AI text detection is the practice of estimating whether a passage was written by a person or generated by a language model like GPT. A detector reads the text, measures statistical patterns in how the words are chosen and arranged, and returns a probability or a label. That is the whole idea in one sentence. The complications start the moment you ask the same detector to read Japanese, Chinese, Spanish, or Arabic instead of English, because almost everything a detector relies on is shaped by the language it was trained on.
This guide explains how these tools actually work, why language matters so much, and how to read a multilingual detection result without fooling yourself. If you are a teacher grading essays from students who write in more than one language, or an administrator choosing a tool for a whole district, the differences below are not academic. They change which results you can trust.
What a detector is actually measuring
Most AI detectors are not magic and they are not reading for meaning. They are measuring two statistical properties of the text.
The first is perplexity, which is a measure of how surprised a language model is by each word. Human writing tends to be a little unpredictable. We pick odd words, we hesitate, we repeat ourselves, we make slightly strange choices. Machine writing tends to be smoother and more probable, because a model is built to choose the likely next word. Low perplexity, meaning the text is very predictable, is a soft signal that a machine wrote it.
The second is burstiness, which measures variation in sentence length and complexity. Humans write in bursts. We follow a long winding sentence with a short one. We vary our rhythm without thinking about it. Models, left to their own settings, often produce sentences of more uniform length and structure. Low burstiness is another soft signal.
Some detectors add a third layer: a classifier trained directly on large piles of human text and machine text, learning the fingerprints that separate them. The GPT-2 Output Detector that circulated widely a few years ago was exactly this kind of classifier, fine-tuned to recognize the output of one specific model.
Notice the common thread. Every one of these methods depends on a model that has a sense of what is normal in a language. And that sense is learned from training data. This is where languages diverge.
Why language changes everything
A detector trained mostly on English has a confident, well-calibrated sense of what surprising English looks like. Hand it Spanish and that confidence is borrowed, not earned. Hand it Japanese and it may be guessing.
Three concrete reasons explain the gap.
Tokenization differs. Detectors break text into tokens, the small chunks they actually count. English splits cleanly on spaces. Japanese and Chinese do not use spaces between words at all, so a tokenizer has to segment the text first, and it often does this badly for non-English scripts. A model that was optimized for English may slice a Chinese sentence into awkward pieces, which scrambles the perplexity math before any judgment is made.
Training data is lopsided. The public web is dominated by English. The models underneath most detectors saw far more English than any other language, so their notion of normal is sharpest in English and fuzzier everywhere else. For a lower-resource language, the model's expectations are shakier, which means its surprise signal is noisier and less meaningful.
The grammar itself behaves differently. Burstiness assumptions built around English sentence rhythm do not transfer cleanly to languages with different sentence structures, different punctuation, and different norms for length. A perfectly natural piece of Japanese prose can look statistically unusual to a tool that learned its instincts from English essays.
The result is predictable. The same detector that performs respectably on English can become unreliable in another language, sometimes flagging fluent human writing as machine generated, sometimes waving machine text through. The competitor tools that ship separate Japanese and Chinese detector pages are tacitly admitting this. One detector does not fit all languages.
The main families of detection tools
It helps to know what kind of tool you are looking at, because the category tells you a lot about its limits.
Model-specific output detectors. These are trained to catch one model's output, like the original GPT-2 detector. They can be sharp against their target and nearly useless against anything newer, because every new generation of model writes a little differently. A 2019-era detector has little to say about today's text.
General classifiers. These are trained on a broad mix of human and machine writing and aim to generalize. They are more flexible but tend to be less certain, and their accuracy drops fastest when they meet a language or genre they did not train on.
Perplexity-based scorers. These run your text through a language model and score how predictable it is. They are transparent and do not need a separate trained classifier, but they inherit every blind spot of the underlying model, including its weakness in non-English languages.
Watermark and provenance methods. A newer and fundamentally different approach. Instead of guessing after the fact, the model that generates the text embeds a subtle statistical signature as it writes, which a checker can later detect. This is more reliable when it is present, but it only works if the generating model cooperated, which most do not. It is provenance, not forensics.
For schools, the practical takeaway is that the dramatic single-number verdict you sometimes see is almost always coming from one of the first three families, and all three are softer than they look, especially outside English.
Worked example
Imagine a student submits a short paragraph in Spanish. A detector reports 88 percent AI. What does that number mean?
It means the tool found the passage more predictable and more uniform than its internal sense of typical writing. But that internal sense was probably calibrated on English. The Spanish text might be flagged because it is fluent and clean, the exact qualities of a strong human writer working in a second language. A careful, grammatically tidy essay from a diligent student can read as machine-like to a detector that equates smoothness with automation. The number is a statistical estimate, not a confession.
Now run the same student's English paragraph through the same tool and you might get 30 percent. Same student, same effort, different language, wildly different score. Nothing about the student changed. The detector's confidence in the language changed.
Common misconceptions
A high percentage is a verdict. It is not. It is an estimate of statistical similarity to machine text, with an error rate that grows in non-English languages. Treat it as one input, never as proof.
One detector covers all languages. Almost never true. A tool tuned for English will behave differently, and usually worse, in Japanese or Chinese. If you teach in multiple languages, ask explicitly which languages a tool was evaluated on, not just which ones it accepts.
Non-native writing is safe from false flags. The opposite risk is real. Several lines of research have found that detectors disproportionately flag writing by non-native speakers, because that writing can be simpler and more regular in exactly the ways detectors reward. This is the single most important fairness issue in multilingual detection, and it falls hardest on the students who can least afford a wrong accusation.
Newer models are easy to catch. Each new generation writes more like a person, with more varied, more surprising prose. Detection has been getting harder, not easier, and a tool's headline accuracy figure is only as current as the models it was tested against.
How to read a multilingual result responsibly
A few habits protect you and your students.
Check the language coverage before you trust a score, and be more skeptical the further the text sits from English. Look for a confidence range or an explanation rather than a lone number, because a tool that admits uncertainty is being honest about how this technology works. Never let a percentage stand as the sole basis for an academic integrity decision, in any language. Pair it with what you know: drafts, version history, the student's voice across past work, and an actual conversation.
Detection is a useful signal and a poor judge. Across languages that gap only widens. The right way to use these tools is as a prompt to look closer, never as the last word, and the further you travel from English the more that caution matters.