How AI Detector Checker Analyzes Text

What you get

• A 0-100 AI-writing signal score with its score stability range.
• One of three bands — Likely human-written, Uncertain, Likely AI-written — decided by frozen thresholds that depend on text length.
• The measured false-flag rate for that length: about 2 in every 1,000 human texts of 150 words or more, about 2.4 in every 100 at 100 to 149 words.
• The patterns observed in your text (deterministic measurements, not explanations of the classifier).
• English only, 100 words minimum; pasted text or an uploaded PDF, DOCX, TXT or RTF file.

AI Detector Checker is a free AI detector that reviews any text for AI-like writing signals. You paste your text, the tool analyses it, and it returns a 0-100 AI-writing signal score with a score stability range and one of three bands (Likely human-written, Uncertain, or Likely AI-written). The score is an estimate of AI-writing signal strength; its stability range shows how much the score mapping varies across resampled source families. This page explains, in plain language, what happens between pasting your text and reading the result — and, just as importantly, what the result can and cannot tell you.

Using the tool: step by step

1. Open the detector. No account or sign-up is required. You use the detector through the web interface; submitted text is sent securely to our server for analysis and processed as described in our Privacy Policy.
2. Paste your text. Copy your text into the box, or type directly. Submit at least 100 words; shorter input returns a request for more text. The current validation is strongest for English text; other languages have not been validated for the current detector, so results on them are unsupported.
3. Start the review. The tool first checks that the text meets the minimum and maximum length it can work with, then runs its analysis. Processing time varies with text length and server load; the result is shown when processing is complete.
4. Read the result. You see a document-level AI-writing signal score with its stability range, one band (Likely human-written, Uncertain, or Likely AI-written), the measured false-flag rate for that length, the patterns observed in the text, and the reminder that the result does not prove AI or human authorship. The result is for the whole text; no part of it is marked separately. Use the label to decide whether the text deserves a closer human read — not to reach a verdict about who wrote it.

What happens behind the scenes

The detector uses a multi-signal approach. Instead of relying on any single test, it measures several independent linguistic and statistical patterns and then combines them inside a single document-level classifier that produces the final label. Here is the general idea behind each stage.

• Receiving and validating the text. The tool checks the input format and length. Text that is too short carries very little signal, so the result for short passages should be read with extra caution.
• Word patterns. The tool measures which words and short word sequences appear, and how often, compared with patterns learned from large samples of human-written and AI-written text. This is a signal, not a measurement of authorship.
• Character and phrasing patterns. The tool also looks at character-level sequences, which capture spelling, punctuation habits, and phrasing that tend to differ between machine-generated and human prose.
• Statistical patterns. The tool looks at measures such as variation in sentence length (burstiness), word-repetition patterns, vocabulary richness, and punctuation rhythm. These checks work on the structure of the text rather than its meaning.
• Producing the label. A single document-level classifier combines these signals into one internal value. That value is mapped to the 0-100 score by a frozen monotonic mapping and to one of three bands by frozen thresholds that depend on text length, chosen to keep wrong AI signals on human writing rare. In a sealed test on human-written texts of 150 words or more, from sources the system had never seen, it gave a wrong AI signal for about 1.7 in 1,000 texts. For texts of 100 to 149 words it was about 2.4 in every 100. The internal value is not the chance that AI wrote the text; the displayed score is a fixed mapping of it, and the band is decided by the raw thresholds only. The label is not proof of authorship.
• Reporting the result. The label is shown with the reminder that it does not prove AI or human authorship. The tool does not mark individual sentences; read the text yourself to decide where a closer look is warranted.
• Interpreting in context. The result should be read together with what you already know about the text — its length, language, subject, genre, and editing history. Heavily edited AI text and carefully structured human text can both sit in the middle of the range.

We describe the general methodology openly, but we do not publish the exact internal weightings or thresholds that would make the tool trivial to game.

Limitations you should keep in mind

• The current validation is strongest for English text of at least 100 words. Other languages have not been validated for the current detector; results on them are unsupported.
• Short texts, quotations, lists, and translated or heavily edited passages are harder to assess and can produce misleading results in either direction.
• A strong-signal label is not confirmation that AI wrote the text, and a no-signal label is not confirmation that a human did. The tool cannot identify which specific person or AI system produced a piece of writing.

The label is a pattern-based review signal. It is not proof that a specific person or AI system wrote the text.

Now that you know how it works, try the free AI detector on your own writing. Treat the result as one input for review alongside your own judgement — never as a conclusion about who wrote the text.

For the full technical account — the safeguards applied after scoring, how the detector is trained and validated, and the exact engine currently deployed — see the methodology, the training and validation page, and the model card.

What V3 is, in one paragraph

V3 uses the same trained core as V2. Nothing new was trained for it. What changed is the decision layer around that core: new thresholds that depend on the length of your text, a frozen mapping from the internal value to the 0-100 AI-writing signal score, and a stability range for that score computed by resampling the source families behind the mapping. The verdict band is decided by the raw thresholds alone, so the number you see never changes which band you get. Every one of those thresholds was fixed before the sealed test and has not moved since; the measured results, including the two targets the test did not meet, are on the evidence page.