AI Detection Glossary
Every term you'll meet in the AI-detection world — explained simply.
AI Detector
A tool that estimates how likely a text was generated by a language model, based on statistical and stylistic signals. Output is a probability, never a certainty.
AI Humanizer
A tool that rewrites machine-sounding text into a natural human voice — varying rhythm and removing robotic patterns — while preserving the original meaning.
Perplexity
A measure of how surprised a language model is by each next word. Human writing tends to have higher perplexity (more surprising word choices) than machine writing.
Burstiness
How much sentence length and complexity vary across a passage. Humans naturally alternate short and long sentences; AI text is often suspiciously uniform.
Lexical Diversity
The ratio of unique words to total words (type-token ratio). Low diversity — the same words recycled again and again — is a common machine-writing signal.
AI Tells
Stock phrases that language models overuse: "furthermore", "in today's world", "it is important to note". A high density of tells raises AI-likelihood scores.
False Positive
When a detector flags genuinely human writing as AI-generated. The most damaging kind of detector error — and the reason scores must never be treated as proof.
LLM (Large Language Model)
The technology behind ChatGPT, Claude, and Gemini: a neural network trained on vast text to predict the next word, which lets it generate fluent language.
Register
The formality level of writing — academic, professional, casual. Good humanizing keeps meaning while shifting register; in Arabic this spans فصحى رصينة to فصحى مبسّطة.
Academic Integrity
The ethical rules of scholarship: submitting your own work and disclosing assistance when required. Style tools must be used within your institution's policies.