An AI glossary is a controlled vocabulary: a curated set of terms with definitions, disambiguation notes, provenance, and links into deeper guides. It is a navigation and hygiene layer—not an encyclopedia that replaces the Knowledge library. This page owns how to use and maintain an AI glossary well: glossary versus encyclopedia, homonyms and overloaded terms, definition provenance, linking terms to Knowledge guides, versioning, anti-patterns such as buzzword inflation, and clear boundaries to “what is AI” and domain guides. It does not publish a 500-term dump, and it does not invent a fake definitions encyclopedia of everything.
Readers who need conceptual on-ramps should use dedicated guides such as artificial intelligence, machine learning, generative AI, and large language models. Glossary entries should point into those treatments rather than competing with them.
Glossary versus encyclopedia
A glossary answers “what does this term mean here, and where do I go next?” An encyclopedia answers “teach me the whole subject.” Confusing the two produces either thin stubs that pretend to be complete or sprawling articles stuffed into a dictionary UI.
Glossary scope should be bounded: terms that appear repeatedly across Brel Knowledge, product UI, and buyer conversations; terms that are routinely confused; and terms that need local house meaning (for example, how Brel uses “model” versus “product”). It should not attempt to define every paper neologism or every vendor trademark.
Length discipline: a glossary definition is typically a short paragraph plus disambiguation and links—not a multi-thousand-word essay. When a term needs systems treatment, the glossary owns the pointer; the guide owns the depth.
Audience discipline: glossary readers may be newcomers or practitioners skimming under time pressure. Write for precision without requiring prior coursework, but do not flatten contested terms into false simplicity. Flag contestation instead of erasing it.
Maintenance discipline differs too. Encyclopedic guides update when practice changes materially. Glossaries update when language drifts, when homonyms collide, or when house usage changes. Different cadences; different owners.
Success metric for a glossary is reduced confusion and faster routing—not page count. A smaller, consistent glossary beats a large, contradictory one.
Think of the glossary as an index with opinions about language, not as a second Knowledge hub. Hub pages list guides; glossary pages stabilize words those guides share. If the glossary grows navigation features that compete with the hub, you have an IA problem, not a writing shortage.
Inclusion votes should be boring. A term earns a slot when it appears in at least a few guides, causes repeated reader confusion, or is required for consistent metadata. Novelty alone is not enough. Yesterday’s launch buzzword is tomorrow’s dead entry.
Write for skimming. Lead with the sense most used on Brel, then alternate senses. Readers under deadline should not need to parse an essay to avoid a homonym trap.
Homonyms and overloaded terms
AI English is dense with homonyms. “Model” can mean a trained artifact, a product SKU, a statistical model, a business model, or a fashion reference in unrelated content. “Agent” can mean a reinforcement-learning policy, a tool-using LLM loop, a human customer-support agent, or a legal agent. “Platform” can mean almost anything sold with a dashboard.
Overloaded terms need structured disambiguation: numbered senses, short glosses, and links to the guide that owns each sense. Example pattern: Agent (1) tool-using system—see AI agents; Agent (2) human role in a workflow—out of AI glossary scope or cross-linked to operations language.
Context labels help: “in ML training,” “in product packaging,” “in governance policy.” Without labels, readers import the wrong sense into contracts and PRDs.
Prefer stable sense IDs over rewriting history. If sense (1) was “classifier,” do not silently redefine sense (1) as “foundation model.” Add a new sense or version the entry.
Also watch near-synonyms used as status symbols: “AI,” “ML,” “deep learning,” and “generative AI” are related but not identical. Glossary hygiene points to machine learning, deep learning, and generative guides rather than treating them as interchangeable badges.
When vendors redefine common words (“memory,” “reasoning,” “knowledge”), record the vendor sense as a labeled dialect—not as the universal meaning. House glossary should protect shared language from marketing capture.
“Inference,” “training,” “fine-tuning,” and “alignment” each carry research and product dialects. Glossary senses should separate “model training” from “staff training,” and “alignment” as safety research goal from “alignment” as sales jargon for “matches our roadmap.” Label dialects explicitly.
“RAG,” “agent,” “embedding,” and “vector” are especially collision-prone for newcomers. Point each primary sense to its owning guide—embeddings, RAG, agents—rather than restating pipeline tutorials inside the dictionary.
Legal and policy language collides with ML language: “decision,” “profiling,” “automated processing.” When glossary terms sit near regulation discussions, add a caution that legal senses may differ and point readers to regulations literacy rather than freelancing legal definitions.
Definition provenance
Every glossary definition should have provenance: who wrote it, what sources informed it, whether it is a Brel house definition, and when it was last reviewed. Provenance is how you prevent silent drift and ghost edits.
