What AI detectors measure, and why we promise nothing

An AI text detector is a classifier. It was shown a large pile of human writing and a large pile of machine writing, and it learned to tell them apart on average. The good ones are right most of the time on the kind of text they were trained on. None of them reads your intention, your sources or your history. They read patterns.

Two things follow. First, a detector can be wrong about any single text. A careful non-native writer who learned English from textbooks produces exactly the even, well-connected prose the detectors associate with machines. Several universities have documented this, and some have dropped detectors because of it. Second, detectors are trained on the models of their time. When a new model or a new rewriting method appears, the detector has not seen it, and its vendor retrains. The cycle repeats every few months.

That is why Samesay shows you a style score and a meaning check, and never a detector score. A style score is ours: it counts habits a reader notices, like stock openers, flat sentence rhythm and dashes. It is stable, explainable, and it has nothing to do with what a detector will decide next week.

If you are a student, the only reliable position is the rules of your institution. Many allow AI help for some tasks and not others, and some ask you to declare it. A text that was allowed does not need to pass anything; a text that was not allowed is a problem whatever a detector says. We wrote our terms around that line, and we mean them.

If you are a professional, the question is whether the text sounds like you and says what you meant. That is what the diff view and the meaning check are for. Use them, read the result once, and sign it with your own name.