AI Text Detector: How Detection Works and How to Read a Score
What AI text detectors measure, why two tools disagree about the same paragraph, where detection fails, and how to read a pattern score without over-reading it.
Teachers, editors and SEO teams all want the same thing from these tools: a straight answer about who wrote a paragraph. That is the one thing no detector can give them.
An AI text detector is a tool that scans writing for statistical patterns associated with language-model output, such as uniform sentence length and predictable word choice, and then reports a score. The score is an estimate about pattern, not a record of authorship, and no detector can confirm who wrote a text.
What an AI text detector looks at
Commercial detectors differ in implementation, but the signals they reach for overlap heavily. Four come up again and again, and knowing them changes how you read any score you are shown. None of the four is evidence of authorship on its own. Each one describes a property of the prose, and properties of prose can come from a model, from a hurried writer, or from a style guide that demands formality.
Perplexity
Perplexity measures how surprising a text is to a language model. Model output tends to be low-perplexity, because the model keeps choosing the word it considers most likely. People choose the less likely word more often, so their writing tends to sit higher on the same scale. Genre pulls this around too: a policy document written by a committee of humans is not a surprising text.
Burstiness
Burstiness is the variation in sentence length and complexity across a passage. A person drafting freely will drop a four-word sentence between two long ones without thinking about it. Raw model output usually marches along at a steady rhythm instead, and that flatness is the single pattern these tools pick up most consistently.
Repeated openers
Models reach for a small pool of transitions and sentence starts: "Furthermore," "Additionally," "It is important to note that," "In conclusion." One of those in a document means nothing. Nine of them in a document of eight hundred words is a pattern, and every detector on the market counts it.
Vocabulary spread
Writers restate an idea from different angles across a piece, so their vocabulary fans out. Model output often reuses the same formula for the same idea each time it comes up, which narrows the spread. This signal is the weakest of the four and the easiest to move by accident.
How the estimate on this site is built
Our rule-based pattern estimate doesn't use a neural classifier. It is deterministic (the same text always gets the same score), which says nothing about accuracy: it has not been validated against labelled human and AI text. It scores text using:
- Burstiness: sentence length variance
- Predictability: lexical variety and phrase diversity
- AI marker penalty: known clichés, filler phrases and stock openers
- Contraction and informality index: how often contractions appear
Because every component is visible, you can see which passage moved the number. That is the only thing the breakdown is good for. It tells you where your prose is uniform; it does not tell you who typed it.
Why detection is hard
The honest summary is that no AI text detector is reliable enough to act as proof, ours included. The distance between careful model output and ordinary human prose is small, and the signals above sit on the wrong side of that gap. Heavily edited model text reads as human because it has become good writing. Rushed human text reads as machine-written because it has become flat writing.
Some categories get hit harder than others:
- Non-native English writers, whose sentences are often deliberately regular
- Academic and legal prose, where formal register is the requirement
- Short submissions, where there is not enough text for a variance measure to mean anything
- Technical documentation written to a house template
That failure pattern is why we show the specific signals behind an estimate rather than a bare percentage, and why the label attached to a score talks about patterns rather than authors.
What we do and do not claim about other tools
GPTZero, Originality.ai, Turnitin, Copyleaks and ZeroGPT all use some form of neural classifier, and several combine it with plagiarism matching. We have not run an accuracy comparison against any of them, so we publish no ranking, no accuracy figure and no grade for their methods. Anyone who does publish such a table owes you a method: the labelled dataset, its size, where the text came from, and how ties were handled. Treat a comparison without that method the way you would treat a statistic without a source.
The practical consequence for you is simple. Two tools can read the same paragraph and return opposite verdicts, and neither result settles the question. If a decision matters, it needs more than a score behind it.
How to read the score
The score describes writing patterns. It is not a verdict on who wrote the text, and it does not predict what another detector will report.
| Score | Label | What it means |
|---|---|---|
| 85–100% | Strongly human-like patterns | Varied sentence lengths, few stock phrases |
| 65–84% | Mostly human-like patterns | Some uniform or formulaic passages |
| 45–64% | Mixed patterns | Noticeably uniform rhythm or repeated openers |
| 0–44% | Many AI-like patterns | Uniform sentences, stock phrases, repeated openers |
Short texts, formal writing and non-native English often score low even when a person wrote every word. There is no number in this table you should be aiming for, and treating any band as a pass mark misreads what the tool does.
What to do when your text reads as machine-written
A low band is worth acting on for one reason only: uniform, cliché-heavy prose is harder to read. Fix it as a writing problem and ignore the number.
Our rewriting tool edits for clarity. It replaces stock phrases with plain wording, turns passives that name an agent into active voice, unpacks noun-heavy phrases into verbs, and splits long coordinated sentences. It leaves a sentence alone when it can't rewrite it safely, and it uses contractions in Casual mode only. The score may go up, stay the same, or go down afterwards; the tool reports whatever it measures.
By hand, the changes that matter most are structural rather than lexical:
- Mix sentence lengths on purpose, including at least one very short sentence per section
- Delete stock transitions outright and see whether the sentence still needs a bridge
- Add one specific detail per section that only you could supply
- Use contractions where the register allows them
- Rewrite passives that hide who acted
The readability checker will show you sentence-length variance directly, which is more useful than a detection band when you are editing. For a longer treatment of the editing work, including what to do with an entire draft rather than a paragraph, see our guide to humanizing AI content. If your worry is copied source material rather than machine rhythm, removing AI plagiarism covers that separately, because the two problems need different fixes.
AI detectors and academic integrity
Detectors are widely used in education and widely argued about. If you are a student, four things are worth knowing:
- AI detection is not infallible: a false positive can happen even with 100% human writing
- Most institutions treat a detector result as one signal among several, not as automatic proof of cheating
- If you draft with AI and then substantially rewrite, whether the result counts as your own work depends on your institution's policy
- Read that policy before you use any tool, including this one
If you are on the receiving end of a flag, the useful response is evidence of process: drafts, notes, version history. A counter-score from a second detector proves nothing, because the second detector is no more reliable than the first.
Does Google penalize AI-generated content?
Google's published position is that it aims to reward helpful content regardless of how it was produced, while its spam policies target content produced at scale to manipulate rankings. Nothing about a detector score feeds into that. Our guide to AI content and SEO works through the policies in detail.
Frequently asked questions
Is FreeTextDetector.com free?
Yes. Both the pattern estimate and the rewriter are free, with no account required, up to 5,000 characters per submission.
Can an AI detector be wrong?
Yes, in both directions. Non-native English writers and academic authors can score as machine-written, and edited model output can score as human. Use any score as one signal, never as proof.
What is the best free AI text detector?
There isn't one. The tools disagree with each other and none is reliable enough to act as proof. Ours is free and shows the signals behind its estimate, which makes it useful for learning what these systems look at.
Does editing for a better score work?
Editing for varied rhythm, plain wording and specific detail makes text read better, which is worth doing on its own terms. No tool can promise a particular result from any detector, and ours doesn't.
Why did two detectors give me opposite answers?
Because they measure different things and were tuned on different text. Disagreement between detectors is the normal case, not a malfunction in one of them.
What to take away
Read a detection score as a description of your prose and nothing more: it tells you whether your sentences are varied and your phrasing is fresh, and it cannot tell you or anyone else who wrote the text. The failure mode to watch for is the one that looks like diligence — treating a band as a target and editing until the number moves, which produces text tuned for a heuristic instead of for a reader. If you only change one thing after reading this, make it sentence-length variance in your own drafts.