The vocabulary of AI detection and humanization, defined in plain English. Use it to read detector results, evaluate humanizer claims, and follow the conversation about AI writing.
Core detection terms
AI detector. A tool that scores text on the likelihood it was generated by a large language model. Modern detectors score statistical features like perplexity and burstiness rather than matching against a corpus. Leap's detector runs free in your browser and highlights the sentences that weigh most on the score.Perplexity. How predictable the next word is given the previous ones. Low perplexity means the text reads as what a language model would have written next, a strong AI signal. High perplexity means unexpected word choices, a human signal.Burstiness. The variance of sentence length and structure in a passage. Human writing swings between short and long sentences; AI writing tends to stay in a middle band. Low burstiness is a strong AI signal.Token probability distribution. The shape of word-choice probabilities across a passage. AI reaches for a narrow band of mid-formal vocabulary, and the distribution of choices tells detectors whether that band is AI-shaped.False positive. Human writing flagged as AI. The central failure mode of detection. Non-native English, technical prose, and formal register are particularly prone to false positives, which is why a detector score should never be the only basis for an accusation.False negative. AI text missed by a detector. Edited AI text, short passages, and outputs from newer models often produce false negatives.Watermarking. Embedding invisible statistical patterns in AI output so detectors can reliably identify it later. Research labs have described watermarking work; adoption is voluntary, and watermarks can typically be stripped by light editing.
Humanization terms
AI humanizer. A tool that rewrites AI-generated text to reduce common detection signals. Leap's humanizer runs a cleanup pass in your browser: it replaces stock AI phrases, removes em dashes and invisible characters, adds contractions, and marks sentences of uniform length so you can vary them. It does not guarantee passing any detector.Paraphrasing. Swapping synonyms and reordering clauses without changing statistical patterns. An early technique that mostly does not shift modern detection scores, because the underlying rhythm and vocabulary distribution stay the same.Rewriter or rephraser. Generic terms for tools that reword text. Not necessarily detection-aware, so a rewriter may or may not affect detection scores.Detect-then-revise loop. The common workflow: score text with a detector, rewrite the flagged sentences, then score again. Leap's detector highlights the heaviest sentences so you know where to revise.
Academic and editorial terms
E-E-A-T. Experience, expertise, authoritativeness, and trustworthiness. Google's framework for evaluating content quality. Generic AI content tends to be weak on experience and expertise.COPE. The Committee on Publication Ethics. Sets the international standard for academic publishing ethics, including guidance on AI use and disclosure in scholarly work.Plagiarism detector. A tool that matches text against a corpus of known sources. Different from an AI detector, which scores writing patterns rather than matching sources.Turnitin. An academic integrity product that combines plagiarism checking with an AI writing indicator. Widely embedded in learning management systems at universities.
AI model terms
LLM (large language model). The class of model that produces text by predicting the next token. ChatGPT, Claude, Gemini, and Llama are all LLMs.RLHF (reinforcement learning from human feedback). The training technique that tunes LLMs to produce responses human raters prefer. It contributes to the hedge-heavy, mid-formal register of typical AI output.System prompt. The instruction given to an LLM that sets its behavior. Different system prompts produce different output patterns, which affects what detectors see.Temperature. A setting that controls how predictable or creative an LLM's output is. Lower temperature produces more predictable text; higher temperature produces more varied output.
Industry terms
Stock AI vocabulary. Words LLMs reach for at rates much higher than human writers do, such as "leverage," "multifaceted," "delve," and "ensure." A heavy frequency of these words is one of the writing signals Leap's detector looks at.The em-dash problem. AI models, especially chat assistants, tend to over-use em dashes, which has become one of the clearest visual tells of AI-drafted text. Leap's humanizer removes em dashes as part of its cleanup pass.Detector bias. Research, including a widely cited 2023 Stanford study, has shown that AI detectors misclassify non-native English writing at high rates. A reminder that detector scores are signals, not proof.The AI detection arms race. The ongoing dynamic between detection tools and rewriting tools. Each new detector release prompts rewriting changes, and each rewriting improvement prompts detector retraining.
Where to go next
If you are new to the space, start with how AI detectors work and how to humanize AI text. For background, read the history of AI writing detection. If you want to try the tools, the detector and the humanizer are both free and run in your browser with no signup.