Artificial intelligence vocabulary is easier to understand when every technical term connects to a real task. These 20 AI words explain what people mean by a model, dataset, prompt, token, inference, hallucination, retrieval and agent, without assuming you already work in technology.
If you see a headline saying an AI model passed a benchmark, was fine-tuned on data or uses retrieval to ground its answers, this page shows what those words actually mean—and what they do not prove. The goal is to help you read AI news, use AI tools and question confident claims without being misled by jargon.
Quick answer: an AI model uses learned patterns to produce an output for a given input. Its usefulness depends on the task, data, design, testing and safeguards. Learn the terms below in context, distinguish what a system can generate from what it can verify, and check important outputs against reliable sources.
Quick answer: the 20 words at a glance
Use this glossary as a translation guide to AI documentation, news articles and product demonstrations. The meanings are deliberately concise; each entry below adds a real-world example and an important limit.
| Word or term | Plain-English meaning |
|---|---|
| Model | A trained computational system that maps inputs to outputs. |
| Training | The process of adjusting a model using data and an objective. |
| Dataset | An organised collection of examples or records. |
| Parameter | A numerical value inside a model that affects its behaviour. |
| Inference | The act of using a trained model to produce an output. |
| Prompt | An instruction or input provided to a model. |
| Token | A unit of text representation processed by many language models. |
| Context window | The amount of input and sometimes generated content a model can consider in one run. |
| Hallucination | An AI-generated statement that is false, fabricated or unsupported. |
| Grounding | Tying an AI output to relevant, checkable input or evidence. |
| Retrieval | Finding relevant material from an external collection. |
| Fine-tuning | Further adapting a trained model using additional training. |
| Multimodal | Able to work with more than one type of input or output. |
| Agent | A system that uses a model, tools and control logic to pursue tasks. |
| Evaluation | A systematic check of how a system performs on defined tasks. |
| Benchmark | A common test or set of tasks used to compare performance. |
| Latency | The delay between a request and a response. |
| Bias | Systematic patterns of unfairness or skew in data or outputs. |
| Privacy | Control over personal information and how it is collected, used and shared. |
| Automation | Having a system perform steps with reduced manual effort. |
What is artificial intelligence vocabulary?
Artificial intelligence vocabulary is the set of terms used to describe how AI systems are built, supplied with information, tested, deployed and used. Some words name technical processes; others describe limitations or risks. Understanding the distinction matters: “training” and “inference” are different stages, for example, and “benchmark score” is not a promise that a system will work equally well for every user.
How to use this list in real life
When a company says its AI can perform a task, separate the claim into five questions. You do not need to become a programmer to ask them; you only need vocabulary precise enough to locate where the uncertainty lies.
- What is the input: text, images, audio, a dataset or another source?
- What is the model expected to do: classify, predict, draft, search, summarise or take action?
- Where does its information come from: prior training, the current prompt, retrieved records or an external tool?
- What evidence supports its performance: task-specific evaluation, independent tests or only a product demonstration?
- What happens when it makes a mistake: can a human review, correct, stop or undo the result?
An important rule is that fluent output is not the same as verified knowledge. AI-generated text may sound certain even when the underlying information is missing or wrong. Use the terms “grounding”, “evaluation” and “privacy” as practical questions, not decoration.
The 20 vocabulary words, with examples and the mistakes to avoid
1. Model
Meaning: A trained computational system that maps inputs to outputs. Different models have different tasks and capabilities.
Say or write: “This model can summarise documents, but we still check the summary.”
Watch the difference: A model is not the same thing as the application or website built around it.
Try this: Ask which model is used and what task it has been tested for.
2. Training
Meaning: The process of adjusting a model using data and an objective. Training happens before ordinary use, although some systems can be updated later.
Say or write: “The model was trained to recognise patterns in examples.”
Watch the difference: Using a chatbot does not necessarily train its model on your conversation.
Try this: Check the provider’s actual data-use and retention terms.
3. Dataset
Meaning: An organised collection of examples or records. Data quality and coverage influence performance.
Say or write: “The dataset included labelled examples of customer questions.”
Watch the difference: A large dataset is not automatically representative or lawful to use.
Try this: Ask who collected the data and which cases are missing.
4. Parameter
Meaning: A numerical value inside a model that affects its behaviour. Modern models may contain very large numbers of parameters.
Say or write: “The model has learned parameters during training.”
