Skip to content
Convertto

A ContentScale AI detector alternative that shows its working

People searching this way want to know whether a piece of text was written by a language model.

Convertto's AI content detector runs a real language model over your text in your browser and reports the two measurements published detection methods are built on: perplexity, the average surprise the model felt at each token, and burstiness, how much that surprise varied between sentences. It reports the numbers and deliberately refuses to issue a verdict, because stylometric measurement cannot establish who wrote a text.

Method
Per-token log-probability from a language model: perplexity and burstiness
Scoring model
GPT-2 (124M), a one-time 267 MB download, then run on your device
Text uploaded
None. The text never leaves your browser
Verdict issued
None. The measurements are reported with their limits
Cost
Free, no sign-up

What you are looking for, and the tool that does it

The jobConvertto toolWhat it does
An AI detection scoreAI Content DetectorReports per-token perplexity and burstiness from a language model run locally
Surface tells of generated proseAI Writing Pattern CheckerMeasures sentence-length variance, vocabulary richness, marker phrases and typographic tells
Hidden characters in pasted textInvisible Character DetectorFinds zero-width spaces, directional overrides, Unicode tag characters and homoglyphs
Whether two texts are the sameText Similarity CheckerCompares two texts with Levenshtein, Jaccard, cosine and Dice
Readability of a passageReadability Score CheckerScores with Flesch, Gunning Fog, SMOG, Coleman-Liau and ARI

What is actually being measured

Every published detection method rests on the same idea. A language model assigns a probability to each token given the tokens before it. Run the model over a passage and you get a sequence of probabilities, and from those two summary numbers: perplexity, roughly how surprised the model was on average, and burstiness, how much that surprise varied from sentence to sentence.

MeasurementWhat it capturesThe tendency it is based on
PerplexityAverage per-token surprise across the passageGenerated text tends to choose high-probability continuations, so it is less surprising
BurstinessVariation in surprise between sentencesHuman writing tends to swing between predictable and unexpected; generated text tends to stay level

Note the word "tends" in both rows. These are population tendencies, not properties of an individual document, and that distinction is the whole reason this tool does not print a percentage and call it a day.

Why no verdict is issued

The consequences of that error rate are not symmetric. A false negative means a generated passage goes unnoticed. A false positive means a person is accused of dishonesty on the basis of a number they cannot inspect or contest. Reporting the measurements, and what they do and do not support, is the honest form of this tool. Reading them as evidence is a judgement a person has to make with the rest of the context in front of them.

Reading the numbers usefully

  • Compare, do not threshold. The numbers are most informative against other samples from the same writer on the same topic, not against a fixed cutoff.
  • Length matters. A short passage gives too few tokens for either measurement to be stable. Score whole documents, not sentences.
  • Editing moves everything. A generated draft that has been substantially rewritten scores like the person who rewrote it, which is the correct outcome and also why a verdict would be meaningless.
  • The scoring model is not the writing model. GPT-2 is used as a probe. Text from a much larger and more recent model is not something a 124M-parameter probe can characterise reliably.

Running the model on your own device

The scoring model is downloaded once, about 267 MB, and then executed in your browser. Nothing you paste is transmitted. That matters here more than for most tools: the text people want to check is frequently a student submission, a job application, or an unpublished manuscript, and sending those to a third party to be scored creates a copy that outlives the question you were asking.

The AI writing pattern checker works on surface habits rather than probabilities and is a useful second view. The invisible character detector catches zero-width watermarks that some pipelines insert. The rest of the on-device models are in local AI tools.

Frequently asked questions

Does it tell me whether the text was written by AI?

No, deliberately. It reports perplexity and burstiness, the measurements detection methods are built on, and explains what they do and do not support. Turning those into a verdict would hide an error rate that falls on real people.

Is my text uploaded?

No. The language model runs in your browser, so the text never leaves your device.

Why is there a 267 MB download?

That is the GPT-2 scoring model. It is fetched once and cached, because running it locally is what keeps the text on your device.

Can I use this to accuse someone of using AI?

You should not. These measurements are population tendencies and they misfire on non-native speakers and on plain formal writing. They can inform a conversation; they cannot establish authorship.

Are AI detectors accurate?

No published stylometric method establishes authorship reliably. They measure statistical properties that correlate with generated text, and both false positives and false negatives are common enough to matter.

Open a tool

Last updated

Browse local ai tools