AI Vocabulary

AI Vocabulary

This is a current AI vocabulary guide written the way people run into AI at work: when they build a feature, judge output quality, pick tools, or manage risk.

If you’re searching for “AI vocabulary words with meaning”, you’re in the right place. The terms are grouped by how they show up in real conversations in 2024 and 2025, not by academic categories.

If you’re learning AI seriously, shipping products, or deciding what to adopt, this will help you follow discussions faster. This is also why structured learning paths like AI Course matter: if you don’t share the same words, you can’t share the same decisions.

12 Artificial Intelligence terms you need to know

If you only learn one set, start here. These are the words that come up every day:

• Artificial intelligence (AI)
  
• Machine learning (ML)
  
• Deep learning
  
• Model
  
• Training
  
• Inference
  
• Generative AI (GenAI)
  
• Large Language Model (LLM)
  
• Token
  
• Context window
  
• Prompt
  
• RAG (Retrieval-Augmented Generation)

AI basics

Artificial intelligence (AI)
AI is the umbrella term for systems that do tasks people usually do with judgment: language, perception, pattern spotting, and decision support. Most AI you touch today is narrow: built for a specific job.

Machine learning (ML)
ML is a part of AI where systems learn patterns from data instead of being hand-coded rule by rule. Most modern AI products run on ML.

Deep learning
Deep learning is ML built on neural networks with many layers. It powers speech recognition, image models, and modern language models.

Model
A model is a trained system that takes input and produces output: a prediction, a label, text, images, audio, or code. In practice, “which model should we use?” usually means tradeoffs between:

• Output quality
  
• Cost
  
• Speed
  
• Consistency
  

Training vs inference
Training is when the model learns from data. Inference is when you use the trained model to answer users. Most day-to-day cost and wait time problems show up during inference.

GenAI terms 

Generative AI (GenAI)
GenAI refers to models that create new content: text, images, audio, or code. Most popular AI products today fall here.

Large Language Model (LLM)
An LLM is trained on huge amounts of text and predicts the next token. That simple setup can produce summarizing, reasoning-like behavior, and code writing.

Token
A token is a chunk of text the model reads and writes. It might be a whole word, part of a word, or punctuation. Many limits and bills are tied to token counts.

Context window
The context window is the maximum number of tokens the model can consider at once. It usually includes:

• System rules
  
• Developer instructions
  
• User message
  
• Conversation history
  
• Retrieved snippets (if you use retrieval)
  

Prompt
A prompt is the input you give the model. In real products, “the prompt” often means a bundle: system rules + developer instructions + user request.

System instruction
System instructions are top-priority rules that set behavior, tone, and boundaries across a session.

Temperature
Temperature controls randomness. Lower temperature gives steadier outputs. Higher temperature gives more variation, plus more risk of wandering.

Hallucination
A hallucination is when a model produces confident output that is wrong or not backed by sources. It’s one reason teams care about retrieval and repeatable tests.

Grounding and knowledge

RAG (Retrieval-Augmented Generation)
RAG is when your system retrieves relevant material from docs or databases and feeds it into the model so answers stay tied to real text. It helps with:

• Fewer made-up claims
  
• More current info (if your docs are current)
  
• Clearer audit trails (you can show what was retrieved)
  

Embedding
An embedding is a numeric “fingerprint” of text or images. It lets systems search by meaning, not just keywords.

Vector database
A vector database stores embeddings and supports similarity search. It’s often used to power RAG.

Chunking
Chunking splits large documents into smaller pieces before creating embeddings. Good chunking helps retrieval hit the right passage instead of a random paragraph.

Agents and agent workflows

Agent
An agent is an LLM system that can do work, not just respond. It combines the model with tools, memory, and control logic to complete tasks.

Agent workflow
An agent workflow lets the system plan steps, take actions, check results, and try again. This is where “it can do the thing” comes from.

Tool use / tool calling
Tool use means the model can call external systems in a structured way, like:

• Search
  
• Databases
  
• Calendars
  
• Internal APIs
  
• Code runners
  

Orchestration
Orchestration is the glue that coordinates models, retrieval, tools, routing, and policies so the system behaves the same way across cases.

Multi-agent system
A multi-agent system uses several agents with different roles that collaborate or hand off tasks (for example: one agent gathers info, another writes, another checks).

