How to check a blog or essay for AI writing without misreading the score

01Read the class probability before interpreting a percentage

The official GPTZero API reference describes three document classes: human, AI and mixed. A high probability for mixed writing means the model favors that classification; it does not identify the proportion of words written by AI. Adding the AI and mixed probabilities creates a different quantity and must not be labeled an AI word percentage. A model confidence label is also not a measured accuracy rate for your particular document. See how precision, recall and a confusion matrix use known labels before comparing these different quantities.

An accuracy or false-positive rate requires a defined evaluation set with known provenance and a stated decision rule. You cannot infer either from a single scan. The independent M4 research benchmark found that generalization to unfamiliar domains and generators remains difficult. That study is a reason to examine conditions, not a certification of the current GPTZero model or this website.

02Prepare a reviewable submission for your scenario

For a blog, use the article body rather than navigation text, comments or an entire page dump. For an essay, preserve the original document and check extracted paragraph order, formulas and tables before submitting. Optional bibliography exclusion starts at the final standalone supported References heading and continues to the end. Code exclusion handles Markdown fences. Both options can change the material being judged, so review the preview and retain the same scope when comparing reports.

KitNelo reads UTF-8 TXT, Markdown, DOCX and unencrypted text PDFs locally, up to 8 MB and 500,000 characters; only PDF has a 30-page limit. A provider scan needs at least 300 trimmed characters for Winston AI or 250 for GPTZero and accepts up to 50,000 characters, subject to shared site credits. A scanned PDF needs OCR and proofreading first. Long documents can be kept intact locally while you choose a suggested segment or custom range. Each submission is explicitly previewed; segment scores are never averaged into a whole-document score. Local counts and citation-marker cues describe text structure; they are not authorship or citation verification.

03Use the tool and preserve its actual evidence

Preview the submission, distinguish class probabilities from authorship evidence and export a clearly labeled local report when no detection account is connected.

Open AI Text Detection & Review and choose the blog, paper, business or other-text scenario. Import or paste your text, then inspect the submission preview. Run the free local review to check preparation without uploading the text. For provider detection, use a configured site account or temporarily enter your own official GPTZero API key. Confirm the disclosure before submitting: your previewed text passes through the server to GPTZero and can consume provider credits. The key is kept in page memory rather than browser storage or exported reports.

A successful scan shows the actual model version, time, language when supplied and provider-returned sentence signals. Flags identify passages to review; unflagged passages are not proof of human writing. Provider exclusions and missing locations remain explicit. Changing the text or scope clears the previous result and consent. A failed request or cancellation retains your input and never generates a substitute score. A cancelled or timed-out request may still be billed, so check provider history before submitting again.

Export JSON for structured evidence, TXT for reading, or printable HTML for browser print-to-PDF. Reports record scope and SHA-256 fingerprints of the source and submission. Source and sentence text are omitted by default; opt in only if the recipient needs them. Review the title and metadata before sharing. A fingerprint links the report to text you can compare; it is not a signed certificate or proof of the author’s identity.

04For papers and editorial decisions, examine the draft history

Compare a flagged passage with notes, dated drafts, source citations and the writer’s explanation. Keep the exact provider, model version and scope when discussing a result. A score from one provider cannot be converted into a school’s Turnitin result, and a low score does not establish factual accuracy, originality or acceptable use under an institution’s rules.

Turnitin’s own guidance says detection should not be the sole basis for adverse action against a student. Apply human review and the relevant policy rather than treating a percentage as a verdict. For blogs, separately check facts, attribution and permission to publish. This tool does not provide plagiarism checking, search-ranking guarantees or institution-certified AI percentages. Official API documentation was checked on November 2, 2026; no paid detection account was available to validate detector accuracy during implementation.