Source classes: (a) house operational definition tied to how Brel pages use the term; (b) paraphrase of a widely taught technical meaning with pointer to a Knowledge guide; (c) contested term with multiple schools noted; (d) vendor dialect clearly marked. Do not invent academic citations. Do not fabricate standards clause numbers.
When a definition depends on a deeper guide, the guide is the canonical expansion. The glossary should not fork a second incompatible deep definition. If practice changes in RAG or embeddings, update the guide first, then align the glossary gloss.
Record known disagreements. “Alignment” and “hallucination” are used inconsistently across research and product teams. A glossary that pretends consensus where none exists mis-educates.
Reject generative autodrafts as sole provenance. Models can propose candidate glosses; humans must accept, cite class, and own errors. Unreviewed machine definitions are how contradictions multiply.
Provenance also covers deprecation: when a term is discouraged in house style (“AI-powered” as empty adjective), say so and point to preferred language.
| Provenance class | What it means | Reader expectation | Update trigger |
|---|---|---|---|
| House definition | Brel usage for IA consistency | Normative for our pages | IA or style change |
| Technical gloss | Points to Knowledge guide | Short; depth elsewhere | Guide revision |
| Contested | Multiple senses noted | No false consensus | New dominant usage |
| Vendor dialect | Labeled product language | Not universal | Product rename |
Store provenance internally even if the public page shows only a short gloss and date. Editors need to know whether a sentence came from a deliberate house style decision or from a temporary paraphrase of a guide intro. Otherwise drive-by edits recreate conflicts.
When borrowing phrasing from a Knowledge guide, prefer paraphrase plus link over verbatim duplication. Duplication guarantees drift. The guide remains free to evolve; the gloss remains a stable pointer with a reviewed summary.
Conflict protocol: if two guides disagree on a term, do not let the glossary silently pick a winner. Open an ownership discussion, update the losing guide or split senses, then publish the glossary change with a note.
Linking glossary to Knowledge guides
Glossary entries should be doorways. A strong entry names the term, gives a precise short definition, lists senses, and links to one primary guide plus optional secondary guides. It should not try to replace large language models or generative AI with a mini-article.
Link economy: one primary “owning” guide per sense when possible. Too many equal links create decision fatigue. Secondary links are for adjacent mechanisms—for example, an embeddings gloss primary-links embeddings and may secondary-link RAG.
Bidirectional discipline: guides may define terms inline, but the glossary remains the index. When a guide introduces a specialized meaning, add or update the glossary sense rather than leaving a trap for searchers who only hit the dictionary.
Avoid circular stubs: glossary → thin page → glossary. If no guide yet owns the topic, either keep the gloss minimal with “guide pending” honesty or do not include the term until ownership exists. Do not invent placeholder encyclopedias.
For agentic and retrieval topics, keep boundaries clean: AI agents owns control-loop literacy; glossary senses should not smuggle in a second agents tutorial.
Search and UX: glossary pages should make sense isolation visible so a reader landing from web search sees that multiple meanings exist before they paste the wrong one into a contract.
Anchor text in glossary links should use the guide’s public title language, not clever variants, so readers recognize they are entering a deep article. After the jump, the guide’s first paragraphs should not assume glossary context; each URL must stand alone.
For umbrella terms like “artificial intelligence” and “machine learning,” the glossary should state the relationship in one or two sentences and defer taxonomy fights to the guides that own them. The glossary’s job is orientation, not settling every boundary dispute in fifty words.
Measure link health periodically: broken anchors, guides that renamed H2s, and orphan senses with no primary link. Link rot turns a glossary into a maze.
Versioning terms
Terms version because technology and fashion move. Versioning practices:
Entry version. Increment when the definition’s meaning changes, not when you fix typos. Show last-reviewed dates on public glossary UIs when feasible.
Sense stability. Prefer adding sense (3) over silently rewriting sense (1). Downstream docs that cited sense (1) should not break invisibly.
Alias and redirect. Maintain aliases (“LLM,” “large language model”) without duplicating conflicting definitions. One canonical entry; aliases point in.
Deprecation path. When a term becomes misleading, mark deprecated, explain preferred term, and keep the old entry findable for a while so historical docs remain interpretable.
Change log. Internal notes should record why a definition changed—new house IA, community shift, or correction of an error. Without a log, debates restart every quarter.
Versioning also applies to examples. Examples age faster than definitions. Prefer durable structural examples over vendor-version-specific screenshots in glossary space.
Coordinate with guide versioning. If artificial intelligence revises boundary language, glossary entries that summarized those boundaries need a pass in the same editorial batch.
Communicate material glossary changes to authors who rely on the terms in flight. A Slack or editorial note beats silent breakage. Authors should know when “platform” stopped meaning “any dashboard” in house style.