Memory
Memory is how a system keeps information across turns or sessions.

• Short-term memory lives in the context window
  
• Long-term memory is stored outside (like a database) and retrieved when needed
  

Training and adaptation terms 

Fine-tuning
Fine-tuning is additional training on your data so a model better matches your domain, style, or task.

Instruction tuning
Instruction tuning trains a model to follow directions and common chat norms more consistently.

RLHF and RLAIF
RLHF uses human feedback to shape behavior. RLAIF uses AI-generated feedback. (You’ll also hear about approaches that use written principles rather than lots of example ratings.)

Evaluation and quality terms 

Eval (evaluation)
An eval is a repeatable test suite that measures performance across many realistic cases. It’s how teams stop arguing over a few screenshots.

Benchmark
Benchmarks like MMLU or GPQA compare models on standard tasks. They can help you get a rough sense of model behavior, but they don’t replace product-specific tests.

Rubric
A rubric is the scoring guide. It makes “good” and “bad” concrete so two reviewers (or a judge model) score outputs the same way.

Human in the loop
Humans review or approve outputs, often for high-stakes or regulated work.

Regression
A regression is when a change makes some outputs worse. Evals are built to catch regressions before users find them.

Tools and workflows here often overlap with skills covered in Tech Certification, especially for teams making architecture calls.

Safety and security words 

Goal fit
This means model behavior matches your intended goals and constraints: what it should do, what it must not do, and how it should respond when it’s unsure.

Jailbreak
A jailbreak is an attempt to bypass safety rules through clever prompting or manipulation.

Prompt injection
Prompt injection is when malicious instructions are hidden in user input or retrieved content to override system rules (for example, a doc snippet that says “ignore previous instructions”).

Overrefusal
Overrefusal is when the system blocks safe requests too often because safeguards are too strict or too blunt.

Deployment and cost terms 

Latency
Latency is how long responses take. Teams often split it into:

• Time to first token
  
• Total response time
  

Throughput
Throughput is how many requests the system can handle per second.

Cost per request
Cost per request depends on:

• Token counts in and out
  
• Model choice
  
• Tool calls (search, retrieval, code, etc.)
  

Observability
Observability means logging and monitoring what the system did so you can debug and improve it. It often includes:

• Prompts and system rules used
  
• Tool calls and outputs
  
• Errors and timeouts
  
• User feedback signals
  

These tradeoffs aren’t only technical. They affect product decisions, pricing, and growth, which is why they show up in Marketing and Business Certification conversations too.

AI Vocabulary IELTS

If you’re studying AI Vocabulary IELTS, focus on using terms in context, not listing definitions. Examiners reward clear use, not buzzwords.

Try answers built around real situations:

• “During inference, latency matters because users wait.”
  
• “RAG reduces hallucinations by grounding responses in documents.”
  
• “An eval catches regressions after a prompt change.”
  

A simple practice loop:

• Pick 5 terms
  
• Write 3 sentences for each that describe a real scenario
  
• Swap in simpler words where possible
  
• Read it out loud until it sounds normal

How to learn AI vocabulary?

If you want this vocabulary to stick, learn it through workflows, not flashcards.

Start with the words you see every day:

• Prompt
  
• Tokens
  
• Context window
  
• RAG
• Tool calling
  
• Evals
  

Then learn the pairs that show real tradeoffs:

• Context window vs RAG
  
• Temperature vs consistency
  
• Human eval vs automated eval
  

Translate each term into a question you can ask in a meeting:

• Are we answering from sources, or guessing?
  
• Can this agent take actions safely?
  
• Did the last change help across real cases?
  
• What did we trade to get lower latency?
  

Final takeaway

AI vocabulary is now basic job language for teams building, buying, or managing AI. Once you understand how these pieces connect—models, prompts, retrieval, tools, and evals—you can follow AI product discussions faster, spot weak claims sooner, and make clearer calls.

Blockchain is increasingly referenced in AI discussions due to its role in secure data sharing and model accountability. Key AI terminology often overlaps with concepts from Blockchain Technology such as decentralization and consensus mechanisms. Many learners strengthen their technical foundation by enrolling in a Blockchain Course to understand these cross domain concepts.