Dang tai... / Loading...
We publish exactly what this tool can and CANNOT do — including its weaknesses. The result is an estimate of flagging RISK, not a verdict that says "this is AI".
The risk score combines three signals. Each is only a proxy (an indirect clue), none is proof. Their weights were re-calibrated to match the benchmark below.
These are Phase 0 spike results: a minimal engine running on the Qwen2.5-0.5B model, measured on 107 Vietnamese + English samples seeded with hard cases (humanized text, hybrid, short text, formal/ESL writing).
AUC closer to 1.0 is better; 0.5 is random guessing. Perplexity alone scores AUC 0.573 (near a coin flip); stylometry scores AUC 0.754 (better, but still not reliable enough to judge). The false-positive rate (innocent human text flagged) at the shared threshold is 10.3%.
| AI text group | Perplexity alone | Combined (ensemble) | Note |
|---|---|---|---|
| Clean AI (unedited) | 56.7% (17/30) | 100.0% (30/30) | Easy case — caught well |
| Humanized AI | 0.0% (0/20) | 0.0% (0/20) | Core blind spot |
| Short AI text | 10.0% (1/10) | 70.0% (7/10) | Weak, low confidence |
| Hybrid (human + AI mixed) | 0.0% (0/8) | 0.0% (0/8) | Core blind spot |
The moment text is rewritten (humanized) once, or mixes human + AI (hybrid), recall drops to 0% across all three methods — none of our methods catch it. We state this plainly instead of hiding it, because it is the truth.
The gauge shows flagging risk from 0-100%, split into three color bands. The label is about RISK, not about "percent AI".
Measures writing-style features (sentence length, phrase repetition, word rhythm). Better than perplexity, but still far from reliable enough to judge.
Measures how "surprising" the text is to a language model. Barely better than a coin flip — human and AI distributions overlap heavily, so it no longer carries high weight.
Asks an LLM "is this AI?". Flags too many innocents (23% of human text, especially formal and ESL writing), so it is turned OFF by default.
Every result carries a confidence level (high / moderate / low). Confidence is lowered automatically when the text falls into the cases where the benchmark shows the engine is least stable:
Low confidence does NOT mean "definitely human" — it means "do not rely on this number to conclude anything".
The detector only estimates the chance another system flags your text. Because humanized/hybrid text yields 0% recall, we cannot — and must not — declare "this is AI".
Thresholds are calibrated tight to keep false flags low. When in doubt, the system leans toward assuming a human wrote it — because a false accusation harms more than a miss.
This tool included. That is the genuine limit of every detector today, not an excuse. We publish it so you never put your trust in the wrong place.
We do not promise to "beat Turnitin/GPTZero", we do not promise "100% undetectable", and we do not sell ourselves as a cheating tool. Being transparent about the limits is our advantage: you know exactly how much this number is worth.
Try it free. Remember this is a reference risk estimate meant to protect you from wrongful suspicion — not a sentence.