How ScanForAI works

ScanForAI reads a piece of writing the way a machine reads it and estimates how much of it looks AI-written. It returns a 0–100 score, a plain-language band, and the specific patterns behind that score — a signal, not a verdict.

What the scanner measures

No single word or sentence proves anything. ScanForAI weighs dozens of independent language signals at once and looks at how densely the machine-associated ones cluster. They fall into five families:

• Rhythm. People instinctively vary short and long sentences; unedited machine writing tends to stay evenly paced. The scanner measures that variation.
• Vocabulary. Models reach for a recognisable set of words and stock phrases far more often than people do — see the words AI overuses.
• Phrasing. Contractions, personal asides, and formulaic constructions (the “it’s not just X, it’s Y” cadence and its relatives) read differently by hand than by model.
• Predictability. Machine writing is unusually easy to predict word to word — a statistical fingerprint of how it was produced.
• Structure. Default scaffolding — tidy lists, signposted sections, formulaic openings and closings — is one of the clearest tells of unedited generation.

Each family contributes several measurements. Which signals matter, and how much, is tuned against a labelled corpus of human and AI writing that we refresh as models change. That weighting is the part that keeps the score accurate — and it’s not something we publish.

How the score is formed

The individual signals are combined by a trained classifier into one number, from 0 (reads human) to 100 (reads like AI). Crucially, the result doesn’t stop at the number: it shows which signals fired on your text and how strongly, so you can see the reasons instead of trusting a black box. That per-result explanation is the promise — not a published recipe that a model trying to slip past detection could train against.

Alongside the score, every result reports a confidence — high, moderate, or low — so a read built on too little text, or one sitting in the ambiguous middle, says so plainly instead of dressing a guess up as certainty.

Register awareness

Many tells are specific to a kind of writing. Formal and academic writing naturally uses few contractions and even sentences; flagging it for that alone would be wrong. ScanForAI first works out the register of the text and eases off the signals that only mean something in casual writing, so genuinely formal — or non-native — human writing isn’t punished for being formal. Guarding against that false positive matters more to us than squeezing out a slightly higher hit rate.

Runs on our own hardware

Scanning happens on infrastructure we control. ScanForAI never sends your text to a third-party AI API, and anonymous scans are never stored — the text is processed to produce the score and then discarded. See the Privacy Policy.

Honest limitations

• No detector is 100% accurate — including this one. Every score is an estimate.
• Formal, technical, or non-native human writing can read high; edited or newer AI can read low.
• A light paraphrase or editing pass removes most surface tells, and the newest human-tuned models can evade detection almost entirely. A low score means the text reads human — it can’t certify how it was made.
• Tells decay as models change, which is why we keep re-tuning against fresh writing.
• It is most reliable on 300+ words of writing, and least reliable on short or highly technical text.

Treat the score as one input, never the sole basis for a decision about a person, and never as grounds to accuse. See interpreting your score for how to read a result responsibly.