The AI detection glossary

What the words in a detection report actually mean, defined so each one stands on its own. Every entry carries a worked example with real text, because a definition you cannot see applied is a definition you cannot use. 50 terms so far.

These pages exist because the vocabulary is doing real damage. A student told their essay scored 87% is rarely told what was measured, and the number sounds like a probability of guilt when it is nothing of the kind. Most of the terms below describe how predictable a piece of writing is — which is a property of the prose, not a fact about its author.

We are adding terms as each one gets a real example rather than publishing a list of stubs. If a term you need is missing, tell us and it goes to the front of the queue.

A

Academic integrity — The standard of honest scholarly practice an institution sets and enforces.
Academic misconduct — The umbrella term for breaches of an institution's academic integrity rules.
AI detector — A category name for tools that estimate whether text was machine-generated.
AI disclosure statement — A written declaration of how AI tools were used in producing a piece of work.
AI slop — Low-effort generated content published in volume with no one accountable for it.
AI writing score — The share of a document a detector flagged as likely machine-written.
Appeal — A formal challenge to a decision, heard on stated grounds rather than reargued.

B

Burstiness — How much sentence length and complexity vary across a passage.

C

C2PA — An open standard attaching signed provenance data to digital content.
Confidence interval — The range a measured rate could plausibly be, given how little was tested.
Contract cheating — Paying a third party to produce work you submit as your own.

D

Detection threshold — The score above which a tool reports a document as AI-written.
Document-level detection — Reporting one score for a whole document rather than for its parts.
Due process — The procedural protections you are owed before a decision is made against you.

F

False negative — AI-generated text that a detector fails to flag.
False positive — Text a detector flags as AI-generated when a person actually wrote it.
False positive rate — The proportion of human-written documents a detector wrongly flags as AI.
Fine-tuning — Further training of an existing model on a narrower dataset to change its behaviour.
Flesch Reading Ease — A 0–100 readability score where higher means easier.
Flesch–Kincaid Grade Level — A readability score expressed as a US school year.

G

Generative AI — Systems that produce new content rather than classify existing content.
GPT — Generative pre-trained transformer — the architecture behind most text models.

H

Hedging — Language that softens a claim to match the strength of the evidence behind it.
Honour code — An institution-wide commitment to academic honesty, often signed and student-enforced.
Humanized text detection — Detection aimed specifically at text that has been through a rewriting tool.

I

Invisible character — Unicode characters that occupy no visible space but are present in the text.

L

Large language model — A model trained to predict the next piece of text, used to generate prose.
Lexical diversity — How varied a text's vocabulary is, and why the usual measure of it cannot be compared.

N

N-gram — A run of n consecutive words, used to compare texts for overlap.
Nominalization — A verb turned into a noun, usually needing an empty verb to prop it up.

O

Originality report — The document a text-matching service produces about a submission.

P

Passive voice — A construction where the thing acted on becomes the subject of the sentence.
Perplexity — A measure of how surprised a language model is by a piece of text.
Plagiarism — Using someone else's work or ideas without crediting them.
Precision and recall — Two separate measures that a single accuracy figure averages away.
Proctoring — Supervision of an assessment to confirm the candidate is working unaided.
Prompt — The instruction and context given to a model to produce a response.

R

Readability — How much effort a text takes to read, estimated from sentence and word length.
Register — The level of formality a text adopts for its audience and situation.
Retrieval-augmented generation — Fetching real source material and giving it to a model before it answers.

S

Sentence-level detection — Reporting a score per sentence rather than one figure for the document.
Similarity score — The share of a document that matches text in a database of sources.
Stylometry — Identifying an author from measurable habits in how they write.
Syllabus AI policy — The course-level statement setting what AI use is permitted in that module.
SynthID — Google DeepMind's watermarking system for AI-generated content.

T

Temperature — A setting controlling how much randomness a model uses when choosing each word.
Token — The unit of text a language model actually reads — usually a word fragment.
Transition word — A word or phrase signalling how one sentence relates to the one before it.

V

Viva voce — An oral examination used to establish whether you can account for your own work.

W

Watermarking — Hiding a detectable signal inside AI output at the moment it is generated.

Where to go from here

If you are here because something you wrote was flagged, the glossary is the wrong place to start — read what to do when you are falsely accused first, then come back for the vocabulary you need to argue with. If you are here to understand the tools, start with how detectors work and how accurate they are.