Watch the difference: More parameters alone do not prove better performance.
Try this: Look for real task results, not just model size.
5. Inference
Meaning: The act of using a trained model to produce an output. Inference is generally the stage users encounter.
Say or write: “The assistant performs inference when it generates a reply.”
Watch the difference: Inference is not the same thing as updating model weights through training.
Try this: Ask about response time and resource requirements for the use case.
6. Prompt
Meaning: An instruction or input provided to a model. Prompts may include examples, constraints or documents.
Say or write: “I prompted the model to draft a short summary for beginners.”
Watch the difference: A clever prompt cannot guarantee truth or override every limitation.
Try this: State the audience, task, source material and desired format.
7. Token
Meaning: A unit of text representation processed by many language models. A token is not always equal to one word.
Say or write: “The document used more tokens than the model’s input limit allowed.”
Watch the difference: Token counts vary by language and tokenisation scheme.
Try this: Check the model’s documented usage limits rather than estimating from words.
8. Context window
Meaning: The amount of input and sometimes generated content a model can consider in one run. Limits and counting rules differ by system.
Say or write: “The long report exceeded the available context window.”
Watch the difference: A large context window does not imply perfect recall or reasoning across it.
Try this: Break complex tasks into verifiable stages when needed.
9. Hallucination
Meaning: An AI-generated statement that is false, fabricated or unsupported. In technical risk discussions, this is sometimes called confabulation.
Say or write: “The answer included a hallucinated citation to a report that does not exist.”
Watch the difference: Confident wording is not evidence that a claim is true.
Try this: Verify dates, quotations, citations and high-stakes facts externally.
10. Grounding
Meaning: Tying an AI output to relevant, checkable input or evidence. Grounding aims to reduce unsupported generation.
Say or write: “We grounded the answer in the supplied policy document.”
Watch the difference: Even a quoted source can be misunderstood or out of date.
Try this: Open the cited passage and check that it actually supports the claim.
11. Retrieval
Meaning: Finding relevant material from an external collection. Retrieval may be combined with generation, often called RAG.
Say or write: “The assistant retrieved the manual before answering the repair question.”
Watch the difference: Retrieval is not the same as proving the source is correct.
Try this: Check source relevance, freshness and access permissions.
12. Fine-tuning
Meaning: Further adapting a trained model using additional training. It often targets a particular task, style or data distribution.
Say or write: “We fine-tuned a smaller model on approved examples.”
Watch the difference: Fine-tuning is different from simply adding information to a prompt.
Try this: Ask whether better instructions or retrieval would solve the task first.
13. Multimodal
Meaning: Able to work with more than one type of input or output. Examples include combinations of text, images and audio.
Say or write: “The multimodal system interpreted the chart image and the written question.”
Watch the difference: Not every multimodal model supports every format or output type.
Try this: Test each mode separately and check accessibility and privacy.
14. Agent
Meaning: A system that uses a model, tools and control logic to pursue tasks. Agents may read information, plan steps or take permitted actions.
Say or write: “The agent checked the inventory and prepared a draft order.”
Watch the difference: Calling something an agent does not prove autonomy or reliability.
Try this: Set permissions, review points and limits before allowing actions.
15. Evaluation
Meaning: A systematic check of how a system performs on defined tasks. Evaluation depends on good criteria and appropriate test cases.
Say or write: “The team evaluated whether summaries preserved key facts.”
Watch the difference: A single demo is not a robust evaluation.
Try this: Ask about failure rates, edge cases and human review.
16. Benchmark
Meaning: A common test or set of tasks used to compare performance. Benchmarks are useful but narrower than real life.
Say or write: “The model scored well on the benchmark used for reasoning tasks.”
Watch the difference: A high score can coexist with poor results on your own documents.
Try this: Try task-specific examples and inspect the method.
17. Latency
Meaning: The delay between a request and a response. Latency matters for live conversations and interactive software.
Say or write: “We measured the latency of responses during peak usage.”
Watch the difference: Fast systems are not necessarily more correct.
Try this: Decide whether your task needs speed, accuracy or both.
18. Bias
Meaning: Systematic patterns of unfairness or skew in data or outputs. Bias can enter through training data, design or deployment context.
Say or write: “The team checked whether the tool performed differently across groups.”
Watch the difference: Bias is not fully solved by one instruction in a prompt.
Try this: Test the relevant population and review consequential decisions.
