Skill.re
AI Terminology Every Learning Pro Should Know
A head of learning is in a tool-selection meeting, and the vendor is talking fast. "Our LLM is fine-tuned on your domain, fully grounded, low hallucination, with an agentic workflow and adaptive paths, and you can tune the temperature." Every L&D leader in the room nods. Not one of them could stop the meeting and ask "wait, which of those words protects us, and which is marketing." Six weeks later the tool ships a module with a fabricated policy threshold, and the post-mortem reveals the team bought a system whose key safety word, "grounded," was never actually verified. The vocabulary was the whole defense, and nobody in the room could wield it. This lesson hands you the words, defined in learning terms, each with a one-sentence reason you care.
Why the Words Are the Job
Vocabulary is not trivia in this field. It is the instrument you use to interrogate a vendor, scope a build, brief a SME, and answer an auditor. Every one of the terms below names a place where AI either becomes safe to use in learning or becomes a liability, and the difference between the two is usually a single word that a vendor says smoothly and an unprepared buyer accepts without testing. When you can hear "fully grounded" and immediately ask "grounded in which source, which version, and show me a claim tracing back to it," you have turned a marketing phrase into an inspectable promise. That is the entire purpose of learning this language.
The terms cluster into three families. There are the machine terms that describe how the AI works (LLM, token, context window, temperature, model drift, fine-tuning). There are the safety terms that describe where it goes right or wrong (grounding and RAG, hallucination, agent). And there are the learning terms where AI meets your discipline (prompt, adaptive learning). Read them as a set, because they interlock: an LLM predicts tokens within a context window at some temperature, which is why it hallucinates, which is why you ground it with RAG, which is why a prompt must be disciplined, which is why an agent that acts on its output needs human oversight. The vocabulary is a chain of cause and effect, not a list.
In learning AI, the word is the control. If you cannot define "grounded," you cannot verify it, and a word you cannot verify is a word a vendor can use to sell you a liability.
The Glossary, Defined in L&D Terms
Here is the working glossary. Each row gives the term, a plain definition in learning language, and a one-sentence "why you care" that ties it to a decision you actually make. Keep this table; it is the reference you will return to every time a vendor or a SME uses a word you need to pin down.
| Term | What it means in plain L&D language | Why you care (the one-sentence stake) |
|---|---|---|
| LLM (large language model) | An AI trained on huge amounts of text to predict and generate language one chunk at a time; the engine behind chatbots and authoring assistants. | It predicts plausible text rather than retrieving verified facts, so its output is always a draft to verify, never a source to trust. |
| Token | The unit of text the model works in: a short chunk, roughly three to four characters, often a word or word fragment. | A policy number or threshold is just tokens to the model, with no protected status, so it can be invented as easily as any other word. |
| Prompt | The instruction and context you give the model to steer its output, including the audience, the source, the objective, and the constraints. | The prompt is your main lever over quality, so a vague prompt produces a vague, ungrounded draft and a disciplined one produces a usable, traceable draft. |
| Context window | The amount of text the model can hold in view at once, the prompt plus the material you give it plus its own answer. | If your policy or SOP is longer than the window, the model cannot see all of it at once, so it may answer from a fragment and miss the clause that matters. |
| Hallucination | Fluent, confident output that is not grounded in any real source: an invented statistic, a fabricated threshold, a citation to a standard that does not exist. | In compliance or safety training a hallucination is a wrong instruction shipped at scale, not a typo, and it arrives in the same confident tone as everything true. |
| Grounding / RAG | Forcing the model to answer from your approved material rather than its training-data memory; RAG means retrieval-augmented generation, retrieving the real passage first, then phrasing an answer constrained to it. | This is the single most important upgrade for regulated content, because it turns an unsourced prediction into an answer an auditor can trace to a document and a version. |
| Agent | An AI set up not just to answer but to take actions in a sequence of steps: searching, calling a tool, updating a record, enrolling a learner, sending a message. | An agent acts, not just drafts, so an unverified agent can propagate a wrong action across systems before a human ever reads the output, which is why agentic workflows need human gates. |
| Adaptive learning | An experience that changes what a specific learner sees or does next based on their performance or profile: skipping, repeating, branching, nudging. | It is a decision about a person, not a paragraph, so its risks are misrouting and unfair outcomes across groups, which you audit in aggregate data, not by proofreading. |
| Model drift | The gradual decline in a model's usefulness as the world, your policies, or the vendor's underlying model change while your setup stays still. | A tool that was accurate at launch can quietly degrade, so "we verified it once" is not a permanent guarantee and AI content needs re-verification over time. |
| Fine-tuning | Further training a base model on a specific body of examples to nudge its style or behavior toward a domain or house voice. | Fine-tuning can shape tone and format, but it does not guarantee factual accuracy, so a fine-tuned model can still hallucinate and still needs grounding and verification. |
| Temperature | The randomness dial controlling how varied the model's output is; lower is safer and more repetitive, higher is more creative and less predictable. | It is why the same prompt can give different answers, so one good run never proves reliability for a regulated procedure, and you test across runs. |
The Distinctions That Trip People Up
A glossary is only useful if you can hold the slippery pairs apart, because vendors and well-meaning colleagues blur exactly these. Three distinctions are worth slowing down on.
