Accuracy Disclaimer

A number between 0 and 100 that expresses how closely a text matches the statistical patterns of machine-generated writing. It is an estimate about a text. It is not a determination about a person, and no threshold on it constitutes proof.

Inkprint OS publishes the eight measured signals behind every score, with their actual values, precisely so that you can judge the result rather than take it on faith.

• Short texts. Below 60 words no detector on the market is reliable. We label those results low confidence rather than pretending otherwise. Reliability rises with length and is reasonably stable above roughly 300 words of continuous prose.
• Formal and technical writing. Academic papers, legal documents, policy text and technical documentation are structurally even and vocabulary-dense, which is exactly what the signals measure. Human authors of this kind of writing score higher.
• Writing by non-native speakers. People writing in a second language tend to use more predictable vocabulary and more uniform sentence structure. Every detector in this category flags them more often than native speakers. This is a documented bias of the whole approach, not a flaw unique to us, and it does not have a fix we can honestly claim to have made.
• Text a person edited. An AI draft that a person genuinely rewrote should read as human, because at that point it substantially is. The score reflects the text in front of it, not its history.
• Text through a paraphraser. Paraphrasers change vocabulary and weaken the lexical signals. Sentence rhythm survives better, which is why it carries the most weight, but detection of laundered text is meaningfully harder.

A false positive is human writing flagged as machine written. This is the error that damages people, and it is the one we tune against: the thresholds are set to keep it low, at the cost of missing some genuinely generated text.

A false negative is generated text that passes. It happens, particularly with short passages, edited drafts and text put through a rewriting tool.

Neither can be eliminated. Any product telling you otherwise is selling you something.

If someone has shown you a score about your work and drawn a conclusion from it, the following are all true and all worth saying out loud:

• A percentage is not evidence. Ask what it was measured on and how long the text was.
• A high score on a short passage means very little.
• A high score on formal writing, or on writing in your second language, is a known failure mode of these tools.
• Drafts, version history, notes and the ability to talk through your own argument are stronger evidence of authorship than any detector output.

Using a score to make automated accusations, or as the sole basis for an academic, disciplinary, editorial or employment decision, breaches the Acceptable Use Policy and can have the account closed.

Before acting on any result, check the rules your institution or organisation sets for disclosing AI assistance. Those rules, not our number, are what governs the situation.

Nothing produced by the service is legal, academic, editorial or employment advice, and no relationship of that kind is created by using it.

Think a result is wrong? Send it to support@inkprintos.com. Cases where the engine got it wrong are how it gets better, and we would rather have the example than be told we are accurate.