40 AI Terms Explained: A Plain-English Glossary

7 min read · updated 2026-08-17 · free, no sign-up

The 10 core terms (start here)

If you only learn ten, learn these.

• 1. AI (Artificial Intelligence) — software that does tasks we used to think needed a human brain.
• 2. Model — the trained system itself. ChatGPT, Claude and Gemini are models (or families of them).
• 3. LLM (Large Language Model) 💰 — a model trained on huge amounts of text to predict the next word. This is what powers chatbots.
• 4. Prompt 💰 — the instruction you give the model. Better prompt, better output.
• 5. Token — a chunk of text (roughly ¾ of a word). Models are priced and limited by tokens.
• 6. Context window — how much text the model can "see" at once. Bigger window = it remembers more of your conversation.
• 7. Training — the process of teaching a model using data.
• 8. Inference — the model actually running and giving you an answer. Training is expensive; inference is what you pay for daily.
• 9. Hallucination — when the model states something confidently that's just wrong. Common. Always verify facts.
• 10. Fine-tuning — taking a general model and training it further on specific data to specialise it.

How models learn — 8 terms that demystify the black box

You don't need to code to understand these.

• 11. Machine Learning (ML) — the broader field. AI is the goal; ML is one way to get there.
• 12. Deep Learning — ML using many-layered neural networks. Behind most modern AI.
• 13. Neural Network — layers of simple math units loosely inspired by brain cells.
• 14. Parameters — the internal numbers a model adjusts during training. GPT-4-class models have hundreds of billions.
• 15. Dataset 💰 — the collection of examples used for training. Someone has to prepare and clean these.
• 16. Labeling / Annotation 💰 — marking up data so a model can learn from it. This is literally what DataAnnotation and Surge AI pay for.
• 17. RLHF (Reinforcement Learning from Human Feedback) 💰 — humans rank model outputs to teach it what "good" looks like. The core of most paid AI-training work.
• 18. Bias — when a model's outputs unfairly favour or disadvantage certain groups, usually because the training data did.

Working with AI — 8 terms you'll use weekly

These show up the moment you start using tools.

• 19. Chatbot — a conversational interface to a model (ChatGPT, Claude, Gemini).
• 20. API 💰 — a way to plug an AI model into other software. Freelance automation work usually means wiring APIs together.
• 21. Prompt engineering 💰 — the craft of writing prompts that reliably produce what you want.
• 22. System prompt — the hidden instruction that sets the model's role ("You are a helpful assistant…").
• 23. Temperature — a setting from 0 to ~1 controlling how random the output is. Low = predictable, high = creative.
• 24. Multimodal — a model that handles text, images, audio or video together.
• 25. Embedding — turning text into a list of numbers so a computer can compare meanings.
• 26. RAG (Retrieval-Augmented Generation) 💰 — giving the model access to your documents before it answers. Huge in business automation work.

Building with AI — 7 terms worth knowing

You'll hear these in job posts on Upwork.

• 27. Agent 💰 — an AI that can take actions (search the web, send emails, run code), not just chat.
• 28. Workflow / Automation 💰 — chaining AI with other tools to do a job end-to-end. Tools like n8n, Make and Zapier live here.
• 29. Vector database — a database designed to store embeddings. The plumbing behind RAG.
• 30. Open source model — a model whose weights you can download and run yourself (Llama, Mistral, DeepSeek).
• 31. Closed / proprietary model — you can only use it via someone else's service (GPT-4, Claude).
• 32. Guardrails — rules that keep a model from doing unsafe or off-topic things.
• 33. Evaluation (evals) 💰 — measuring whether a model actually does the job. Increasingly paid work.

The money terms — 7 you'll see on job platforms

Straight from real listings.

• 34. Data labeler — the general job title on platforms like DataAnnotation, commonly reported at around $20/hr base, with specialist tasks higher (source).
• 35. AI trainer — often means RLHF work. Independent reviews put verified general-task pay on Surge AI at $14–$20/hr (source) — below the $25–$30 you'll see quoted elsewhere.
• 36. Domain expert — someone with professional knowledge (medicine, law, finance). Mercor reportedly pays these from ~$25/hr up to $200/hr (source). Top rates go to existing expertise, not to beginners.
• 37. Freelance AI work — Upwork's own guidance: entry-level $10–$25/hr, intermediate $25–$75/hr, specialised AI/dev $75–$150+/hr (source).
• 38. Productised gig — a fixed AI offer on Fiverr. Sellers report meaningful revenue only after several completed orders build ranking (source).
• 39. Assessment — the unpaid test gating entry to DataAnnotation and Surge. This is the most common failure point.
• 40. Availability — how many hours the platform actually gives you. Two hours at $30/hr is less than eight at $15/hr. Most people earn little or nothing at first.

So — is learning this worth your time?

Honest answer: it depends on which term above made you sit up.

If #16 labeling or #17 RLHF did, the fastest path is an assessment on DataAnnotation or Surge — days to weeks to a first payout, but acceptance is not guaranteed.

If #21 prompt engineering, #26 RAG or #28 automation did, you're looking at Upwork or Fiverr. No gate to join; the real barrier is winning your first contract with no reviews. Expect 1–4 weeks and a lot of proposals before the first win.

If #36 domain expert described you and you already have a profession, Mercor is worth an application — but treat it as a month-3 target, not week-1.

What's not honest is telling you a number you'll hit per month. Nobody controls that. The terms are learnable in an afternoon. Whether the work shows up is a separate question.

Free forever · no account needed · nothing you make is shared