Why we replaced tropa-2
tropa-2 caught almost all raw AI text. But when we scored a much larger set of human writing, it flagged too much of it: 3.02% of 239,490 held-out human texts at its production threshold, and 6.07% of news articles. Those are real people being told their writing looks like AI.
For tropa-3 we added training data in three places: more human writing across lengths, with extra weight on the domains where tropa-2 misfired most (news, speeches, student and learner essays, biomedical papers, technical docs); texts from the newest models; and output from more commercial humanizers.
False positives on 239,490 human texts
None of our models saw these texts in training. They cover web pages in seven languages, news, speeches, biomedical papers, docs, e-mail, forums and reviews. At the live threshold, a score of 90 or more, tropa-3 flags 0.10% of them.
To compare the two models fairly, the table sets each one to the threshold where it catches 97% of raw AI text and counts how many human texts it flags there.
| Human texts flagged at 97% raw-AI detection | tropa-2 | tropa-3 |
|---|---|---|
| All 239,490 texts | 1.99% | 0.32% |
| News | 6.07% | 0.30% |
| Speeches | 4.01% | 0.22% |
| Biomedical | 2.63% | 0.35% |
| Web pages | 1.68% | 0.35% |
| Technical docs | 0.90% | 0.10% |
Detection on held-out test sets
Here both models are set to flag 1% of human texts, and we measure how much AI text each one catches at that point.
| AI text detected at 1% false positives | tropa-2 | tropa-3 |
|---|---|---|
| Raw AI text | 90.2% | 98.6% |
| Raw text from the newest models | 89.8% | 99.1% |
| Paraphrased AI text | 48.9% | 85.2% |
| Humanized AI text | 28.5% | 94.2% |
These test sets come from the same generators and humanizers as the training data. No text overlaps, but this is the model's home ground. On the public benchmark, scored at the live threshold of 90, tropa-3 caught 58% of humanized text, not 94%.
The newest models
Share of raw text detected per model at a 1% false-positive rate, on the same held-out test sets.
| Model | tropa-3 |
|---|---|
| GPT-6 Sol | 100% |
| GPT-6 Luna | 100% |
| GPT-6 Astra | 100% |
| Claude Fable 5.1 | 100% |
| Claude Opus 5.5 | 98.0% |
| Gemini 3.8 Flash | 100% |
| Grok 4.7 | 100% |
| DeepSeek V4.1 Flash | 98.0% |
| Qwen 3.8 Max | 97.6% |
| GLM 5.3 | 95.9% |
| MiMo V2.6 Pro | 100% |
Compared with other detectors
In our September 2026 benchmark, on 513 texts that every detector scored, tropa-3 detected 95.6% of raw AI text and flagged none of 247 human texts. GPTZero detected 61.6% and Pangram 44.0%, also with no human text flagged. The benchmark page has the paraphrase, humanizer and "Mixed" results, and every text is public.
How to read a result
The site shows a score from 0 to 100 and a label for the band it falls in.
If you build on the API
model: "tropa-3"is the default. You do not need to send it.- Requests that name
model: "tropa-2"are served by tropa-3, so integrations that pinned tropa-2 keep working. model: "tropa-1"is still available at the same price if you need time to recalibrate thresholds.- The response format has not changed. Field reference in the API docs.
Known weaknesses
- Some humanizers still get through. At a score of 90, tropa-3 caught 25.5% of UndetectedGPT texts and 31.2% of WalterWrites texts in the public benchmark.
- A few small domains got slightly worse than with tropa-2: recipes, wikiHow articles and e-mail. The samples are small and false positives stay under 2.5%.
- It reads the first 512 tokens of a text. For a long document, check sections separately.
- A result is evidence, not proof. Show the flagged sentences to the person concerned and let someone decide, above all in academic settings.
FAQ
What changed from tropa-2 to tropa-3?
Mainly the false-positive rate. At a score of 90, tropa-3 flags 0.1% of 239,490 held-out human texts. tropa-2 flagged 3.02% of the same texts at its production threshold. tropa-3 also catches more paraphrased and humanized AI text on our held-out test sets.
I pinned "tropa-2" in my API integration. Do I need to change anything?
No. Requests that name tropa-2 are now served by tropa-3. If you want to stay on an older generation while you recalibrate thresholds, send model "tropa-1".
How does tropa-3 compare with GPTZero and Pangram?
In our September 2026 benchmark, on 513 texts all detectors scored, tropa-3 detected 95.6% of raw AI text, GPTZero 61.6% and Pangram 44.0%. tropa-3, GPTZero and Pangram flagged none of the 247 human texts. The AI texts come from models our detector was trained on, so treat our numbers as an upper bound.
Can tropa-3 detect humanized AI text?
Often, not always. In the public benchmark it caught 58% of humanized text (57.7% of 286 texts). UndetectedGPT (25.5%) and WalterWrites (31.2%) are the hardest humanizers for it.
How much text does tropa-3 read?
It reads the first 512 tokens of a text, a few hundred words. Longer documents are scored on their opening unless you check sections separately.
Try it on the text you were not sure about
tropa-3 runs every text and document check on this site and in the API. Paste the one that made you wonder.
Test tropa-3 Free