AI Writing Glossary
The vocabulary of AI writing and AI detection, defined in plain English — so when a detector report says "low perplexity", you know exactly what it means.
AI humanizer
An AI humanizer rewrites machine-generated text so it reads like a person wrote it — by changing sentence structure and rhythm, not just swapping synonyms.
AI detector
An AI detector estimates whether text was written by a human or a language model, using statistical signals like perplexity and burstiness plus trained classifiers.
Perplexity
Perplexity measures how predictable text is to a language model. Low perplexity — predictable word choices — is the classic statistical signature of AI-generated text.
Burstiness
Burstiness measures variation in sentence length and structure. Humans write in bursts; AI models keep a steady rhythm — making low burstiness a machine tell.
Large language model (LLM)
A large language model is a neural network trained on vast text to predict the next word — the technology behind ChatGPT, Claude, Gemini and modern AI writing.
Token
Tokens are the word-fragments language models actually read and write. What they are, why limits are counted in them, and what they have to do with detection.
Temperature
Temperature controls how adventurously a language model picks its next word — low is safe and repetitive, high is varied and risky. It shapes how detectable output is.
AI watermark
AI watermarking hides a statistical signature in generated content at creation time. How it works for text, where it's deployed, and why it's fragile.
Stylometry
Stylometry identifies authors by their measurable writing habits — centuries older than AI detection, and the foundation it quietly stands on.
Paraphrasing
Paraphrasing restates text in different words while keeping the structure — which is exactly why paraphrased AI text still gets flagged by detectors.
Prompt engineering
Prompt engineering is crafting model instructions for better output. What it can do for writing quality — and why it can't make text undetectable on its own.
Hallucination
AI hallucination is when a model states false information fluently — invented citations, fake facts, plausible nonsense. Why it happens and why no humanizer fixes it.
Training data
Training data is the text corpus a language model learns from — and the reason AI writing sounds the way it does, from 'delve' to the five-paragraph template.
Zero-shot detection
Zero-shot detection identifies AI text without training a classifier on examples — using a model's own probability landscape, like DetectGPT's curvature method.
More from Humanize Studio
AI detector guides
Every AI detector flags text differently. These guides explain what each one measures, why it flags AI writing, and how to humanize your text and verify your score before anyone else checks it.
Humanizers by model
ChatGPT doesn't write like Claude, and detectors know it. These guides cover the telltale patterns of each major AI model and how to rewrite them into text that sounds like a person.
Use cases
An email, an essay and a product description fail for different reasons when AI writes them. Pick your situation for specific guidance on making the text genuinely yours.
Blog
How detectors actually work, why they sometimes flag human writing, and what genuinely makes text read like a person wrote it — explained without the hype.
Alternatives
Shopping around? Fair enough. Here's how the popular AI humanizers approach the problem, and where Humanize Studio takes a different path: built-in detection, never-stored text, and an iOS app.
Comparisons
When you're choosing between two specific tools, generic reviews don't help. These are direct head-to-heads: what each one actually does, where each wins, and how to verify either one's verdict yourself.
Try it on your own text
Humanize a passage and check its AI-probability score in one place — 3-day free trial, and your text is never stored on our servers.