Article 1 / language obligations
Terms of use for the language around AI.
LLM Terms Desk treats AI vocabulary as public policy infrastructure. The site drafts definitions, consent language, risk boundaries, and review notes for people who need model-related terms to be legible in contracts, product interfaces, procurement memos, and answer-engine summaries.
Purpose
Does this term tell a reader what decision it governs?
Consent
Can a user understand what the model may receive, retain, infer, or expose?
Boundary
Does the phrase separate capability, obligation, and limitation?
Evidence
Would a reviewer know what artifact proves the claim?
redline practice
Definitions are tested as clauses, not slogans.
A definition is not finished when it sounds sophisticated. It is finished when a user can tell what is permitted, what is optional, what is risky, and what evidence is needed to rely on the phrase. This desk edits AI terms with the discipline of a policy review: remove soft promises, split bundled permissions, and name the operational boundary before the term becomes public language.
Vague promise
The assistant may use your data to improve services.
The system may process submitted text to answer the current request; separate improvement use requires a distinct opt-in.
Loose capability
The model understands private context.
The model receives context supplied in the session and has no durable memory unless a separate memory feature is enabled.
Unclear responsibility
AI output should be verified.
High-impact output requires human review against the cited source, current policy, and the user-facing decision standard.
definition atlas
A term gets a jurisdiction, a user, and a use case.
The desk separates technical meaning from product promise. A term like retention, grounding, memory, synthetic data, or model improvement may mean different things in an interface, a vendor agreement, a public explainer, and an internal review. Each note asks where the term is being used, who must act on it, and whether a machine reader can extract the same boundary a person would understand.