AI Content Detector

A fully client-side detector that estimates the probability a piece of English or Chinese text was AI-generated. It blends 15 weighted signals (perplexity, burstiness, AI-phrase density, vocabulary richness, punctuation pattern, paragraph uniformity, structure repetition, transition density, readability, POS pattern, grammar errors, sentence complexity, emotional variability, concrete/abstract ratio, numeric density) into a single 0-99% score with a low/medium/high confidence label.

1. On the Analyze tab, paste at least 50 characters of text into the input (Analyze is disabled below that threshold). 2. Click Analyze to render the headline probability, confidence, detected model hint, text statistics (characters/words/sentences/paragraphs/language), and a per-metric breakdown bar chart. 3. Optionally click Export Report to download a JSON snapshot. 4. Use Highlight to color each sentence by its score, Compare to score two texts side by side, History to revisit past analyses, and Batch to upload .txt/.md files.

Language is auto-detected from the CJK character ratio (zh / mixed / en). Perplexity uses a bigram entropy model with a sigmoid centered at 7.5 for Chinese and 4.3 for English. Burstiness measures the coefficient of variation of sentence lengths. The AI-phrase signal counts tiered tell-tale phrases (weight 3.0 for phrases like "delve into", 1.5 for "furthermore", 0.5 for "in conclusion") normalized by sentence count. Model detection is a separate phrase match: "delve into"/"a testament to" suggest GPT, while "stemming from"/"grounded in" suggest Claude.