Perplexity GPTZero checker and its limits

This tool sends your text to our server, which analyses it under a fixed instruction and returns prose. The instruction is worth reading before you interpret anything it says: analyse only the supplied text for the qualities implied by the tool's name, give a concise result with clear findings, and never claim certainty about authorship, plagiarism or which model wrote the text, describing observable signals and limitations instead.

Every boundary on this page follows from that. The output is a description of properties present in a document. It is not a classification of the document, and it is not a forecast of how any third party will classify it.

The phrase joins two unrelated referents, and conflating them produces most of the bad advice in this category. Perplexity is a commercial answer engine. Perplexity is also an information-theoretic measure of how predictable a sequence of tokens is, and it is the measure GPTZero was originally built on.

GPTZero's own explainer defines its version as how likely an AI model would have chosen the exact same set of words found in the document, and defines burstiness as how much that measure varies across the document, with human writing varying more. Neither definition has any bearing on the answer engine. A page that answers the metric question when the reader asked the product question, or the reverse, has answered nothing.

A large share of the writing about GPTZero explains it in terms of perplexity and burstiness. GPTZero states on its own pages that as of autumn 2023 it no longer uses perplexity and burstiness for AI detection, having migrated to a deep-learning architecture, and that the pair now function as one of seven indicators within that model alongside text search and deep learning.

Its support documentation goes further in tempering them, describing the metrics as one part of the whole story and one element of several. GPTZero also notes that the retired approach remains the basis of dozens of other detector applications, which is the useful part of the disclosure: a third-party tool quoting you a perplexity figure may be running the method the vendor itself moved on from.

The procedure is trivial. The interpretation afterwards is where the work is, which is why step 3 is written as a constraint rather than an action.

Step 1

Submit 1 to 12,000 characters. The ceiling is enforced in the request handler before anything is dispatched, so an oversized submission is refused rather than silently cut.

Step 2

Optionally supply up to 300 characters of instruction to scope the analysis, for example to a named section rather than the whole document.

Step 3

Read each finding as a statement about the text and check it against the passage it describes. A finding you cannot locate is not a finding you can act on.

Every property these tools measure is a property of the writing, and writing properties are not exclusive to their most common cause. Even sentence lengths, level register, transitions that always arrive, word choices that are consistently the expected one: those describe a great deal of machine output and they also describe a competent second-language writer, a technical author following a house style, and anyone writing to a template.

This is the structural reason detection is contested rather than a matter of tuning. A classifier can report that a document has the properties associated with generated text. It cannot report where the document came from, because the properties do not carry that information. GPTZero publishes a false positive rate below 1 percent, and experienced users publicly dispute it. We have tested neither claim and cite neither as fact, which is itself the position: the number is not ours to assert.

Perplexity is not computable by a reader, and a figure you cannot reproduce is not much use for editing. Its correlates are visible, and unlike the number they survive a vendor changing models.

Uniform sentence length

Variance across a paragraph, not average length, is the property. A page of sentences all between eighteen and twenty-four words reads level regardless of their quality.

Complete syntactic execution

Human drafts contain sentences that change direction partway through and get finished anyway. Their absence across a long document is conspicuous.

Invariant register

Formality that does not drift between the argument and the aside. People get blunter when they care about something.

Obligatory connectives

A transition at the head of every paragraph, supplying continuity that the content should already provide.

Modal word choice

Vocabulary sitting consistently at the most probable acceptable option, never at the fifth, which is where specificity usually lives.

No, and the tools in this result set that say otherwise are describing something they cannot be doing. One competitor states plainly that it uses models which mimic the exact detection mechanics of GPTZero and Copyleaks, and reports an AI probability accordingly. Those detection mechanics are proprietary. A number generated by an unrelated model is not the number the detector will produce, whatever it is labelled.

We have no affiliation with GPTZero, Turnitin, Copyleaks or Originality.ai, we run no tests against them, and we publish no figure that implies we do. If you need a GPTZero result, the only legitimate source of one is GPTZero.

Listed plainly, including what is absent, because a reader comparing this page against a competitor's feature grid deserves the comparison to be honest in both directions.

No score

No percentage, no confidence value, no threshold. Nothing in this codebase computes one and none is displayed.

No document highlighting

The result is a read-only text area containing prose. There is no red and green sentence map on this route.

No model attribution

The runtime is instructed never to claim which model produced a text. Nor will it claim a text is human.

12,000 characters

Per run, enforced server-side. Roughly 1,800 to 2,000 words of English prose.

5 runs per rolling 24 hours

Counted per visitor by our server, rolling from the moment of each run rather than resetting on a calendar boundary.

Text is transmitted

This is a remote analyser. Your text goes to our AI server, which is how the analysis happens rather than a caveat about it.