AI vocabulary sounds deceptively human. Models are trained, systems learn, networks have attention, and outputs may be called hallucinations. These metaphors are useful shorthand, but beginners understand AI more accurately when they keep the human association separate from the technical mechanism.
Artificial intelligence is discussed as though everyone shares the same vocabulary, yet familiar words such as model, learning and training have specialised meanings. Understanding the terms makes it easier to judge what a system can do, what it cannot do and where confident claims require evidence.
Artificial intelligence and machine learning
Artificial intelligence (AI) is a broad label for computer systems designed to perform tasks associated with capabilities such as perception, language, prediction, planning or decision-making.
Machine learning is an approach in which a system learns patterns from data rather than relying only on rules written explicitly for every case. Not all AI uses machine learning, and not every machine-learning model behaves like a conversational assistant.
Algorithm and model
An algorithm is a defined computational procedure or set of steps. A model is a mathematical or computational representation used to make predictions, classifications or generated outputs.
The distinction is simplified in everyday speech, but useful: an algorithm describes a process; a trained model contains patterns or parameter values learned through that process.
Training data, features and labels
Training data is the material used to adjust a model. Its quality, coverage and provenance influence what the model learns.
A feature is an input variable or representation used by a model. A label is the target answer attached to an example in supervised learning, such as whether an email is spam.
Data is not neutral simply because it is large. Historical bias, missing groups, errors and uneven sampling can all affect performance.
Parameters and training
Parameters are values inside a model that are adjusted during training. In a neural network, they influence how strongly information passes through the network’s connections.
Training is the process of updating these parameters so that the model performs better on an objective. The model is not being taught in exactly the way a person is taught. The word describes an optimisation process.
Neural networks and deep learning
A neural network is a model built from layers of connected computational units. The name is historically inspired by biological neurons, but a digital neural network should not be mistaken for a working replica of a human brain.
Deep learning uses neural networks with multiple layers to learn complex representations. It has been particularly influential in image, speech and language systems.
Generative AI and large language models
Generative AI produces new content such as text, images, audio or code by learning statistical patterns in existing data.
A large language model (LLM) is trained to process and generate language, commonly by predicting tokens in context. Its responses can be fluent because it has learned intricate patterns of language. Fluency does not guarantee truth, judgement or human-like understanding.
Tokens, prompts and context
A token is a unit into which text is divided for processing. A token may be a whole short word, part of a longer word, punctuation or another text fragment.
A prompt is the input or instruction given to a generative system. The context window is the amount of material the model can take into account in one interaction. Clear prompts help, but they cannot make an unsuitable model reliably know information it lacks.
Inference
Inference is the stage at which a trained model is applied to new input to produce an output. Training builds or adjusts the model; inference uses it.
In a language model, inference may generate one token at a time according to probabilities influenced by the prompt and preceding context.
Fine-tuning and retrieval
Fine-tuning is additional training that adapts a model to particular behaviours, tasks or domains.
Retrieval-augmented generation (RAG) gives a generative model relevant information retrieved from external sources before it answers. Retrieval can improve grounding and freshness, but the system can still select poor sources, misunderstand them or produce an unsupported synthesis.
Overfitting and generalisation
Overfitting occurs when a model fits its training data too closely and performs poorly on genuinely new examples. Generalisation is the ability to perform well beyond the examples seen during training.
A high score on familiar benchmark data does not automatically show robust performance in a different population or real-world setting.
Hallucination
A hallucination is a plausible-looking output that is false, fabricated or unsupported by the available evidence. The term can make the behaviour sound mysterious, but the practical point is simple: generative systems produce likely sequences, not guaranteed facts. Important claims need verification.
Bias, fairness and explainability
Bias can refer to systematic error in data, design or output. Fairness concerns how benefits, errors and decisions are distributed across people and groups. There is no single fairness measure that resolves every trade-off.
Explainability concerns whether people can understand why a system produced an output. An explanation may describe influential inputs, a decision rule or model behaviour, but a persuasive explanation is not necessarily a complete causal account.
Benchmark and evaluation
A benchmark is a standard task or dataset used to compare systems. Evaluation should also consider reliability, safety, cost, latency, accessibility and performance in the actual context of use.
Use the vocabulary to ask better questions
When an AI claim sounds impressive, ask: What data was used? What task was measured? How was the model evaluated? Does it generalise? What happens when it is wrong? Precise vocabulary does not remove uncertainty, but it makes the uncertainty visible.
Always ask for the operational meaning
When an AI term sounds psychological, translate it into what the system actually does. What changes during “learning”? What calculation implements “attention”? What output behaviour is being labelled a “hallucination”? The mechanism protects you from taking the metaphor too literally.