How AI text detection works today

We asked four different AI assistants about this and compared the answers. They agree: there is no single test, and the answer is always a probability.

Seven approaches

1. Statistics: perplexity (how predictable the words are) and burstiness (how uneven the sentence rhythm is). Fast and clear, but polished human text is predictable too.
2. Zero-shot probability methods (DetectGPT, Binoculars): compare the likelihood of a text with that of its rewrites. Paraphrasing or swapping look-alike characters weakens them badly.
3. Stylometry: vocabulary variety, sentence length, connectives, punctuation. Transparent and more robust to edits, but depends on genre and language.
4. Trained classifiers (RoBERTa and similar): the core of commercial systems. Accurate on familiar data, quickly outdated when new models appear.
5. Watermarks: a signal planted at generation time. Works only if the generator added one, and disappears when the text is rewritten.
6. Writing process: drafts, edit history, pauses and pastes. The most direct evidence, but you need access to the process.
7. Ensembles: several signals plus human review. More reliable, but costlier.

Where all the sources agree

• A detector estimates a probability, not a fact of authorship.
• It is weakest on mixed texts, where a human edits AI or AI edits a human. In the studies cited, accuracy dropped to roughly 54-71%.
• False positives remain the main risk. Formal and academic writing and non-native writers suffer most.
• Paraphrasing and "humanizers" sharply reduce accuracy.
• Vendor-claimed figures (99% and up) and independent tests differ noticeably.

The percentage: what can honestly be said

If all you have is the finished text, you cannot reconstruct how it was written. So it is more accurate to say "65% of the text shows signs of machine generation" than "AI wrote 65%". A person's ideas and edits do not show up in that number.

A good system also:

• treats false positives on human text as the main error;
• calibrates thresholds per language, genre and length;
• shows which passages raised suspicion, with uncertainty bounds;
• says "not sure" when the text is short or the language is poorly studied.

Languages

English is the best studied. Russian has a lot of formal text that is "smooth" on its own and triggers false alarms. Ukrainian has the least open data, so reliability is lower. Portuguese is in better shape here.

What we already have

Here is how that maps to the seven approaches above:

• Stylometry. Our core method: variation in sentence and paragraph length, connectives, lexical diversity, punctuation, and three tiers of AI phrases in Russian, Ukrainian, Portuguese and English.
• Explainability. Not just a number: specific passages with line numbers, quotes and a note on why the signal matters. That is where the field is heading, according to the studies.
• Fragment-level view and hints. You can see which stretches are monotonous or templated, and how to fix them. Rules, no neural networks.
• Hidden characters. We find and remove the invisible characters sometimes used to mark text (259 kinds) and normalize typography. These are literal characters, not a statistical watermark in the words themselves.
• Honest limits. We abstain on short texts and warn about smooth genres (news, business writing, promo) where humans often look "machine-like".
• Measurement. Validation on a corpus with confidence intervals and breakdowns by language and length. The sample is still small: threshold 50 on 34 AI texts and 20 human ones gave 0% false positives, but that is not final accuracy.

What we want to add

• A bigger corpus: at least 600 human texts per language, plus non-native writers, paraphrases and mixed texts. Every detector is weakest on exactly these today.
• Calibration by language and length. "Human / uncertain / AI" zones from corpus distributions instead of hand-picked thresholds.
• A range instead of one number. A result with an uncertainty band and a "no data" mark when no conclusion is possible.
• Share by fragments. "This part of the text shows signs of generation" instead of one overall percentage.
• "What if". Compare the score of a passage before and after an edit.
• Text type (news, review, post, article) with its own thresholds, once genre corpora exist.
• An optional language-model signal of the Binoculars class, but only if evaluation shows a clear gain, for example on non-native writers and paraphrases. Otherwise we leave it out.

We do not promise "indistinguishable from human" and we are not building a tool to evade checks: the goal is livelier text and a more honest assessment. Treat the result as a hint about where to reread more carefully, not as a verdict.

What to read next

Start with how to tell if a text was written by AI for the signs you can see yourself, then why modern AI detectors get it wrong for the limits. For recent research see what is new in AI detection. To try it on a text, open the checker.