For public consumers, consider showing “updated on” at entry level. Transparency builds trust that language is curated, not frozen in 2017 blog prose.
Snapshot exports can help large editorial programs: a dated glossary dump for writers. Snapshots are references, not excuses to stop linking live entries.
Anti-patterns (buzzword inflation)
Buzzword inflation is the glossary’s natural enemy. Anti-patterns include:
Synonym stuffing. Adding every marketing variant as if it were a distinct technical term (“AI solution,” “intelligent automation,” “cognitive platform”) without distinguishing senses.
Definition theater. Long poetic definitions that never constrain usage. If a definition cannot fail, it is not a definition.
Scope creep into encyclopedia. Turning each entry into a full guide duplicate, guaranteeing drift between copies.
False precision. Inventing neat distinctions that practitioners do not use, then enforcing them pedantically.
Vendor capture. Letting a single vendor’s coinage overwrite shared meaning without a dialect label.
Autogenerated sprawl. Bulk-creating hundreds of entries from model output without owners, provenance, or links—producing a junkyard that looks complete.
Metric cosplay. Defining “accuracy,” “hallucination rate,” or “alignment score” as if universal formulas existed, without pointing to evaluation context.
Countermeasures: inclusion criteria, owner assignment, maximum entry length, mandatory primary link for technical terms, and periodic deletion or merge passes. Shrinking a glossary can increase quality.
Inflation also happens through prestige stacking—defining every term as “advanced,” “next-generation,” or “enterprise.” Strip intensifiers. If the noun needs an adjective to sound real, the noun is doing no work.
Another anti-pattern: comic precision—“exactly 17 types of AI”—that invents taxonomy theater. If a count is not load-bearing for navigation, omit it. Counts become fake citations.
Resist translating every vendor SKU into a glossary entry. Product names belong on product or company entities; glossary owns shared language. SKU spill turns the dictionary into an ad index.
Boundary to what-is-AI and domain guides
This page owns glossary practice—how to maintain term hygiene across the library. It is not the beginner on-ramp and not the systems map.
For accessible conceptual orientation, use the what-is-AI on-ramp when published in the library path, and use artificial intelligence for deeper systems treatment. For learning methods, use machine learning. For generative systems, use generative AI and large language models. For representation and retrieval building blocks, use embeddings and RAG. For tool-using systems, use AI agents.
Do not replace those articles with glossary text. Do not invent a mega-glossary that pretends to finish AI education. Do not mint fake citations to decorate definitions.
A healthy glossary makes shared language safer: fewer homonym accidents, clearer handoffs into Knowledge, and less buzzword fog in requirements. Maintain it like critical infrastructure—small, versioned, evidenced, and linked—not like a novelty word zoo. When in doubt, define less, disambiguate more, and send readers to the guide that owns the system.
Editorial onboarding should include a thirty-minute glossary clinic: how to add a sense, how to request a deprecation, how to avoid encyclopedia creep, and how to file a conflict when two guides disagree. Without onboarding, each new author reinvents language privately.
QA checks before publishing a guide: scan for undefined jargon, confirm glossary senses exist for recurring terms, and ensure the guide does not invent a private meaning for a shared word. If a private meaning is necessary, add a labeled sense or choose a different word.
Reader-facing help text can say: “Short definitions live in the glossary; deep explanations live in Knowledge guides.” That one sentence prevents disappointed expectations when a glossary entry refuses to be a textbook.
Internationalization caution: translating glossary entries is not a word-swap. Homonyms differ by language, and vendor dialects may not translate cleanly. If you localize, re-run disambiguation workshops with native-speaking practitioners rather than machine-translating contested terms blindly.
Metrics for glossary health: number of conflicting senses resolved per quarter; percent of technical entries with primary links; age of oldest unreviewed entry; and count of entries deleted or merged. Growth alone is a vanity metric; resolution and link health are quality metrics.
The glossary succeeds when a newcomer can survive first contact with AI jargon without false certainty, and when experts can cite a house meaning without starting a flame war. Maintain the instrument; keep it small enough to trust; and let domain guides carry the weight of teaching systems, not the dictionary.
When stakeholders demand “just add everything,” answer with the inclusion criteria and a merge plan for near-duplicates. Expanding without merges is how glossaries become landfills. A quarterly delete-and-merge ritual is a feature of serious language hygiene, not an admission of failure.
Finally, remember that glossary work is cultural. Teams that laugh off definitions invent private dialects in tickets and contracts. Teams that over-police language create pedantry theater. Aim for shared enough meaning to coordinate, with explicit senses where the stakes require precision—and always a clear door into the Knowledge guide that owns the system behind the word. That door into depth is the entire point of the glossary as a navigation layer.