AI literacy glossary

Core AI terms in plain language — what each means and where it shows up in practice

Terms

  • LLM (large language model)

    A model generating text, trained on huge corpora. ChatGPT, Claude, Gemini. Doesn't 'know' — it statistically continues text.

  • Prompt

    Your request to the model. 80% of answer quality is prompt quality: context, format, examples, constraints.

  • Context window

    How much text the model 'sees' at once — from ~8k to 1M tokens. Long documents get truncated from the start.

  • Token

    The text unit for models ≈ 3/4 word in English, ≈ 1/2–1 word in Ukrainian. API pricing is per token.

  • Hallucination

    A confident but fabricated answer — facts, figures, citations. Cause: models optimize for plausibility, not truth.

  • RAG (retrieval + generation)

    The model first searches YOUR documents, then answers with citations. The basis of corporate bots.

  • Fine-tuning

    Training a model further on your data for style/format. Doesn't add facts (use RAG for that), adjusts tone.

  • Embeddings

    Turning text into number vectors — 'coordinates of meaning'. Powers semantic search and RAG.

  • Multimodality

    The model handles text + images/audio/video in one request. GPT-4o, Gemini are multimodal.

  • Agent

    A model given tools (search, calculator, APIs) and a think → act → verify loop. Can run multi-step tasks.

  • Temperature

    Randomness parameter: 0 — stable answers (facts/code), 0.7–1 — creative (copy).

  • Zero-shot / few-shot

    Zero-shot — task with no examples; few-shot — 2–5 examples in the prompt sharply improve quality.

How to use it

  • 1.Ask the model to cite sources — then verify them: the citation itself can be a hallucination
  • 2.Facts, figures, dates, names — always verify manually before publishing
  • 3.Don't paste client/employee personal data into public chats without enterprise protection
  • 4.Knowing terms saves money: 'RAG' and 'fine-tuning' in an estimate differ 10x in budget

The glossary is deliberately simplified: the goal is to understand what is being sold to you and ask the right questions, not to learn the math. For depth — read specific model documentation.