AI Content Detector

Paste any passage — get an instant verdict. Fully client-side inference, no text ever leaves your browser.

🚀 Live Demo · API Docs · Benchmark

Note: AI-text detection is probabilistic. Results are signals for review, not proof of authorship. Short, paraphrased, translated, or heavily edited passages can be difficult to classify reliably.

---

✨ Features

• Privacy-first — static frontend scores text entirely in-browser; no text is uploaded to any server
• 11-signal diagnostic grid — perplexity, burstiness, type-token ratio, POS entropy, readability, and more
• Sentence-level highlighting — see exactly which sentences push the score toward AI or human
• Three-tier verdict — calibrated probability bands (Likely Human / Mixed / Likely AI)
• No build step — single index.html file, served anywhere
• Optional FastAPI backend — same calibrated pipeline exposed as a REST API for integrations
• Drag-and-drop file loading — plain-text files supported directly in the browser
• Session history — recent analyses persist in tab session storage, never sent anywhere

---

📊 Verdict Bands & Benchmark

Consistent across browser and backend:

Verdict	Calibrated AI Probability
Likely Human	≤ 0.30
Mixed / Uncertain	> 0.30 and < 0.70
Likely AI-generated	≥ 0.70

Local benchmark (30 iterations per sample, 3 warmups):

Sample	Predicted	AI Probability	Median latency
AI text	Likely AI	0.8962	42ms
Human text	Likely Human	0.0719	52ms
Mixed text	Mixed	0.6398	91ms

Accuracy: 100% on benchmark suite · p95 latency: 92ms

---

🛠 Tech Stack

Layer	Technology
Frontend	Vanilla HTML / CSS / JavaScript — zero build step
Browser scoring	Exported LinearSVC weights + reference LM (model.json, lm.json)
Backend API	FastAPI, scikit-learn, NLTK, NumPy
Model	LinearSVC with 11 statistical features, calibrated probability output
Deployment	Vercel (frontend) · Render / Docker (backend)

---

📁 Project Structure

ai-content-detector/
├── index.html                  # Full frontend — single file, no build step
├── model.json                  # Exported LinearSVC weights (browser-safe)
├── lm.json                     # Reference language model (browser-safe)
├── ui-state.js                 # Frontend state module
├── render.yaml                 # Render deployment config
├── Dockerfile                  # Root / Hugging Face Spaces image
├── test/
│   └── ui.test.js              # Frontend regression tests
├── benchmarks/
│   ├── benchmark.py            # Inference latency + accuracy benchmark
│   └── local_results.json      # Latest benchmark results
├── docs/
│   ├── API.md                  # Full API contract
│   ├── BENCHMARK.md            # Benchmark methodology
│   ├── INTEGRATION.md          # Integration guide
│   └── SECURITY_AUDIT.md       # Security audit notes
└── backend/
    ├── app.py                  # FastAPI service
    ├── inference.py            # Model loading, calibration, scoring
    ├── features.py             # 11-feature extraction pipeline
    ├── train.py                # Classifier training + artifact export
    ├── reference_lm.py         # Reference language model construction
    ├── export_lm_json.py       # LM → browser JSON export
    ├── export_weights.py       # Classifier → browser JSON export
    ├── requirements.txt
    ├── tests/test_app.py       # FastAPI + inference regression tests
    └── model/                  # Trained model artifacts

---

🧠 How It Works

Browser Inference (no server needed)

Text input
    ↓
11 statistical features extracted locally
  • Perplexity against reference LM
  • Burstiness (sentence-length variance pattern)
  • Type-token ratio, avg sentence/word length
  • Stopword ratio, punctuation density
  • Repetition score, POS entropy, readability
    ↓
Exported LinearSVC weights applied in-browser
    ↓
Calibrated probability → Verdict + sentence-level highlights

Everything runs in-browser from model.json and lm.json. No API call is made.

Optional Backend (API consumers)

The FastAPI backend runs the same feature pipeline server-side with additional endpoints for batch analysis, model metadata, and health/readiness checks.

---

🚀 Getting Started

Run the frontend

npx --yes serve .

Open http://localhost:3000 — no install, no build.

Run the backend locally

cd backend
pip install -r requirements.txt
uvicorn app:app --reload --host 127.0.0.1 --port 8000

See docs/API.md for the full API contract.

Run the full test suite

python3 -m compileall -q backend
node --check ui-state.js
npm test
PYTHONPATH=backend pytest -q backend/tests

---

🐳 Docker

docker build -t ai-content-detector .
docker run -p 8000:8000 ai-content-detector

---

🔌 REST API

POST /analyze

curl -X POST http://localhost:8000/analyze \
  -H "Content-Type: application/json" \
  -d '{"text": "The concept of emergence refers to properties..."}'

Response:

{
  "verdict": "likely_ai",
  "ai_probability": 0.87,
  "confidence": "high",
  "features": {
    "perplexity": 11240.5,
    "burstiness": 4821.3,
    "type_token_ratio": 0.74
  },
  "sentence_scores": [...]
}

See docs/API.md for all endpoints including /analyze/batch, /model/info, and /health.

---

🚨 Known Limitations

• Short passages (< 150 words) produce less reliable scores — the feature pipeline needs enough signal to work with
• Heavily paraphrased, translated, or domain-specific text can fool the classifier
• The rate limiter is process-local — multi-worker or multi-replica deployments should add a shared edge limiter
• Model metrics reflect training-distribution accuracy, not a guarantee on unseen writing styles

---

📄 License

MIT