AI Text Detector

Single-model perplexity detectors are both inaccurate and unfair — GPT-2 alone flags ~50% of human text as AI and penalizes non-native English writers at up to 26× the rate of native speakers. This tool uses multiple signals, shows its reasoning, and measures fairness explicitly.

• 01Binoculars cross-perplexity (gpt2 + distilgpt2) cancels topic bias for the primary verdict.
• 02GPT-2 perplexity and NLTK n-gram signals provide transparent secondary evidence.
• 03Ensemble mode returns a structured verdict: confidence, per-signal metrics, and narrative reasoning.
• 04Local-first — no required external APIs; Dockerized for reproducibility.

1. Text input
2. Binoculars · gpt2 + distilgpt2 cross-perplexity
3. GPT-2 perplexity + NLTK n-gram signals
4. Ensemble verdict + reasoning
5. Streamlit UI · Dockerized

• Fairness: measuring and reporting false-positive rates by population, not just headline accuracy.
• Explaining a probabilistic verdict instead of emitting an opaque score.

• Binoculars cross-perplexity detection
• Per-population fairness evaluation
• Structured verdict + narrative reasoning
• Local-first, no external APIs required
• HC3 reproducibility as a test gate