ChatGPT GPTZero checker and method

This checker reads submitted text and reports the stylistic regularities associated with model-generated prose: hedging constructions, paragraph-initial connective stacking, tricolon enumeration, unattributed appeals to evidence, and sentence length converging on a mean. Each finding carries the reason it was flagged.

It is independent of GPTZero. We hold no affiliation, make no API call to them, reproduce no score of theirs and predict no verdict. Naming their product in the heading of a page is a description of the search intent, not a claim of correspondence between their output and ours.

The only parameter worth planning is length, because several of the signal classes are undefined on short inputs.

Step 1

Submit between 1 and 12,000 characters of contiguous text. Distributional signals require several hundred words to exist at all.

Step 2

Optionally supply a constraint of up to 300 characters, narrowing the check to a section or a specific concern.

Step 3

Interpret findings against the author's baseline. A pattern present throughout that writer's earlier work is a habit, not evidence.

The perplexity and burstiness account is the most widely repeated explanation of AI detection and it is no longer current. Perplexity described how probable a model would judge the observed word sequence, with GPTZero publishing a threshold above which a document was described as more likely than not human. Burstiness described the variance of that measure across sentences, on the theory that human writing is discontinuous and model output holds a steady level.

GPTZero's own published explanation states that as of autumn 2023 it no longer uses perplexity and burstiness for detection, having migrated to a deep learning architecture, with the statistical measures retained as one of several indicators. The current service outputs a three-way classification, fully human, fully AI or mixed, with sentence-level analysis and confidence scoring, from a classifier described as trained across creative, scientific and news writing. The practical consequence is worth stating: an argument that appeals to a perplexity threshold as the basis of a 2026 flag is appealing to a retired method.

No, and the reason is structural rather than modest. A deep learning classifier's decision boundary is a function of weights that are not published, applied at a threshold that is not published, in a model that is revised without notice. Predicting its output from outside would require replicating all three.

We therefore make no prediction, run no tests against GPTZero, Turnitin, Copyleaks or Originality.ai, and tune nothing against their scoring. This is not a limitation specific to us. Any product claiming a guaranteed verdict from a third-party detector is claiming knowledge of a system it does not operate, and that claim should be read as marketing wherever it appears.

The result set for this query contains the original service alongside multiple distinct domains presenting under the GPTZero or ZeroGPT name. They are separate operators running separate models, and there is no reason to assume any agreement between their outputs.

This has a consequence in a dispute. Where an institution's policy permits detector evidence, it typically names a specific service, which means a result produced elsewhere falls outside the policy that was invoked. A screenshot showing a percentage is not self-identifying, so asking which domain produced it is a reasonable and frequently decisive question. The more general point: a score with no provenance is not evidence, regardless of how confident the interface looks.

The reliability question has a published record. A University of Maryland paper in March 2023 concluded that several AI text detectors are not reliable in practical scenarios. The Washington Post has reported that GPTZero suffers from false positives, cautioning that even small error rates produce wrongly accused students at scale. Futurism, working from GPTZero's own stated error rate, reported that teachers relying on the tool would falsely accuse close to a fifth of innocent students. These are attributed reports about a third party, and they are cited here because a student defending their work is better served by a checkable record than by an opinion.

The underlying reason is a base rate problem. The features these classifiers key on, uniformity of sentence length, formal connectives, predictable structure, describe a very large population of honest writers: students following taught essay form, technical authors writing for clarity, and anyone composing in a second language using the register they were drilled in. When the honest population sharing a feature is large, even a low false positive rate produces a substantial absolute number of wrong accusations.

Stated explicitly, because a page criticising overclaiming should not overclaim.

It computes no perplexity figure and no burstiness figure. It is a server-side classifier that reads text and reports observable stylistic signals with reasons, and it does not output a probability, a percentage or a score of any kind. It has no access to a document's revision history, no knowledge of who typed it and no ground truth to calibrate against, so it cannot establish authorship and does not attempt to. It accepts 1 to 12,000 characters per run with 5 runs per rolling 24 hours, which means it is a review tool for individual pieces rather than an invigilation instrument, and it should not be used as one.

Writers auditing a draft ahead of submission into a process where detection is in use, who want the uniform passages identified while there is still time to edit them. Professionals under a client or employer AI policy doing the same for contractual rather than academic reasons.

Educators, in a narrower and more defensible use: reading what these signals actually consist of before deciding how much weight to give a detector output. A named pattern supports a conversation with a student about their writing. A confidence percentage supports only an accusation, and the documented record above is a reason to be cautious about making one on that basis alone.