A multimodal AI content detection pipeline that analyses audio, video, and text to determine whether content was generated by an AI or created by a human.
Checker AI runs three independent analysis modules in parallel and combines their scores using a weighted confidence matrix:
| Module | Weight | What it checks |
|---|---|---|
| Audio | 45% | Voice cloning, synthetic pitch, unnatural silence patterns |
| Visual | 35% | Deepfake artifacts, spatial anomalies, generative blending |
| Text | 20% | Sentence burstiness, LLM filler phrases, lexical diversity |
The final result is a single prediction with a percentage score.
- Node.js 18+
- FFmpeg installed on your system (
sudo apt install ffmpeg/brew install ffmpeg)
git clone https://github.com/your-username/checker-ai.git
cd checker-ai
npm installCreate a .env file in the project root:
GEMINI_API_KEY=your_gemini_api_key_here
ELEVENLABS_API_KEY=your_elevenlabs_api_key_here # optional
AUDIO_CLASSIFY_ENDPOINT=http://localhost:8000/classify-audio # optionalGet your free Gemini API key at aistudio.google.com.
Never commit your
.envfile. It is already listed in.gitignore.
Analyse a video file:
node pipeline.js --file ./video.mp4Analyse text:
node pipeline.js --text "Paste your text here"Analyse a video and its transcript together:
node pipeline.js --file ./video.mp4 --text "The transcript of the video"Get raw JSON output (for integrations):
node pipeline.js --text "Some text" --json┌─────────────────────────────────────────┐
│ CONTENT ANALYSIS RESULT │
├─────────────────────────────────────────┤
│ Prediction: [ AI ] Very Likely AI │
│ AI Score: ████████░░ 78% AI │
├─────────────────────────────────────────┤
│ Checked: Audio, Video, Text │
└─────────────────────────────────────────┘
What we found:
• Voice pitch is unnaturally consistent
• Facial warping patterns associated with deepfakes
• Writing style closely matches AI-generated text
| Score | Verdict |
|---|---|
| 85 – 100% | Almost Certainly AI |
| 65 – 84% | Very Likely AI |
| 50 – 64% | Likely AI |
| 35 – 49% | Likely Human |
| 15 – 34% | Very Likely Human |
| 0 – 14% | Almost Certainly Human |
All modules fall back to local heuristics when no API key or endpoint is available, so the pipeline always produces a result even without credentials.
MIT — see LICENSE