Why we optimized for false positives first
The worst thing an AI detector can do is not missing AI text — it's accusing a real person of using AI when they didn't. A missed AI text is an inconvenience; a false accusation can cost a student their academic standing or a writer their job. So when we trained tropa-1, the first question wasn't "how much AI text does it catch?" but "how often is it wrong about humans?"
To measure that, we scored 1,314 guaranteed-human texts — pre-2017 arXiv abstracts from twelve scientific domains, written years before modern language models existed, freshly fetched and deduplicated against all training data.
Human texts wrongly flagged as AI (threshold 0.5)
Wrongly flagged at strict threshold (0.1)
The false-positive numbers in full
| On 1,314 verified human texts | o.3 (previous) | tropa-1 |
|---|---|---|
| Flagged as AI (threshold 0.5) | 9.7% (127 texts) | 1.2% (16 texts) |
| Flagged at strict threshold (0.1) | 31.7% | 2.0% |
| Mean AI-score assigned to human text | 0.138 | 0.014 |
The mean-score row is the one we care about most. tropa-1 is strongly bimodal: human writing clusters near 0.01, raw AI output near 1.00. The previous model smeared human texts across the whole 0–0.9 range — which is exactly how borderline cases and false accusations happen. On the external benchmark's website texts, o.3 flagged almost every third human page (32%); tropa-1 flags 0.5%.
Humanized AI text — compared honestly
"Humanized" AI text — output paraphrased by tools like Rephrasy to evade detectors — is the hardest case. Here a naive comparison would flatter the old model: at a fixed threshold, o.3 catches a lot of paraphrased AI simply because it flags every fifth human too. The honest way to compare is at an equal false-positive budget: tune both models to wrongly flag the same (small) share of humans, then measure who catches more AI.
Humanized AI text caught at equal 1% false-positive rate (359 paraphrased texts)
| False-positive budget | o.3 | tropa-1 |
|---|---|---|
| 0.5% of humans wrongly flagged | 76.3% | 83.6% |
| 1% | 79.1% | 86.4% |
| 2% | 80.2% | 89.7% |
tropa-1 wins at every operating point — it doesn't trade false accusations for recall, it improves both at once.
Unedited AI text: still 100%
Both models detect 100% of raw, unedited output across all twelve tested generators — including GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro and, notably, two Qwen models that tropa-1 never saw during training. tropa-1 holds that 100% even at a very strict 0.9 threshold, which is what makes the low false-positive rate possible in practice.
Methodology
- Training: tropa-1 was fine-tuned on 442,000 texts from 16 sources, including outputs from current 2026 models (Claude Opus 4.8, gpt-5.4-mini, Gemini 3.1 Pro and others).
- Evaluation: everything reported here is fully held out — a 739-text external benchmark (200 human, 539 AI from 12 models, 359 of them paraphrased) plus 1,314 fresh pre-2017 human abstracts. None of it appears in training data.
- Production parity: all figures were measured on the exact model and code path now serving this site — not a lab configuration.
Honest caveats
- Our human test corpus is English, at least 50 words per text, and mostly formal writing (papers, articles, reviews). Short casual text and non-English content are out of scope for these numbers.
- "Humanized" here means paraphrased by the twelve listed models. Dedicated humanizer tools like Rephrasy may behave differently — and roughly 14–16% of humanized text still slips through at production settings. No detector catches everything; anyone claiming otherwise is selling you something.
- Detection results are evidence, not proof. That's why we show sentence-level scores: a flag should start a conversation, not end one — especially in academic settings.
tropa-1 is live now
Every text and document check on this site already runs on tropa-1 — via the web detector and the API. Try it on your own writing.
Test tropa-1 Free