AI Text Detector

Fine-tuned DistilBERT classifier (66M params) for detecting AI-generated text, achieving 98% validation accuracy.

Fine-tuned DistilBERT (66M parameters) on the GPT-wiki-intro dataset (~150k human vs. AI-generated text pairs) for binary text origin classification, achieving 98% validation accuracy after 3 epochs.

Key features:

  • Full ML pipeline: AdamW optimizer with linear warmup (10% steps), gradient clipping, mixed-precision training (float16), and best-model checkpointing by validation accuracy
  • Apple Silicon MPS GPU acceleration support
  • Gradio web interface with live confidence scores, probability bar chart, and example texts
  • CLI prediction mode and importable Python API (from predict import TextDetectorPredictor)
  • Documented limitations: domain-specific to Wikipedia-style text, unreliable on short texts (<100 chars), trained on GPT-2 era output

Developed as part of AI literacy research at VU Amsterdam, with a focus on responsible deployment.

Tech stack: Python, PyTorch, Hugging Face Transformers, Gradio, scikit-learn