Artificial intelligence vocabulary has an unusual power to mislead because so many of its technical terms are already human words. A model learns, is trained, has attention, uses memory and may hallucinate. These words are useful shorthand, but they can quietly smuggle human psychology into a description of computation.
AI has always borrowed from human cognition
Early artificial-intelligence research deliberately asked whether machines could perform tasks associated with reasoning, problem-solving or intelligence. Terms such as neural network also borrowed inspiration from biology. The relationship is analogical: artificial neural networks are mathematical systems, not small digital brains.
As the field matured, some metaphors became precise technical labels even though their everyday associations remained much broader.
Training is optimisation, not schooling
In machine learning, training generally means adjusting model parameters using data and an optimisation process. The model is not attending lessons in the human sense. Learning describes changes in statistical behaviour produced by that process.
This distinction matters when discussing what a system “knows”. Performance can reflect patterns in training data without implying human-like understanding, experience or belief.
Attention became a mathematical mechanism
In modern neural networks, attention refers to mechanisms that weight relationships among representations, especially in sequence modelling. The term was inspired by selective emphasis, but it does not mean a model consciously notices something.
Transformer, the architecture name associated with attention-based sequence models, is another example of a word whose technical meaning is now mostly detached from ordinary transformation.
Hallucination is memorable but controversial shorthand
When a generative model produces plausible-sounding content that is unsupported, fabricated or inconsistent with the source, people often call it a hallucination. The metaphor is vivid, but human hallucination is a perceptual experience. A language model does not need to have such an experience in order to generate an incorrect statement.
Alternative terms such as fabrication or confabulation have their own problems. The lesson is to define the failure behaviour rather than infer a mental state from the label.
New vocabulary changes public expectations
Words such as agent, reasoning, memory and autonomy can make systems sound more independent or human than their actual design warrants. Conversely, overly mechanical language can hide genuinely complex capabilities.
Good AI communication therefore separates the technical definition, observed behaviour and metaphor. That makes it easier to discuss both capabilities and limitations without either mystification or dismissal.
A five-word trail through this subject
These terms are worth opening next because each one carries part of the history or distinction described above.
- training: Parameter adjustment using data and optimisation, not human schooling.
- model: A computational system whose outputs reflect learned statistical structure.
- attention: A mathematical weighting mechanism, not conscious focus.
- token: A unit into which text or other data is segmented for processing.
- hallucination: A metaphor for unsupported or fabricated model output, not a human perceptual experience.
How to read the vocabulary more deeply
For each specialist term, compare its ordinary meaning, historical origin and current technical definition. Then look at its nearest neighbours. The most revealing question is often not “What does this word mean?” but “Why does this field need this word when everyday English already has another one?” That comparison exposes the distinctions the discipline considers important.
The idea to keep
AI vocabulary is full of human metaphors because human language is the toolkit available for describing new systems. Use the metaphor, but always return to the mechanism before drawing conclusions about what the machine is doing.