19. Privacy
Meaning: Control over personal information and how it is collected, used and shared. Policies and technical protections vary by service.
Say or write: “We removed personal details before sending examples to the tool.”
Watch the difference: A private-looking interface does not prove data is not retained.
Try this: Read the provider’s privacy policy and avoid unnecessary sensitive data.
20. Automation
Meaning: Having a system perform steps with reduced manual effort. Automation can be rule-based, AI-assisted or a combination.
Say or write: “We automated the first draft of routine summaries for human review.”
Watch the difference: Automation is not proof of intelligent judgment.
Try this: Identify which steps need approval, logging and a human fallback.
Worked example: decode a news headline about an AI product
Imagine you read: “New multimodal AI agent beats a benchmark after fine-tuning and retrieval upgrades.” That sentence combines five different ideas. “Multimodal” concerns kinds of input or output. “Agent” suggests software can use tools or follow a workflow. “Benchmark” refers to a particular test. “Fine-tuning” changes a model through further training. “Retrieval” brings external information into the workflow.
None of these words, separately or together, tells you the system is always correct, safe for confidential data or effective on the problem you care about. To judge the announcement, ask for the evaluation setup, test data, error cases, privacy controls and whether humans can supervise tool actions.
Now consider a practical user question: “Can this AI summarise our 80-page policy guide and quote the exact sections?” A sensible test supplies the official guide, requests page-level references, checks three difficult sections manually and records any omitted details. That is a meaningful evaluation of one real task.
The takeaway is simple: use AI vocabulary to ask better questions. Knowing “token” will help you understand limits; knowing “hallucination” will remind you to verify; knowing “agent” will make you ask what the system is allowed to do.
Common mistakes and quick repairs
- Treating “AI”, “generative AI” and “agent” as interchangeable → identify whether the system predicts, generates or takes actions.
- Assuming a high benchmark score guarantees workplace performance → test with your documents and edge cases.
- Believing every cited AI answer → open the original evidence and check the passage.
- Assuming all chats are used to train a provider’s model → read the actual product and account privacy settings.
- Confusing retrieved material with a training update → ask what came from the prompt, a tool or fine-tuning.
- Sharing confidential records for a quick trial → remove sensitive details and use authorised systems.
A practical 15-minute vocabulary exercise
For fifteen minutes, examine a real AI product description. Underline one claim about the model, one about the source of information, one about speed and one about quality. Write a question for each: What model? Which source? What latency? Evaluated on which task?
Next, ask the tool to summarise a short document you already understand. Compare the output with the document. Note one correct point, one omission and one sentence that needs evidence. This exercise teaches how AI terms connect to actual reliability rather than just marketing vocabulary.
Frequently asked questions
What are the most important AI terms for beginners?
Start with model, training, dataset, prompt, inference, hallucination and evaluation. These explain the basic workflow and why outputs must be checked. Add retrieval and agent when using software that searches sources or operates tools.
What is the difference between AI and generative AI?
AI is a broad category of computational systems performing tasks associated with intelligence. Generative AI refers to systems that produce content such as text, images, audio or code. Not every AI system is generative.
Does a larger AI model always give better answers?
No. Model size is only one design characteristic. Data, task fit, evaluation, inference approach, tools and safeguards all matter. Check performance on the actual use case.
What does an AI hallucination mean?
It means an AI system produced content that is false, fabricated or unsupported. It can include invented sources, incorrect figures or misrepresented quotations. Verify consequential claims using evidence you can inspect.
What does AI agent mean for an ordinary user?
An AI agent usually means software arranged to pursue tasks using a model and sometimes tools. The important questions are what actions it can take, which permissions it has, and when a human must approve or correct it.
Further reading and reliable sources
- NIST Generative AI Profile — definitions and practical discussion of generative AI risks.
- NIST AI Risk Management Framework FAQ — guidance on valid, reliable, secure and accountable AI systems.
- NIST glossary: AI — authoritative broad definitions of artificial intelligence.
Keep exploring vocabulary on eduKateSG
Read the eduKateSG Vocabulary hub to move from meanings to accurate everyday use, browse Vocabulary Articles & Knowledge for more subject-specific language, or use the 2,000 essential adult English vocabulary words for a wider foundation.
A useful word is one you can understand, retrieve and use without forcing it into a sentence. Bookmark this page and return to the examples when the real situation comes up.