AI Writing Glossary

AI detector: A tool that estimates whether a piece of text was written by AI, usually as a percentage or a label. Detectors use statistics, trained classifiers or both, and their results are estimates, not proof. See How AI detectors work, or try our free AI detector.
Burstiness: How much a text varies, especially in sentence length and structure. Human writing tends to mix short and long sentences. AI writing tends to keep an even pace. See Perplexity and burstiness.
Chatbot leftovers: Lines an AI tool writes to the person using it, not to the final reader, such as Certainly! Here's your essay or I hope this helps. When they're pasted into finished work, they're the clearest sign of AI text.
Classifier: A model trained on labeled examples to sort new text into categories, such as "human" or "AI". Most commercial detectors include one.
Context window: The amount of text a model can take into account at once, including your prompt, any pasted material and its own reply. Long style sheets and samples use up part of it.
Custom instructions: Standing instructions saved in an AI tool that apply to every chat. A good place to keep your personal style sheet.
DetectGPT: A detection method published by Stanford researchers in 2023. It makes small rewrites of a text and checks whether the original is unusually likely according to a model, which is typical of AI output.
E-E-A-T: Experience, expertise, authoritativeness and trustworthiness. The qualities Google's quality guidelines ask raters to look for in content. See AI content and SEO.
False positive: When a detector labels human writing as AI-generated. Short texts, formal writing and writing by non-native English speakers are flagged this way more often. See Falsely flagged as AI?
Hallucination: When an AI model states something false with confidence, such as an invented statistic, quote or source. The reason every fact in an AI draft needs checking.
Humanizer: A tool that rewrites AI text, usually by swapping words and shuffling sentences, to lower detector scores. The results often read strangely or change the meaning. Editing with your own details works better. See The 10-minute de-robot edit.
Language model (LLM): The kind of AI behind writing tools. It's trained on large amounts of text to predict which word, or token, should come next. "LLM" stands for large language model. See How an AI model picks the next word.
Negative parallelism: The it's not just X, it's Y sentence pattern. Common in AI writing, especially social posts. See 12 tells that give away AI writing.
Perplexity: A measure of how surprising a text is to a language model. Low perplexity means the words were easy to predict, which is typical of AI-generated text. See Perplexity and burstiness.
Prompt: The instructions and material you give an AI tool. Prompts built on your own notes produce far more specific drafts. See Brain-dump before you prompt.
Rule of three: Grouping things in threes (fast, simple and reliable). Natural in moderation. When nearly every list has exactly three items, it reads as AI-written.
Scaled content abuse: Google's name, since March 2024, for mass-producing pages mainly to rank in search, whether by AI, people or both. It's against Google's spam policies.
Stock vocabulary: Words and phrases that appear far more often in AI output than in human writing, such as delve, tapestry and plays a crucial role. See Banned words and style rules.
Style sheet: A one-page description of how you write that you give to an AI tool. See Build a personal style sheet.
Temperature: A setting that controls how adventurous a model is when choosing words. Low temperature picks the most likely word almost every time. High temperature adds randomness, not meaning.
Token: The unit a language model reads and writes. Usually a word or part of a word. As a rough guide, 100 tokens is about 75 English words.
Training data: The text a model learned from. A model's habits, including the words it overuses, come from patterns in its training data and from later tuning.
Version history: A record of every saved change to a document, kept by tools like Google Docs and Word on OneDrive. Strong evidence of your writing process. See Use version history as proof.
Voice brief: A short list of concrete rules describing how you write, built from samples of your own work. See Write a voice brief.
Watermark (AI text): A hidden statistical pattern some AI companies can build into the words their model chooses, which their own tools can later detect. Google's SynthID is one example. Heavy editing weakens it.