Fine-Tuning Is Not Grounding
These two get conflated constantly, and the confusion is expensive. Fine-tuning adjusts how a model writes by training it further on examples; it shapes style, tone, and format. Grounding changes what a model answers from by forcing it to use your approved source at the moment of the question. Why the distinction matters: a vendor who says "we fine-tuned the model on your industry" has improved how the output sounds, not guaranteed that any specific claim traces to your policy. A fine-tuned model can still confidently hallucinate a threshold. Only grounding gives you a source to point to. When a vendor leans on "fine-tuned," your follow-up is always "and is it grounded in our actual source documents." The first word is about voice. The second is about truth.
Hallucination Is Not a Bug to Be Patched
People hear "hallucination" and imagine a defect the vendor will eventually fix in the next release. That mental model is dangerous. Hallucination is a direct consequence of how an LLM works: it predicts plausible text, so where it has no grounded fact, it must still produce plausible-sounding content. You manage hallucination with grounding and verification; you do not wait for it to be patched away. A vendor who promises "no hallucinations" is either grounding every claim and constraining output to the source, in which case ask to see exactly how, or is overselling. The realistic standard is not "a model that never invents" but "a workflow where every load-bearing claim is grounded and verified before it ships."
Adaptive Is a Decision, Not a Feature
"Adaptive" and "personalized" are sold as warm, learner-friendly features, and they describe something colder underneath: an automated decision about what a specific person is allowed to see. Why you care: the moment a system is making decisions about people based on data, you have inherited fairness and privacy obligations, and the failure mode is not a visible error but a quiet pattern of who-was-routed-where across many learners. Hearing "adaptive" should make you ask for the routing logic and the aggregate outcome data, not nod at a feature. The friendliest-sounding word in the glossary hides the job with the most invisible failure mode.
A Worked Example: Reading a Vendor Sentence With the Glossary
Return to the opening sentence and run it through the vocabulary, word by word, the way an AI-literate learning leader would. Watch a smooth pitch turn into a checklist of inspectable promises.
"Our LLM is fine-tuned on your domain." Translation: it is a next-token predictor whose style was nudged toward your industry. That improves how it sounds. It guarantees nothing about whether a given claim is true, because fine-tuning is about voice, not source. Your move: note that this sentence is about tone, and the truth question is still open.
"Fully grounded, low hallucination." Translation: the vendor claims the model answers from your approved material, which would make it traceable. This is the load-bearing claim in the whole pitch. Your move: "Grounded in which sources, which version, and show me a regulated claim tracing back to the exact line." If they can, this is the real safety mechanism. If they cannot, "grounded" was decoration, and "low hallucination" was a hope.
"Agentic workflow." Translation: the system does not just draft, it takes actions, perhaps enrolling learners or updating records. Your move: "Where are the human gates before an action commits," because an agent that acts on an unverified output can propagate an error across systems faster than a human can catch it.
"Adaptive paths." Translation: it makes automated decisions about what each learner sees. Your move: "Show me the routing logic and the aggregate outcomes across groups," because this is a decision about people with an invisible failure mode, not a feature to admire.
"You can tune the temperature." Translation: you can set the randomness dial. Your move: for regulated content, keep it low and still test across multiple runs, because lower temperature improves consistency, not correctness, and one good run never proves reliability.
