How the GPTHuman AI Detector Produces Its Score

The detector sends the submitted passage with a fixed analysis instruction to a configured AI provider. The model reviews the passage for linguistic signals associated with generated and human writing, then returns two complementary scores, one overall verdict, and sentence-level labels.

This implementation is not a locally calculated perplexity or burstiness classifier. Those concepts can explain some detector research, but the current GPTHuman workflow does not expose or calculate either metric. Read the separate perplexity and burstiness guide for general background.

1. Your browser submits the passage

   The detector accepts plain text through the tool form. The server requires at least 50 words and rejects empty, oversized, or malformed requests.

2. The server validates the request

   Input type, size, tool name, and word count are checked before any model call. A quota reservation records operational usage for the request.

3. A configured provider evaluates the text

   The submitted passage and detector instruction are sent to an available AI provider. Provider and model availability can change the route used for a request.

4. The server validates the response shape

   The response must contain numeric AI and human scores, an allowed verdict, and a sentence array before it is returned to the interface.

5. The result appears in your browser

   GPTHuman returns the analysis without writing the submitted text or generated result to the application’s usage-history table. See Text and usage processing for the policy details.

Sentence variety

Changes in sentence length and structure, including repeated openings or unusually even paragraph construction.

Vocabulary range

Whether word choice is varied and context-specific or relies heavily on predictable, generic phrasing.

Idiomatic usage

Expressions and phrasing that fit the language and context, without assuming that fluency proves human authorship.

Specificity

Concrete examples, details, constraints, and qualifications compared with broad claims that could fit almost any topic.

Personal voice

Subjective perspective or situated experience when the genre would reasonably contain it. Its absence is not automatically an AI signal.

Formulaic patterns

Stock transitions, excessive hedging, repeated sentence shapes, generic examples, and mechanically balanced sections.

A displayed “80% AI” result should not be translated into “an 80% chance this person used AI.” The number is a model output under the current instruction, provider, text, and threshold. GPTHuman has not published a calibration study that supports that personal-probability interpretation.

• Check whether highlighted sentences repeat the same opening, use generic transitions, or lack the detail promised by the surrounding section.
• Compare the sentence with drafts, notes, citations, and document history before making an authorship claim.
• Keep exact facts, names, dates, quotations, URLs, and technical terms intact when revising.
• Do not add errors, random synonyms, or awkward fragments merely to move a detector score.
• For a disputed result, preserve the exact passage, date, screenshot, tool, and any visible settings before retesting.

If verified human work is flagged, follow the evidence-preservation and appeal workflow in AI detector false positives.

Short text

The tool accepts 50 words, but the current instruction treats passages under 100 words as lower-signal evidence.

Mixed authorship

A human-edited AI draft or AI-edited human draft may not fit a clean binary origin label.

Language and genre

Performance can vary across languages, formal templates, code, quotations, reference lists, and specialized writing.

Provider changes

Configured providers, model versions, and service availability can change. GPTHuman does not currently expose a public detector version identifier with each result.

Repeat runs

Model-based analysis can vary between runs. A second score is another model output, not independent evidence of provenance.

Consequential decisions

Grades, employment actions, accusations, or publication decisions require stronger evidence and a fair human review process.

The 500-sample test protocol is a preregistration-style plan with blank result tables. It defines sample groups, thresholds, error metrics, stress tests, and reporting requirements, but it is not a completed benchmark.

• No accuracy, precision, or false-positive percentage is claimed on this methodology page.
• A future result must identify the frozen dataset, detector configuration, test date, exclusions, failed runs, and uncertainty intervals.
• Any material change to the detector instruction or score rule should update this page’s reviewed date and version note.