Back to home

AI Text Detector

Text classifier that distinguishes between human writing and AI-generated text

AI Text Detector

Overview

A DeBERTa-v3 model fine-tuned to classify whether an essay was written by a human or generated by an AI model.

The problem

Modern language models produce fluent writing that easily bypasses basic n-gram or perplexity checks, making it difficult to detect AI-generated submissions without running into high false positive rates.

Notes

Trained DeBERTa-v3 on a dataset of encyclopedia articles and paired with generated text from AI models like GPT-Neo-1.3B. The trained checkpoint is published on Hugging Face.

Key Metrics

Reached 97.4% validation accuracy and a 0.989 ROC-AUC score, keeping false positives low on formal human writing.

Tech Stack

Python, PyTorch, Hugging Face Transformers, Natural Language Processing(NLP), NumPy, GPT-Neo-1.3B-Model, DeBERTa-v3-Model

Next steps

Build an interactive web demo with token perplexity highlights and test performance against frontier models like GPT-5, Gemini 3 and Claude 5