Glossary
·AI Humanizer & AI Detection Glossary
Every term you need to understand how AI detectors and humanizers work, defined in one sentence each.
- Perplexity
- A measure of how predictable a piece of text is to a language model. Low perplexity means each word closely matches what a model would statistically expect next — a pattern common in raw LLM output and one of the two core signals AI detectors score.
- Burstiness
- The variation in sentence length and rhythm across a passage. Human writing tends to alternate short, punchy sentences with longer ones; unedited AI output tends toward more uniform sentence lengths, producing lower burstiness.
- AI likelihood score
- The percentage or band (e.g., 'likely AI', '82% AI') a detector assigns to a passage. It is a probability estimate from a classifier, not proof of authorship, and can vary significantly between different detection tools.
- False positive (AI detection)
- When a detector flags genuinely human-written text as AI-generated. Formal academic writing, non-native English prose, and heavily self-edited text are the most commonly reported false-positive categories.
- AI humanizer
- A tool that rewrites AI-generated text to increase natural sentence variation and reduce templated phrasing, lowering the likelihood scores assigned by AI detectors while preserving the original meaning.
- Text watermarking
- A technique some AI providers embed in generated text (via statistically biased token selection) intended to make machine-generated content identifiable later. Watermarking is separate from, and generally more reliable than, third-party statistical detectors.
- Paraphrase-origin detection
- A specialized detection signal that identifies text run through a paraphrasing or spinning tool (like QuillBot) by spotting unnatural synonym substitution patterns that survive after a spin, distinct from raw AI-generation detection.
- Ensemble detector
- An AI detection system that combines multiple underlying models or signals (rather than a single classifier) to produce a final score, generally intended to reduce single-model false positives.
- Token predictability
- How statistically expected each individual word (token) is given the preceding context, as scored by the language model a detector uses internally. It is the mechanical basis of perplexity scoring.
- Meaning-safe rewrite
- A rewriting approach that changes sentence structure, rhythm, and word choice while deliberately preserving the original facts, citations, and arguments — the approach Neonhumanizer uses instead of altering content meaning.
- Institutional AI likelihood band
- The tiered risk categories (e.g., 'low', 'moderate', 'high') that academic-focused detectors like Turnitin assign to submissions, used by instructors as one input among several rather than automatic proof of misconduct.
- Content fingerprint
- The distinctive statistical pattern (word choice frequency, sentence structure, transition-word usage) a specific language model tends to leave in its output, which detectors are trained to recognize.