The game moved. So did the model.
A year ago the hard problem was not accusing real people. tropa-1 made a big dent in it. tropa-2 keeps it: in our August benchmark it wrongly flags 1 of 251 human texts across twelve domains — recipes, poems and short emails included.
But the texts we see today are not the texts of a year ago. Raw model output is rare. What arrives is AI text that has been run through a humanizer — or, cheaper still, sprinkled with characters no human can see and no tokenizer can ignore. A detector that only knows "AI or human" has no vocabulary for either.
tropa-2 has the vocabulary. Three things are new.
It names the kind of AI text
Four classes instead of one score: human, raw AI, lightly edited AI, humanized AI. When the humanized class dominates, the result says so.
It sees the invisible characters
Homoglyphs, zero-width spaces, exotic whitespace — stripped before scoring, and counted, so the trick itself becomes a finding.
Its score means what it says
Calibrated. A 62 is a 62: about six of ten texts scoring there are AI. No more mental conversion from "confidence" to probability.
The numbers, with the false-positive rate nailed to the floor
Every figure below was measured at the same operating point: 0.5% of human texts wrongly flagged. That is the number we refuse to trade. Recall is what the model earns on top of it.
| Held-out test set, August 2026 | tropa-2 |
|---|---|
| Human-written texts wrongly flagged | 0.5% |
| Raw output of current models caught — GPT-5.6, Claude Opus 5 and Fable 5, Gemini 3.6 and others | 97% |
| AI text run through a dedicated humanizer caught — the hardest case there is | up to 78% |
| "Possibly humanized" raised on raw AI text — a false hint | 0% |
| "Possibly humanized" raised on genuinely humanized text | ~40% |
One thing to keep in mind when you compare this with the tropa-1 announcement: this is a harder exam. Newest-generation models, real humanizer products instead of model-made paraphrases. We chose it because it is the test our users actually face.
How to read a result
The site shows the score itself — "62% AI" — and a label for the band it falls in. Because the score is calibrated, the bands are nothing more than what the probability says.
A badge that only speaks when it is sure
The possibly humanized badge is built to be believed. It appears only when the model puts most of the AI signal in the humanized class and the overall score is at least 40. In testing it never once fired on raw AI text.
The price of that precision is reach: it catches about four in ten humanized texts. So the badge means something, and its absence means nothing — which is why you will never see a "not humanized" label or a percentage next to it, here or in the API. When it shows up, open the sentence view. That is where the story is.
The characters you cannot see
A surprising share of "undetectable" AI text is ordinary AI text with a Cyrillic "а" swapped in for the Latin one, or zero-width spaces between letters. To a reader it is identical. To a tokenizer it is a different language, and detectors that score the raw bytes fall apart. tropa-2 normalises the text first, scores the clean version, and reports what it removed — homoglyphs, zero-width characters, full-width characters, unusual spaces. When that warning appears, the obfuscation is usually the most interesting thing about the document.
If you build on the API
/api/v1/detect/textnow returnsscore,verdict,humanizedHint,inputAnomaliesandsubclassProbs. Nothing you already parse has changed;isAIstill means a score above 50.- Not ready to recalibrate? Send
model: "tropa-1"and stay on the previous generation at the same price.mode: "depth"gets you real per-sentence scores. - Responses now take one to two seconds instead of a few hundred milliseconds. Give your client a 30-second timeout.
- Field reference in the API docs.
What it does not do
- Catch every humanized text. Against the best tools roughly one in four still gets through. Anyone quoting "99% accuracy" is measuring raw output only, or not telling you which.
- Cover every language and length. These numbers are for English text of at least about fifty words.
- Always return every field. Under load a fallback path returns the score without the humanizer hint or class distribution — treat those fields as optional.
- Replace a human. A result is evidence. Show the flagged sentences to the person concerned and let someone decide — above all in academic settings.
Try it on the text you were not sure about
tropa-2 is live on every text and document check on this site and in the API. Paste the one that made you wonder.
Test tropa-2 Free