Plain definitions for the vocabulary of personal AI writing: writing profiles, stylometry, register, voice fidelity and the rest. One sentence each, with sources.
An AI writing profile is a structured, reusable description of how a specific person actually writes, measured from samples of their own text and supplied to a language model so its output carries that person's patterns instead of the model's defaults.
Stylometry is the statistical measurement of writing style — sentence length, function-word frequency, punctuation habits, hedging and vocabulary distribution — used to characterise or identify an author from text alone.
A Voice DNA runtime block is a compact, machine-readable rendering of a writing profile, formatted to be loaded directly into a language model's system prompt so an agent can adopt a person's voice without a human pasting instructions.
Prompted AI writing tells a model how you want to sound; trained AI writing supplies evidence of how you already sound, so the difference is between a description you author and a measurement of text you wrote.
Custom instructions are a free-text field inside one assistant where you state your preferences; a style profile is a measured artifact about your writing that can be pasted into that field, into a different assistant, or into an agent's system prompt.
A tone label is an adjective describing how writing should feel, such as professional yet friendly; a voice is the measurable set of habits — sentence length, qualifier placement, transition choice, punctuation — that make one person's writing recognisably theirs.
Writing sample size is the amount of an author's own text a system needs before per-author style measurements stabilise, and the binding constraint is usually variety of contexts rather than raw word count.
AI-flattened English is the narrow band of register that language models converge on by default — mid-formal, evenly hedged, structurally symmetrical — which reads as competent and carries no information about who wrote it.
Keigo is the Japanese honorific system that encodes the relationship between speaker, listener and referent, and AI assistants handle it unreliably because the correct level is determined by social context the model is not given rather than by the sentence being translated.
Function word analysis measures how often an author uses grammatical words such as the, and, of and but, and it is the backbone of stylometric authorship attribution because those choices are made unconsciously and stay stable across topics.
Register is the systematic variation in a writer's language according to situation — audience, medium, purpose and formality — and it is why one person's Slack messages and board memos are measurably different texts by the same author.
Perplexity measures how surprising a text is to a language model and burstiness measures how much that surprise varies from sentence to sentence, and AI-detection tools use both because model output tends to be both unsurprising and uniformly so.
Voice fidelity is how closely generated text reproduces a specific author's measured patterns, evaluated by comparing the output's stylometric features against that author's corpus rather than by asking a reader whether it sounds right.