Key facts
- Price
- Free, 20 checks per day
- Ads
- None
- Sign-up
- Not required
- Language
- Vietnamese (calibrated and validated)
- Length
- 50 to 800 words per check
- Accuracy
- "AI" verdicts right 98% of the time; 0.8% of human texts misflagged (500 test articles)
- Method
- Binoculars with two open Qwen2.5-0.5B models on a DigiAgent server
- Privacy
- Text content is not stored; only a hash and score for 30 days
How to use
- Paste the Vietnamese text to check, 50 to 800 words.
- Click "Check" (or Ctrl+Enter). Results usually take 5 to 10 seconds.
- Read the AI likelihood and the highlighted sentences: the darker, the more AI-like.
How does it work?
AI-written text tends to be more predictable than human writing. The tool uses the Binoculars method (Hans et al., 2024): two small open-source language models (Qwen2.5-0.5B base and chat) read the text and compare how predictable it is with what an AI model would typically produce, then convert that into a percentage.
The models run on a DigiAgent server in Singapore. Text is never sent to OpenAI, Google or any third party.
Accuracy
Measured on 500 held-out Vietnamese news articles (half by journalists, half by GPT, Gemini and Grok) from the ICCIES 2025 dataset:
- When the tool says "likely AI", it is right 98% of the time.
- Only 0.8% of human articles were wrongly flagged as AI.
- When the tool says "likely human", it is right 96% of the time.
- Overall separation (AUROC) is 0.94 on a 0 to 1 scale.
Limits to know
No AI detector is always right. AI text edited by people, very short text, bureaucratic or list-like text and machine translation are easily misjudged, and research shows detectors often misjudge non-native writers.
Treat the result as a signal, never as proof for accusing or penalising anyone.
Frequently asked questions
Is it free?
Yes. Free, no ads, no sign-up. To prevent abuse, each person gets 20 checks per day.
Is my text stored?
No. Text is analysed and discarded. Only a hash of the text and its score are kept for 30 days so re-checking the same text is instant; the text cannot be recovered from the hash.
Which AI models can it detect?
The method is model-agnostic, so it applies to ChatGPT, Gemini, Claude, Grok, Copilot and other AI writers. Accuracy was measured mainly on GPT, Gemini and Grok output.
Why at least 50 words?
Predictability statistics are unreliable on very short text. Longer text (up to 800 words) gives steadier results.
Does it work for English?
It is calibrated and validated for Vietnamese. English text can be analysed but accuracy has not been measured yet.
Sources
- Hans et al. (2024), Spotting LLMs With Binoculars: Zero-Shot Detection of Machine-Generated Text
- ICCIES 2025: Vietnamese news human/AI dataset (CC BY 4.0)
- Liang et al. (2023), GPT detectors are biased against non-native English writers
- Qwen2.5-0.5B model card (Apache 2.0)
Updated: