Top 100 Vocabulary for Adults | AI & Machine Learning Professionals
AI and machine learning vocabulary is the language of systems that learn patterns from data and then act under uncertainty. Professionals need to distinguish training from inference, benchmark from real-world performance, model capability from product reliability, and fluent output from grounded evidence.
This profession-specific flagship belongs to the eduKate Adult Vocabulary for Professionals system. It complements Data & Analytics Professionals, Software & Technology Professionals and Cybersecurity Professionals.
The Four Banks
Models & Learning: artificial intelligence, machine learning, model, algorithm, parameter, feature, label, training, inference, supervised learning, unsupervised learning, reinforcement learning, classification, regression, clustering, neural network, transformer, embedding, token, context window, fine-tuning, pretraining, loss, optimisation, gradient.
Data & Evaluation: dataset, training set, validation set, test set, ground truth, annotation, sampling, imbalance, leakage, benchmark, metric, accuracy, precision, recall, F1 score, calibration, robustness, generalisation, overfitting, underfitting, hallucination, uncertainty, confidence, evaluation, red teaming.
Generative AI & Agents: prompt, completion, system instruction, tool call, retrieval, RAG, vector database, semantic search, agent, workflow, planning, memory, grounding, citation, guardrail, structured output, multimodal, temperature, sampling, reasoning, chain, orchestration, latency, token cost, context.
Deployment, Safety & Operations: deployment, endpoint, API, model serving, throughput, latency, scaling, monitoring, drift, observability, logging, feedback loop, human-in-the-loop, fallback, rollback, versioning, governance, privacy, security, bias, fairness, explainability, alignment, safety, reliability.
Top 100 AI & Machine Learning Vocabulary: Working Meanings
| # | Word | Professional meaning |
|---|---|---|
| 1 | Artificial intelligence | Computer systems performing tasks associated with human cognitive capabilities. |
| 2 | Machine learning | Methods allowing systems to learn patterns from data rather than rely only on explicit rules. |
| 3 | Model | A learned mathematical representation used to make predictions or generate outputs. |
| 4 | Algorithm | A defined computational procedure for solving a problem. |
| 5 | Parameter | A learned numerical value controlling model behaviour. |
| 6 | Feature | An input variable used by a model. |
| 7 | Label | A target value associated with a training example. |
| 8 | Training | The process of adjusting model parameters using data and an objective. |
| 9 | Inference | Using a trained model to produce predictions or outputs. |
| 10 | Supervised learning | Learning from examples paired with target labels. |
| 11 | Unsupervised learning | Learning patterns from data without explicit target labels. |
| 12 | Reinforcement learning | Learning behaviour through interaction and reward signals. |
| 13 | Classification | Predicting a category. |
| 14 | Regression | Predicting a numerical value. |
| 15 | Clustering | Grouping similar observations without predefined labels. |
| 16 | Neural network | A layered computational model composed of interconnected learned units. |
| 17 | Transformer | A neural-network architecture based heavily on attention mechanisms. |
| 18 | Embedding | A numerical vector representing semantic or structural information. |
| 19 | Token | A unit of text or another modality processed by a model. |
| 20 | Context window | The amount of input and generated content a model can consider at once. |
| 21 | Fine-tuning | Additional training adapting a pretrained model to particular tasks or behaviours. |
| 22 | Pretraining | Large-scale initial training before task-specific adaptation. |
| 23 | Loss | A numerical objective representing model error during training. |
| 24 | Optimisation | The process of adjusting parameters to reduce loss or improve an objective. |
| 25 | Gradient | Information describing how model parameters should change to alter the objective. |
| 26 | Dataset | A structured collection of examples used for training or evaluation. |
| 27 | Training set | Data used to fit model parameters. |
| 28 | Validation set | Data used to tune choices during model development. |
| 29 | Test set | Held-out data used for final evaluation. |
| 30 | Ground truth | The reference answer or outcome treated as correct for evaluation. |
| 31 | Annotation | Human- or machine-assigned labels or metadata added to data. |
| 32 | Sampling | Selecting data points or model outputs according to a procedure. |
| 33 | Imbalance | Unequal representation of categories or outcomes in a dataset. |
| 34 | Leakage | Unintended use of information unavailable at real prediction time. |
| 35 | Benchmark | A standard test used to compare model performance. |
| 36 | Metric | A quantitative measure of model performance. |
| 37 | Accuracy | The proportion of predictions that are correct. |
| 38 | Precision | The proportion of positive predictions that are truly positive. |
| 39 | Recall | The proportion of true positives that are successfully identified. |
| 40 | F1 score | The harmonic mean of precision and recall. |
| 41 | Calibration | The degree to which predicted confidence matches observed outcomes. |
| 42 | Robustness | Ability to maintain performance under variation or perturbation. |
| 43 | Generalisation | Ability to perform well on previously unseen data. |
| 44 | Overfitting | Learning training-specific patterns that do not generalise well. |
| 45 | Underfitting | Failing to learn enough structure to perform well even on training data. |
| 46 | Hallucination | A generated output containing unsupported or fabricated information. |
| 47 | Uncertainty | Incomplete knowledge about model correctness or future outcomes. |
| 48 | Confidence | A model or system estimate of certainty, which may or may not be calibrated. |
| 49 | Evaluation | Structured measurement of model behaviour against defined criteria. |
