Vocabulary for Talking About AI and Machine Learning at Work
Essential English vocabulary for AI and machine learning conversations at work: models, training, inference, prompts, evaluation, and the phrases to use in meetings.
AI is now part of almost every engineering conversation, whether you build models or simply integrate an API. To take part confidently, you need the right vocabulary and the ability to use it naturally in meetings. This guide covers the core terms, common collocations, and example sentences you can borrow directly.
Core Concepts
| Term | Meaning |
|---|---|
| Model | A program that learns patterns from data to make predictions. |
| Training | The process of teaching a model using data. |
| Inference | Using a trained model to produce an output (a “prediction”). |
| Dataset | The collection of examples used for training or testing. |
| Parameters / weights | The internal numbers a model learns. |
| Fine-tuning | Adapting a pre-trained model to a specific task. |
| Prompt | The input text given to a language model. |
| Token | A chunk of text the model processes (roughly a word-part). |
“We’re not training a model from scratch — we’re fine-tuning an existing one on our support tickets.”
Talking About Quality
Teams obsess over how well a model performs. Know these terms:
- Accuracy — how often the model is correct.
- Precision — of the items flagged positive, how many truly are.
- Recall — of the truly positive items, how many were found.
- Hallucination — when a language model produces confident but false information.
- Drift — when model performance degrades over time as the real world changes.
- Benchmark — a standard test used to compare models.
“Precision is high, but recall is poor — we’re missing a lot of real fraud cases.” “The model is hallucinating product names that don’t exist, so we’ve added a retrieval step.”
The Language of Large Language Models
| Term | Meaning |
|---|---|
| LLM | Large language model. |
| Context window | How much text the model can consider at once. |
| Prompt engineering | Crafting inputs to get better outputs. |
| RAG | Retrieval-augmented generation — feeding the model relevant documents. |
| Embedding | A numerical representation of text used for search. |
| Temperature | A setting controlling how random the output is. |
| Guardrails | Rules that constrain what the model can output. |
“We’re using RAG so the model answers from our own docs rather than its training data.” “Lower the temperature if you want more deterministic, consistent answers.”
Verbs You’ll Hear Constantly
- to train a model
- to fine-tune on a dataset
- to deploy a model to production
- to evaluate against a benchmark
- to prompt the model
- to ground the answer in real data
- to serve inference requests
“Once we’ve evaluated it against the benchmark, we’ll deploy it behind a feature flag.”
Useful Phrases for Meetings
When you want to raise a concern:
“My concern is that the training data may not represent our actual users.” “How confident are we in these accuracy numbers? What’s the size of the test set?”
When you want to manage expectations:
“This is a probabilistic system, so it will occasionally get things wrong — we need a human-in-the-loop for high-stakes cases.”
When you propose an approach:
“Rather than fine-tuning, I’d suggest we start with prompting and RAG — it’s cheaper to iterate on.”
Words People Often Confuse
| Often confused | Clarification |
|---|---|
| AI vs ML | ML is one approach to building AI. |
| Training vs inference | Training builds the model; inference uses it. |
| Parameter vs hyperparameter | Parameters are learned; hyperparameters are set by you. |
| Supervised vs unsupervised | Supervised uses labelled data; unsupervised finds patterns without labels. |
Getting these distinctions right signals that you understand the field, not just the buzzwords.
Hedging Language for Uncertainty
AI outputs are uncertain, and English has precise ways to express that:
- “The model tends to struggle with long documents.”
- “It occasionally produces incorrect citations.”
- “Results vary depending on the prompt.”
- “We can’t guarantee a correct answer every time.”
Avoid absolute claims like “the AI is always right” — they damage your credibility with technical and non-technical stakeholders alike.
A Quick Self-Test
Try explaining each of these in a single sentence: hallucination, fine-tuning, context window, recall. If you can do that clearly, you can hold your own in almost any AI conversation at work.
With this vocabulary you can move from nodding along to actively shaping AI discussions. Use the example sentences as templates, label your uncertainty honestly, and keep the distinctions between training and inference, precision and recall, crisp. Speaking precisely about AI is itself a competitive advantage in today’s engineering teams.