Academic integrity policy
A school's rules on what counts as your own work, including whether and how AI tools may be used. Policies differ between institutions and even between courses, so the policy for your class is the one that applies.
Glossary
Plain-English definitions of the terms behind AI detectors and AI humanizers, from perplexity and burstiness to false positives and watermarking. Each one links to a deeper guide or a free tool.
A school's rules on what counts as your own work, including whether and how AI tools may be used. Policies differ between institutions and even between courses, so the policy for your class is the one that applies.
A tool that estimates whether text was written by an AI model. Detectors such as GPTZero, Turnitin and ZeroGPT score statistical patterns like word predictability and sentence variation. They give a probability, not proof of who wrote the text.
A tool that rewrites AI-generated text so it reads like natural human writing. Unlike a paraphraser, a humanizer changes the sentence rhythm, structure and word choice that AI detectors measure, while keeping the original meaning.
A detector feature that looks for text that was generated by AI and then reworded by a paraphrasing tool. Turnitin reports this kind of text separately in its AI writing report.
The number an AI detector reports for a piece of text, usually the share of the text it believes is AI-generated or the probability that it is. Scores from different detectors aren't comparable, and none of them prove authorship.
Patterns that make text read as machine-written: stock phrases like 'plays a crucial role', words like 'delve' and 'tapestry', formulaic transitions, rule-of-three lists and heavy em dash use. Readers notice them, and detectors pick up the predictability behind them.
How much sentence length and structure vary through a piece of writing. Human writing tends to mix long and short sentences, while AI output is often even. GPTZero popularized it as an AI signal, and modern detectors still respond to it.
A machine learning model trained to sort inputs into categories. Most AI detectors are classifiers trained on examples of human and AI writing, which learn to label new text as one or the other.
The long dash (—) used to set off a phrase. Human writers use it too, but heavy em dash use has become a well-known sign of unedited ChatGPT output. Removing em dashes alone doesn't change an AI detector score.
When an AI detector labels AI-generated text as human-written. Edited, mixed or humanized text is the most common cause.
When an AI detector labels human-written text as AI-generated. Formal or templated writing, short texts and writing by non-native English speakers are flagged more often, which is why detector scores shouldn't be treated as proof.
Why detectors have false positivesWhy is my writing flagged as AI?
AI systems that create new content such as text, images or code from a prompt. ChatGPT, Claude, Gemini, Copilot and Grok are generative AI tools for text.
Short for generative pre-trained transformer, the family of large language models from OpenAI that powers ChatGPT. People often use 'GPT' loosely to mean any AI text generator.
When an AI model states something false or invents details, such as fake citations, with confidence. Humanizing doesn't fix hallucinations, so always check facts, figures and sources in AI-assisted work.
To rewrite AI-generated text so it sounds like a person wrote it: varied sentence rhythm, natural word choice and fewer stock phrases, with the meaning unchanged. You can humanize by hand or with an AI humanizer.
An AI model trained on large amounts of text to predict the next word, or token, in a sequence. LLMs power chatbots like ChatGPT, Claude and Gemini, and their tendency to choose likely words is what detectors look for.
The Humanizer rewrites the perplexity and burstiness patterns detectors measure, while keeping your meaning. 350 free words every month.
A tool, such as QuillBot, that rewords text by swapping words and rearranging phrases. Paraphrasing usually keeps the underlying sentence structure, so paraphrased AI text can still be flagged by detectors.
A measure of how predictable a text is to a language model. Low perplexity means each word is the one a model would expect, which is typical of AI output. It was an early core signal for detectors, and today's trained models still pick up the predictability it measures.
The instruction or question you give an AI model. Better prompts can make AI drafts sound more natural, but prompting alone rarely removes the patterns detectors measure.
A number estimating how easy text is to read. The Flesch reading ease score runs from 0 to 100, where higher is easier, and grade-level formulas estimate the years of schooling needed. Readable text is not automatically human-sounding.
How much the number of words per sentence changes from one sentence to the next. It's the easiest part of burstiness to see and fix: a run of similar-length sentences is a common AI tell.
Turnitin's plagiarism measure: the share of a submission that matches existing sources. It's separate from the AI writing score, so original AI-generated text can have a low similarity score and a high AI score.
The Humanizer's setting for long texts. Free and Basic plans humanize up to 400 words per run; with Supercharged mode, Pro and Ultimate users can humanize longer texts in one pass with a consistent voice.
A setting that controls how random an AI model's word choices are. Low temperature produces safer, more predictable text, which detectors find easier to flag. Higher temperature adds variety but can reduce accuracy.
The unit of text a language model reads and writes, often a word or part of a word. Models generate text one token at a time by picking from likely options, which is the source of the predictability detectors measure.
AI-assisted text that AI detectors classify as human-written. No tool can promise this for every detector and every text, which is why humanized work should still be checked before it matters.
The record of edits a document keeps over time, as in Google Docs or Microsoft Word. It's one of the strongest ways to show you wrote something yourself if your work is ever flagged as AI-generated.
Embedding a hidden statistical signal in AI-generated text, usually by nudging word choices, so the text can later be identified as AI-written. It's an active research area, and a watermark only helps if the checker has access to the provider's method.
Keep learning
The signals behind GPTZero, Turnitin and ZeroGPT, in plain English.
Read guideThe two numbers detectors lean on, with simple examples.
Read guideCommon reasons human writing gets flagged, and what to do.
Read guidePaste text from ChatGPT, Claude or Gemini and get natural writing designed to pass GPTZero, Turnitin and other AI detectors.