Aletheia – AIGC Text Detector API

中文文档

A FastAPI service for detecting AI-generated text, based on AIGC_text_detector and DivEye.

Quick Start (Docker)

# Using the pre-built image from GHCR
docker compose -f docker-compose.yml up

The API will be available at http://localhost:8000.

Chrome Extension

A Chrome extension is available that automatically detects AI-generated text on article pages you visit.

Install

1. Go to the Releases page and download the latest aletheia-extension-*.zip.
2. Unzip the file.
3. Open chrome://extensions in Chrome, enable Developer mode.
4. Click Load unpacked and select the unzipped folder.

Configure

Click the extension icon to open the popup. Set the API URL to point to your running Aletheia instance (default: http://localhost:8000). For more options (detection strategy, domain whitelist/blacklist), go to the extension's Settings page.

The extension will automatically detect article content on pages you visit and show a floating badge with the result.

Local Development

Requires uv.

# Install dependencies
uv sync

# Run the server
uv run uvicorn app.main:app --reload

Or with Docker Compose:

docker compose up --build

API

POST /detect

Detect whether text is human-written or AI-generated.

Request body:

Field	Type	Required	Default	Description
text	string	yes		Text to detect
lang	string	no	auto-detect	"zh" for Chinese model, anything else for English
model_id	string	no	(by lang)	HuggingFace model ID; overrides lang
strategy	string	no	"truncate"	"truncate", "sliding_avg", "sliding_weighted_avg", or "sliding_vote"
early_stop	bool	no	false	Stop early when confidence is high enough (sliding strategies only)

Strategies:

• truncate – Truncate to 512 tokens. Fast, single forward pass.
• sliding_avg – Sliding window (512 tokens, stride 256). Average softmax scores across windows.
• sliding_weighted_avg – Sliding window. Confidence-weighted average: chunks with higher confidence contribute more.
• sliding_vote – Sliding window. Majority vote on predicted label across windows.

Note: For most use cases, truncate is sufficient. For long texts where you want higher accuracy, use sliding_weighted_avg with early_stop enabled — it gives better results than plain averaging by weighting high-confidence chunks more heavily, and early stopping avoids unnecessary computation when the result is already clear.

Example:

curl -X POST http://localhost:8000/detect \
  -H "Content-Type: application/json" \
  -d '{"text": "This is a sample text to detect."}'

Response:

{
  "label": "human",
  "score": 0.98,
  "model_id": "yuchuantian/AIGC_detector_env3",
  "detected_lang": "en",
  "num_chunks": 1
}

GET /health

Health check endpoint.

Default Models

Language	Model ID
English	yuchuantian/AIGC_detector_env3
Chinese	yuchuantian/AIGC_detector_zhv3

You can use any HuggingFace *ForSequenceClassification model by passing model_id in the request.

Detection Methods

AIGC Text Detector

Transformer-based sequence classifiers fine-tuned for AI-text detection. Source: YuchuanTian/AIGC_text_detector.

DivEye

DivEye detects AI-generated text using surprisal-based statistical features that capture how unpredictability varies throughout a text. Human writing exhibits greater variability in lexical and structural unpredictability compared to LLM outputs. These features feed an XGBoost classifier, making it interpretable and robust to paraphrasing attacks.

Advik Raj Basani, Pin-Yu Chen. Diversity Boosts AI-Generated Text Detection. TMLR 2026.

Source: IBM/diveye