The pitch did not change. Your relationship to it did. Five smooth phrases became five specific, answerable questions, and the one word that actually protects you, "grounded," got tested instead of trusted. That is the difference the vocabulary buys: not the ability to sound technical, but the ability to convert marketing into verification.
Using the Words in the Three Rooms That Matter
The vocabulary earns its keep in three specific rooms, and it pays to rehearse each. The first is the tool-selection meeting, where the words let you separate what a vendor is promising from what they are merely implying. The second is the SME review, where the words let you explain to a furious subject-matter expert why the model "changed" their procedure, so the conversation becomes a fixable grounding problem rather than a fight about whether AI is trustworthy. The third is the audit, where the words let you answer the only three questions an auditor really asks: where is the source behind this claim, who was the human gate before it shipped, and how do you know it is still accurate today. In every room the question underneath is the same, and the vocabulary is simply what lets you ask it precisely instead of vaguely.
The Words That Are Marketing and the Words That Are Controls
Not every term carries equal weight, and a useful habit is sorting the words into the ones that protect you and the ones that merely impress. "Fine-tuned," "agentic," and "adaptive" describe capabilities; on their own they tell you nothing about whether your content is safe, and a vendor can say all three while shipping a hallucinated threshold. "Grounded," "verified," and "human gate" describe controls; they are the words that, if real and demonstrated, actually keep a wrong claim off a learner's screen. When you hear a capability word, your job is to ask which control word backs it. "Agentic" is impressive until you ask "where is the human gate." "Adaptive" is impressive until you ask "where is the fairness audit." "Fine-tuned" is impressive until you ask "and is it grounded." The capability words sell the tool. The control words defend the workforce, and you are the person responsible for the second list.
This is also why the glossary is worth teaching to a whole team rather than hoarding as personal expertise. When everyone who touches an AI tool can hear "grounded" and reflexively ask "in which source, which version, show me a claim," the organization stops being able to accidentally buy or ship an unverified system. The vocabulary becomes a shared immune response. A single literate buyer can be talked past on a bad day; a whole function fluent in the difference between voice and truth, between a capability and a control, is much harder to sell a liability to.
How the Terms Interlock Into One Rule
Step back and the glossary collapses into a single chain you can recite. An LLM predicts tokens within a context window at some temperature, which is why it hallucinates, which is why you ground it with RAG and write a disciplined prompt, which is why a fine-tuned model still needs verification, which is why an agent that acts on its output needs a human gate, which is why adaptive routing needs a fairness audit, which is why model drift means you re-verify over time. Every term is a link, and the chain ends where the whole program ends: a human verifies the load-bearing claims and owns the decision, because the vocabulary explains exactly why no single term in the chain can be trusted on its own.
That is why this glossary is not a quiz to memorize and forget. It is the operating language of an AI-aware learning professional, the set of words that lets you stand in the tool-selection meeting, the SME review, and the audit and ask the one question that matters in each: where is the source, where is the human gate, and how do you know it is still true. Learn the words and you can ask those questions. Skip them and you are nodding along while a vendor decides what your workforce learns.
You do not need to build the model. You need to be the person in the room who can hear "grounded" and ask "show me the source," because that single question is the line between a tool you control and a liability you bought.
Key Takeaways
- Vocabulary is the instrument of the job: each term names a place where AI becomes safe or becomes a liability, and the difference is often one word a vendor says smoothly and a buyer accepts without testing.
- An LLM predicts tokens within a context window at some temperature, so its output is always a draft to verify, and a number is just tokens with no protected status.
- Grounding and RAG are the single most important upgrade for regulated content, because they turn an unsourced prediction into an answer traceable to a document and a version.
- Hallucination is a consequence of how LLMs work, not a bug to be patched, so you manage it with grounding and verification rather than waiting for a fix.
- Fine-tuning shapes voice, not truth: a fine-tuned model can still hallucinate, so "fine-tuned on your domain" is never a substitute for "grounded in your source."
- An agent takes actions, not just drafts, so agentic workflows need human gates before an action commits, or an error propagates across systems before anyone reads it.
- Adaptive learning is an automated decision about a person, so it carries fairness and privacy obligations and an invisible failure mode you audit in aggregate data, not by proofreading.
- The terms interlock into one rule: because of how each works, a human must verify the load-bearing claims and own the decision, and the vocabulary is what lets you ask where the source, the gate, and the proof are.