Plain-language glossary
AI words without the fog.
These explanations are deliberately practical. They are not a complete computer-science course; they are the concepts that help you decide what to ask for, what to trust, and what to check.
Model
A model is a trained pattern-making system. It turns input into a likely continuation or action based on what it learned, not because it has a human-like guarantee of truth.
Generative AI
AI that produces new text, code, images, audio, or other content. “New” does not mean automatically original, accurate, or safe; the output still needs evaluation.
Context window
The working space available for the current conversation, files, instructions, and tool results. More context is not always better if important rules become hard to find.
Hallucination
A plausible-looking claim, citation, code path, or detail that is unsupported or false. Grounding and verification reduce the risk; they do not make it disappear.
Grounding
Connecting an answer to reliable evidence: a repository, opened web source, database, test output, or user-provided document. Grounding is the bridge from “sounds right” to “supported here.”
Retrieval
Finding relevant information and placing it into the model’s working context. Retrieval helps only if the source is relevant, current, complete enough, and interpreted correctly.
From chatbot to agent
A chatbot mainly responds. An agent can plan a path, use tools, observe results, adjust, and repeat until the task is complete or it needs human input. A workflow is more prescribed: the software decides the sequence and the model fills in parts of it. A useful system chooses the simplest pattern that handles the job.
Tools, memory, and autonomy
| Term | In plain English | Question to ask |
|---|---|---|
| Tool calling | The model requests an external operation such as searching, reading a file, running a test, or sending data. | What can the tool change, and what requires approval? |
| Memory | Information preserved beyond the immediate exchange, either intentionally or by a product feature. | What is stored, for how long, and can I correct or remove it? |
| RAG | Retrieval-augmented generation: retrieve source material, then generate an answer using it. | Can I see the source passages and their dates? |
| Reasoning effort | A setting that can trade latency/cost for more internal work. Higher is not automatically better. | Did representative evaluations show a meaningful gain? |
| Guardrail | A boundary in instructions, tools, permissions, validation, or code that prevents or catches unsafe behavior. | Is it enforced where the action happens, or merely suggested to the model? |
| Eval | A repeatable test of expected behavior across representative cases. | Does it include failures, near-misses, and regressions—not only success cases? |
Human control is a design feature
- Give the system access only to what the task needs.
- Separate read actions from write, publish, financial, deletion, or communication actions.
- Require confirmation where the impact is meaningful or recovery is difficult.
- Make uncertainty and source quality visible.
- Keep an audit trail for consequential actions.
- Use privacy, safety, fairness, and reliability as requirements—not as a last-minute polish pass.
For a broader risk vocabulary, the NIST AI Risk Management Framework is a useful source. It is voluntary guidance, not a magic certificate.