Free AI Content Detector

Free AI Content Detector

What this AI content detector does

This page analyses text and returns a score from 0 to 100, a label and a set of writing signals. You can paste text or upload a PDF, DOCX or DOC file. The browser extracts the readable text, then the server scores it. The result is a statistical opinion about patterns in the writing, not proof that a person or a machine wrote it. It is designed for editors, teachers, publishers and anyone who wants a second look at a suspicious passage.

How to submit text

Paste the text into the input area and press the analyse button. The tool expects at least 50 characters, and the server accepts up to 50,000 characters. For a document, choose a PDF or Word file and let the browser extract the text before submitting. Scanned PDFs and image-only documents contain no text layer, so they will not work without OCR. Check the extracted preview if you need to confirm that the right content was read.

What the score means

The score is a probability-like number, not a percentage of words written by AI. A score of 81 or above is labelled AI Generated. Scores from 61 to 80 are Likely AI, 41 to 60 are Uncertain, 21 to 40 are Likely Human, and 20 or below are Human. The label is a summary of the signals. A high score means the text resembles the patterns the model associates with machine writing; it does not identify the author or the tool used.

How the detector reaches a result

The service combines twelve weighted signals. They include sentence-length variation, density of AI-typical wording, vocabulary variety, repeated phrases, contractions, stock transition words, informal markers, lists of three, punctuation habits, trailing -ing clauses and concrete details such as names, numbers and dates. Signals are combined in log-odds space, then scaled by how much evidence is available. A short sample is pulled toward the middle because there is less to go on.

Why the confidence can be low

Reliability depends on how many signals could be measured and how many words the sample contains. Very short text, unusual formatting or a language the detector does not handle well can reduce reliability. The result may show low confidence even when the score looks dramatic. Read the confidence and the signal breakdown together. A single signal is not a verdict, and a missing signal is not evidence either way.

Mixed documents

Long documents are split into segments of at least 40 words, up to 30 segments. If the segments disagree strongly, the overall label can become Mixed. That is useful for a document that contains both human-written and machine-written passages. The segment view shows where the pattern changes. Do not assume that a Mixed label means every paragraph is a blend; it usually means different parts look different.

What the signals can mistake

The detector looks at style, so it can be fooled in both directions. Formal human writing with uniform sentences and few contractions can look machine-like. Lightly edited AI text with personal details and varied rhythm can look human. Templates, legal boilerplate and translated text often produce odd results. The informality signal caps the score when slang or missing apostrophes are present, but that is a heuristic, not a guarantee.

Document and language limits

PDF extraction depends on a text layer. DOCX parsing is generally reliable, while legacy DOC extraction is heuristic and may miss content. There is no OCR, so a photographed page or a scanned PDF needs a separate text-recognition step. The tool is tuned for English; other languages may receive a low-confidence or misleading result. Tables, footnotes and multi-column layouts may be flattened or read out of order.

Privacy and data handling

Pasted text and extracted document text are sent to the Newisty server for scoring. The page does not perform the final analysis locally. Do not submit confidential manuscripts, student records, legal documents or personal data unless you accept that transmission. The browser keeps the document until you clear it or close the tab, and the server may process the text for the request. Read the site's privacy policy if data handling matters to your workflow.

How to use the result responsibly

Treat the score as one piece of evidence among many. Ask for drafts, notes, version history or an interview when the stakes are high. A detector cannot establish intent, authorship or misconduct. If you are an editor, use the signal breakdown to guide a conversation about style rather than to accuse someone. If you are a writer, use it to spot passages that feel generic, then add specific details and your own phrasing.

Improving a low-confidence result

Submit a longer sample with complete paragraphs. Make sure the extracted text is clean and in the expected language. Remove navigation text, repeated headers and pasted references that are not part of the writing. If a document is mixed, review the segment scores instead of relying on the headline number. Run the same text twice only if you want to confirm the service is stable; repetition does not make a weak sample stronger.

What not to do

Do not use the score as the only reason to reject a student, employee or submission. Do not claim that a specific AI tool wrote a passage based on this result. Do not upload confidential material without checking the privacy implications. Do not expect OCR, plagiarism detection or fact-checking. Do not treat a Human label as a guarantee of originality, and do not treat an AI label as proof of misconduct.

Frequently Asked Questions