DeepSeek GPTZero checker

This DeepSeek GPTZero checker produces a written analysis of the distributional and structural properties of the text you submit: rhythm, lexical range, connective frequency, paragraph templating, explicit signposting. It names the passages carrying those properties. It does not emit a probability, a percentage or a binary classification, because the engine behind it produces prose and a number dressed over prose would be a fabrication of precision.

The route is not GPTZero and carries no affiliation with GPTZero, Turnitin, Copyleaks or Originality.ai. It cannot invoke their models and makes no prediction about their output. Processing is server-side, capped at 12,000 characters per request, with an allowance of five requests per rolling 24-hour window and no account.

Two vendor benchmarks dominate this question, and both are worth reading as self-reported. GPTZero's February 2025 test used 1,000 human and 1,000 AI documents spanning books, articles, blogs, essays and social media, and reports 99.8% accuracy, 99.7% recall and a 0.01% false positive rate, against Originality Lite at 98.9% accuracy, 99.9% recall and 2.2% false positives. Originality.ai's own study, 150 DeepSeek-Chat samples under a confusion-matrix method, reports 99.3% recall for its Lite and Turbo models and 97.3% for GPTZero.

Each vendor's benchmark places its own product first, which is the expected result of a company grading its own exam. The finding that survives both is structural: DeepSeek required no detector retraining to be identified, because models not explicitly optimised against detection converge on the patterns detectors already recognise. Those figures are cited here, not adopted. We have run no benchmark and publish no accuracy rate of our own.

Beyond the character cap, segment at section boundaries and accept that each segment is read without sight of the others.

Step 1

Submit the passage, under 12,000 characters.

Step 2

Read the analysis and locate the passages it names.

Step 3

Revise those passages yourself and re-read the whole text for coherence.

Neither ranking vendor page defines its own terms for a reader. These are the properties in question, and none of them is a hidden marker embedded in the text.

Perplexity

The degree to which each token is unexpected given its predecessors. Generated text systematically selects high-probability continuations, depressing the measure.

Burstiness

Variance in sentence length and complexity across a passage. Human composition is irregular; generated composition regresses toward a mean.

Type-token behaviour

Vocabulary breadth relative to length, and the recurrence rate of a small set of connectives.

Paragraph templating

Recurrence of an identical internal shape: claim, elaboration, example, restatement.

Metatextual signposting

Sentences whose propositional content is a description of the document's own structure.

A 0.01% false positive rate and a 2.2% rate describe different products, and both describe a population rather than a person. Applied across a cohort of several hundred submissions, the second figure produces misclassified human work as a matter of arithmetic rather than misfortune.

The compounding issue is selection. The human writing most likely to be misread is the writing that most resembles the target distribution: formal register, even sentence length, heavy editing. That penalises disciplined prose and non-native writers producing carefully constructed English. Neither vendor page addresses this, and the practical consequence is that provenance evidence, meaning drafts, version history and notes, is worth more than any counter-score.

DeepSeek's reasoning-oriented models fix a structure before surface generation. When the task is expository writing, the plan is realised directly in the prose, producing regular paragraph shape and explicit transition marking. Those are the same properties a classifier reads, so the correlation is not coincidental.

It follows that no removal operation applies. Nothing has been inserted into the text that could be stripped out. The measured quantities are properties of the composition itself, which is why revision is the only intervention with a mechanism behind it.

Stated plainly, because the alternative is implying otherwise by silence.

Any third-party outcome

No access to GPTZero's model or anyone else's, so no prediction of their result is possible or offered.

Authorship

Signals are reported, never proof of who wrote a text or which model produced it.

Short passages

Variance measures need length. Below a few hundred words the reading is weak and should be treated as indicative.

Stability of wording

The analysis is generated prose, so phrasing varies between runs even where the identified signals do not.

Policy

A report says nothing about whether your institution permits AI assistance. That is a separate document and it governs.

Writers with a finished AI-assisted draft and time to revise. Editors diagnosing why submitted copy reads uniformly. Anyone facing an allegation who needs to understand what the allegation is measuring before responding to it.

It is the wrong tool for an unfinished draft, for text too short to measure, and for anyone seeking an assurance about a detector we do not operate.