| 50 | Red teaming | Adversarial testing intended to uncover failures, misuse or vulnerabilities. |
| 51 | Prompt | Input instructing or conditioning a generative model. |
| 52 | Completion | Text or other content generated in response to input. |
| 53 | System instruction | A high-priority instruction defining model behaviour and constraints. |
| 54 | Tool call | A structured request from a model to an external function or service. |
| 55 | Retrieval | Fetching relevant external information for use during generation. |
| 56 | RAG | Retrieval-augmented generation: combining external retrieval with model generation. |
| 57 | Vector database | A database designed to store and search numerical embeddings. |
| 58 | Semantic search | Retrieval based on meaning rather than exact keyword matching alone. |
| 59 | Agent | A model-driven system that can plan, choose actions and use tools toward a goal. |
| 60 | Workflow | A defined sequence of AI and non-AI steps producing an outcome. |
| 61 | Planning | Selection and ordering of actions toward a goal. |
| 62 | Memory | Stored information used to preserve relevant context across steps or interactions. |
| 63 | Grounding | Connecting generated outputs to reliable external evidence or state. |
| 64 | Citation | A reference linking a claim to supporting source material. |
| 65 | Guardrail | A control constraining model or application behaviour. |
| 66 | Structured output | Model output conforming to a predefined schema or format. |
| 67 | Multimodal | Capable of processing or generating more than one type of media. |
| 68 | Temperature | A sampling parameter influencing output randomness. |
| 69 | Sampling | The process of selecting the next output from a probability distribution. |
| 70 | Reasoning | Internal or explicit computation used to reach a conclusion. |
| 71 | Chain | A sequence of model, tool or transformation steps. |
| 72 | Orchestration | Coordination of models, tools and workflows. |
| 73 | Latency | The delay between request and usable output. |
| 74 | Token cost | The computational or monetary cost associated with processed tokens. |
| 75 | Context | Information available to the model for interpreting the current task. |
| 76 | Deployment | Putting a model or AI application into operational use. |
| 77 | Endpoint | A network-accessible interface for invoking a model or service. |
| 78 | API | An application programming interface allowing software systems to interact. |
| 79 | Model serving | Infrastructure used to deliver model inference to applications. |
| 80 | Throughput | The number of requests or tokens processed over time. |
| 81 | Scaling | Increasing system capacity to handle more workload. |
| 82 | Monitoring | Ongoing observation of model and application behaviour. |
| 83 | Drift | Change in data, behaviour or relationships that may reduce model performance. |
| 84 | Observability | The ability to understand internal system state from logs, metrics and traces. |
| 85 | Logging | Recording system events and outputs for analysis. |
| 86 | Feedback loop | A process using observed outcomes to improve future behaviour. |
| 87 | Human-in-the-loop | A system design requiring human review or decision at specified points. |
| 88 | Fallback | An alternative action used when the preferred model or workflow fails. |
| 89 | Rollback | Reverting to an earlier model or system version after problems. |
| 90 | Versioning | Tracking distinct model, prompt, data or application versions. |
| 91 | Governance | The structure of ownership, oversight and accountability for AI systems. |
| 92 | Privacy | Appropriate collection, use and protection of personal information. |
| 93 | Security | Protection of AI systems, data and users from malicious or accidental harm. |
| 94 | Bias | Systematic patterns producing distorted or unfair outcomes. |
| 95 | Fairness | Assessment of whether outcomes treat relevant groups appropriately under stated criteria. |
| 96 | Explainability | The degree to which model behaviour or decisions can be meaningfully understood. |
| 97 | Alignment | The degree to which model behaviour matches intended objectives and constraints. |
| 98 | Safety | Reduction of unacceptable risk from AI system behaviour or use. |
| 99 | Reliability | Ability of an AI system to perform as intended consistently. |
| 100 | Model risk | The possibility of harm or loss arising from incorrect, unstable or misused model behaviour. |
Benchmark Performance Is Not Product Reliability
A model can score well on a benchmark and still fail in production because users ask different questions, retrieval is weak, data drifts, latency is unacceptable or the application lacks fallbacks. The unit of reliability is the whole system, not just the model.
Scenario: Fluent but Unsupported Answer
Check whether the model had access to the required evidence, whether retrieval surfaced it, whether citations actually support the claim and whether the workflow should have abstained. Fluency is a presentation property; grounding is an evidence property.
Seven-Day AI & Machine Learning Vocabulary Plan
| Day | Practice |
|---|---|
| 1 | Separate model, training, inference and evaluation. |
| 2 | Map datasets, labels, leakage and generalisation. |
| 3 | Compare accuracy, precision, recall and calibration. |
| 4 | Build a retrieval-grounded generative workflow. |
| 5 | Audit one AI application for fallback, monitoring and safety. |
| 6 | Recall 75+ AI terms by function. |
| 7 | Write a one-page AI system review separating model capability from operational reliability. |
Continue the Digital Profession Wing
- Cybersecurity Professionals
- Data & Analytics Professionals
- Product Management Professionals
- Software & Technology Professionals
Conclusion
AI and machine learning vocabulary helps professionals make a fast-moving field more exact. It separates model from application, training from inference, confidence from correctness, and impressive demos from systems that can be trusted under real conditions.