What GPTZero results mean, and how to review them

GPTZero describes an end-to-end deep-learning system trained on human and generated text, with analysis at document and sentence level. Its technology page explains a three-class approach, human, mixed and AI, and distinguishes classification confidence from the proportion of generated text.

The vendor also describes ongoing model development and techniques intended to reduce false positives. These are the provider’s descriptions, not Cortex test results. A model update can change the assessment of the same passage, so record the date and input when a result matters to a discussion.

A human label means the model favors its human-writing category; it does not establish that no assistance was used. An AI label means the passage resembles the model’s AI category; it does not identify the author or the exact tool used. Mixed is a classification for material the model associates with combined writing, not a forensic record of which keystrokes came from whom.

Confidence describes the model’s certainty under its evaluation framework. A highly confident prediction can still be wrong for an individual document. GPTZero itself advises against treating detection as a final verdict. Read the full context before drawing a conclusion from a label.

The Cortex free detector endpoint sends your pasted passage to GPTZero’s v2 text-prediction endpoint when the service is configured. It maps the returned document classification to a human, mixed or AI label and uses the AI-class probability for its main number. That number is not a similarity or plagiarism percentage.

Cortex also reads the sentence and paragraph detail GPTZero returns, its confidence category and its burstiness figure, and renders them alongside five writing signals it measures itself. API fields and model responses change over time, so this is not guaranteed to surface every field in GPTZero’s own dashboard. An error message is not a low AI score.

GPTZero says longer prose is a stronger basis for analysis than short snippets and notes that no detector is perfect. Poetry, code, quotations and narrowly constrained technical text may not resemble the prose a reader has in mind when they hear an accuracy claim.

Language background also matters. Stanford research showed that older detectors could misclassify non-native English writing. GPTZero describes later bias-reduction work, but neither that improvement nor the older study establishes a universal current error rate. Do not compare a vendor’s benchmark percentage with an unrelated classroom sample as though the tests were identical.

Save the precise passage and the full result, including its date. Compare flagged passages with your real notes, source material and version history. Ask whether an abrupt change in voice has a straightforward explanation, such as a properly attributed quotation or a terminology-heavy section.

If you use Cortex to revise, judge whether the revision improves meaning and readability. Do not rewrite merely to make a number smaller, and do not use Auto-Writer to manufacture evidence of composition. Only submit work you understand, with permission and disclosure appropriate to the assignment.