier syntax produce the same surface statistics, so the error is not randomly distributed. It concentrates on a group already disadvantaged by assessment, and that asymmetry is the strongest argument against treating any result as dispositive.

The consequence is a rule rather than a caveat: if a flagged sample falls into either population, the flag carries almost no information and should be set down.

The server floor is one character, a validation boundary rather than a recommendation. Most of the catalogue above is distributional: uniform sentence length, repeated transitions and structural symmetry are properties of a passage, not of a sentence, and cannot be observed in a sample too short to hold a distribution.

Checkers that state a figure put the useful minimum around 80 words and the reliable range at 300 or more. We have measured nothing, so treat that as the industry's stated floor rather than ours and read short-sample results with proportionate scepticism.

At the other end the 12,000 character ceiling is enforced server-side, and since the response is capped, a maximal input returns a broader, shallower reading than a chosen 2,000 character excerpt.

No probability, no percentage, no confidence interval, no per-sentence highlighting, no three-way split into written, refined and human. Competing pages offer several of these. This route returns one prose response per call, and putting a number on top of it would only dress the same reading in borrowed precision.

No file ingestion, no batch queue, no API, no account, no stored copy to return to. Usage is counted against a hash of the connection rather than an identity. We hold no affiliation with OpenAI, and nothing here forecasts what a third-party checker will report.

What remains is narrow and defensible: a described reading of a passage, reasoning exposed and limits stated, for a reader who intends to make the judgement themselves.