Evaluating AI adoption or listening to a vendor pitch brings a stream of unfamiliar words: LLM, prompt, tokens. When the meaning is unclear, so is what you are supposed to judge.
This article covers only the vocabulary clinic staff actually need, prioritizing usefulness in day-to-day judgment over technical precision.
Disclaimer: This article provides general information. Technologies and services change. See also Using Generative AI in Healthcare.
1. Generative AI
A collective term for AI that creates—text, images, audio.
Earlier AI mainly classified or predicted: judging whether an X-ray shows an abnormality, forecasting next month's patient volume. Generative AI produces new text or images themselves.
In clinic work: producing a chart draft from the consultation, producing referral text from the chart, producing explanatory material for patients. Creating, not judging.
2. LLM (Large Language Model)
Generative AI that handles language. ChatGPT, Claude, and Gemini all use LLMs.
In one sentence: it has learned from enormous volumes of text and continuously calculates what word would naturally come next.
This matters practically because an LLM is not answering "what is true" but producing "natural text"—the root of hallucination below. Because it can produce plausible prose, errors do not look like errors.
3. Prompt
The instruction you give the AI. The same AI produces very different results depending on how you write it.
Assign a role. Opening with "You are a medical administrative staff member" tends to produce answers suited to that position.
State conditions concretely. Not "make it easy to understand" but "in language a middle schooler would understand, about 400 characters."
Specify the output shape. "Five bullet points," "as a table."
Critically, never put patient-identifying information in a prompt—names, dates of birth, addresses, phone numbers, chart numbers. Anonymize first. This belongs in written clinic policy.
4. Tokens
The unit in which AI processes text—roughly, chunks of words or characters. In Japanese one character may be more or less than one token, so it does not match a simple character count.
Two reasons to know this: it is the billing unit, and there is a ceiling on how much can be handled at once—which is context.
5. Context (Context Window)
The range of information the AI can read and consider at once. Think of it as how many documents fit on the desk at the same time. Anything outside that range, the AI has not seen.
Recent AI has expanded this range considerably, making it practical to feed in long documents or a patient's course. But a ceiling always exists:
- In long conversations, the earliest exchanges can fall out of range
- "Read all ten years of this chart" may not work straightforwardly
- When the AI answers about information not in range, it may be writing from guesswork
6. Hallucination
The AI writing something untrue, plausibly.
This is not a malfunction but an unavoidable property of how it works. As noted above, an LLM produces natural text; it is not searching for correct facts.
What deserves particular caution in healthcare: errors are written with confidence. It will not say "this might be wrong."
Three countermeasures:
- Supply grounded information. Summarizing supplied chart content leaves less room for invention
- Use it on the premise of human verification. Treat output as a draft; humans finalize
- Choose verifiable uses. Work you can check by reading (document drafts) is safe; finding things in large datasets makes errors hard to notice
See Where to Draw the Line on Delegating Work to AI.
7. Training (and how your input is handled)
Directly relevant to clinic policy. "Training" means two things:
- Pre-training: the AI acquiring capability from vast text when it was built
- Use of your input for training: what you enter being used to improve the service
The second is the issue. Services differ on whether input is used for training. Free consumer services may default to using it; corporate and healthcare plans commonly contract that it is not.
For healthcare use, what the contract says about your input not being used for training is a mandatory check.
8. API
The interface through which systems connect. Relevant when calling AI directly from your own systems. See What Is an API?.
9. MCP
A shared convention for connecting AI to external data and tools. A newer term, and the foundation for AI that "looks things up" or "operates" something. See What Is MCP?.
10. AI Agent
AI that, given an objective, assembles its own procedure and carries out multiple steps. Told to "prepare this patient's referral," it searches the chart, gathers what it needs, formats the document, and produces a draft—without being scripted.
See How Far Can AI Agents Replace Medical Administrative Work?.
Quick Reference
| Term | In a phrase | Where it matters |
|---|---|---|
| Generative AI | AI that "creates" | Drafting records and documents |
| LLM | Generative AI for language | What ChatGPT and others are |
| Prompt | The instruction | How you write changes results |
| Token | Unit of processing | Billing and ceilings |
| Context | What it reads at once | Whether long material fits |
| Hallucination | Plausible falsehood | Why verification is needed |
| Training | Whether input is reused | A contract checkpoint |
| API | Interface between systems | EMR integration |
| MCP | Standard connecting AI to tools | Foundation for AI that looks things up |
| AI agent | AI that assembles procedures | Workflow automation |
What These Ten Words Let You Judge
Memorizing vocabulary is not the point. The point is seeing what to verify when a vendor explains something.
- "The AI produces it automatically" → What about hallucination? Who verifies?
- "It can read large volumes" → What is the context ceiling? What happens beyond it?
- "It's secure" → Does the contract say input is not used for training?
- "It integrates with existing systems" → Is there an API? How far does integration go?
- "The AI looks things up automatically" → Connected to what data, and how?
Once the words mean something, these questions arise naturally—the best protection against a failed adoption.
Conclusion
- Generative AI creates rather than judges; an LLM is the language-handling kind
- LLMs produce natural text, not truth—the root of hallucination
- Prompts improve with role, concrete conditions, and output shape. Never include patient-identifying information
- Tokens are the processing and billing unit; context is what fits at once, and a ceiling always exists
- Whether input is used for training is a mandatory contract check in healthcare
- Understanding the vocabulary makes the right questions arise naturally
For details on AI Karte or to request a demo, please contact us.