85 AI Terms Explained (In Plain English, For Beginners)

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

How to use this glossary

You do not need all 85 terms. Most people working with AI in a practical way — writing prompts, using tools, doing paid tasks on platforms like DataAnnotation or Mercor — use maybe 20 of them regularly. The rest are useful for reading job posts, understanding what a client wants, or following a tutorial without getting lost.

A fair way to read this page:

• Skim the bold terms. If one is unfamiliar, read its line.
• Star the ones that appear in something you're actually trying to do (a course, a gig description, a tool's docs).
• Ignore the rest for now. They'll come back when they matter.

Everything below is written for someone with no technical background. Where a term has a common myth attached, I've said so.

The 15 core terms (start here)

• AI (Artificial Intelligence): software that does tasks we used to think required human judgement.
• Machine Learning (ML): the main way modern AI works — the system learns patterns from examples instead of following hand-written rules.
• Model: the trained file that does the actual predicting. When people say "the model", they mean this.
• LLM (Large Language Model): a model trained on text. ChatGPT, Claude and Gemini are LLMs.
• Prompt: what you type in.
• Prompt engineering: the skill of writing prompts that get useful output. Less mystical than it sounds.
• Token: roughly a chunk of a word. Models charge and limit by tokens, not words.
• Context window: how much text a model can "see" at once. Bigger = can read longer documents.
• Training: teaching the model, done once by the company that built it.
• Inference: using the model, done every time you send a prompt.
• Fine-tuning: taking an existing model and nudging it toward a narrower task.
• Hallucination: the model confidently making something up. It happens. Assume it will.
• Generative AI: AI that produces new text, images, code or audio.
• API: a way for one piece of software to talk to another. "Using the OpenAI API" means a program (not a person) is sending prompts.
• Chatbot: a conversational interface on top of a model.

Model behaviour and quality (15 terms)

• Parameters: the internal numbers a model uses. More parameters ≠ better, but it's a rough size indicator.
• Weights: the specific values of those parameters after training.
• Temperature: a knob controlling how random the output is. Low = predictable, high = creative.
• Top-p / top-k: other randomness knobs. You rarely need to touch them.
• System prompt: hidden instructions the model gets before you start.
• Zero-shot: asking with no examples.
• Few-shot: asking with a couple of examples first. Often dramatically better.
• Chain-of-thought: asking the model to reason step by step.
• Grounding: making the model answer from a specific source, not its memory.
• RAG (Retrieval-Augmented Generation): giving the model your documents to answer from. The single most useful pattern for real work.
• Embedding: turning text into numbers so a computer can compare meaning. Powers search and RAG.
• Vector database: where those embeddings live.
• Guardrails: rules that block certain outputs.
• Alignment: the effort to make models do what humans actually want.
• Benchmark: a standard test used to compare models. Take them with salt.

Working-with-AI terms you'll see in job listings (15 terms)

These come up on Upwork, Fiverr, DataAnnotation, Mercor and Surge.

• RLHF (Reinforcement Learning from Human Feedback): humans rate model outputs, and the model learns from the ratings. This is what most "AI training" jobs are.
• Data annotation / labelling: tagging data so a model can learn from it.
• Red teaming: deliberately trying to break a model to find its flaws.
• Evaluation (evals): structured testing of model output quality.
• Preference ranking: picking which of two model answers is better. Common paid task.
• Prompt tuning: iterating prompts against a set of test cases.
• AI trainer: the umbrella term for the above roles.
• Domain expert: someone with real professional experience (medicine, law, finance) whose judgement AI labs pay for. Mercor's higher rates target this.
• Human-in-the-loop: a workflow where a person checks or corrects AI output.
• Content moderation: reviewing outputs for safety or policy issues.
• Automation: stringing tools together so something runs without you.
• Workflow: the sequence of steps in that automation.
• No-code / low-code: building software with visual tools instead of writing code.
• Agent: an AI system that takes actions, not just answers.
• Copilot: an AI that assists a human doing a task, rather than replacing them.

Types of AI you'll hear named (15 terms)

• GPT: the model family behind ChatGPT.
• Claude: Anthropic's model family.
• Gemini: Google's model family.
• Llama: Meta's open-weight models.
• Mistral: a French model family, often open.
• Open-source model: a model whose weights are public. You can run it yourself.
• Closed model: accessed only through the company's API.
• Multimodal: handles more than text — images, audio, video.
• Vision model: understands images.
• Text-to-image: produces images from prompts (Midjourney, Stable Diffusion, DALL·E).
• Text-to-speech (TTS): turns text into audio.
• Speech-to-text (STT): transcription. Whisper is the well-known one.
• Diffusion model: the technique behind most image generators.
• Transformer: the underlying architecture behind almost every modern LLM.
• Foundation model: a large general-purpose model others build on top of.

Terms about risk, cost and limits (15 terms)

Real work needs these more than the flashy ones.

• Bias: the model reflecting patterns in its training data, sometimes unfair ones.
• Data drift: the world changing so the model's answers get stale.
• Overfitting: a model that memorised its training data instead of learning from it.
• Latency: how long a response takes.
• Throughput: how many requests per second a system handles.
• Rate limit: the ceiling on how often you can call an API.
• Token cost: what you pay per thousand tokens. Adds up fast at scale.
• Context stuffing: cramming too much into the prompt. Wastes money and hurts quality.
• Jailbreak: a prompt that gets a model to break its own rules.
• Prompt injection: malicious text hidden in a document that hijacks the model.
• PII (Personally Identifiable Information): data you should be careful pasting in.
• Data retention: whether the provider keeps your inputs.
• On-device / local model: runs on your computer, not in the cloud.
• Quantisation: shrinking a model so it runs on smaller hardware.
• Compute: the raw processing power a model needs. Expensive.

The last 10 (worth knowing, not urgent)

• AGI (Artificial General Intelligence): hypothetical human-level AI. Not here.
• ASI: hypothetical superhuman AI. Also not here.
• Narrow AI: what all current AI actually is — good at specific things.
• Neural network: the maths structure underneath most models.
• Deep learning: neural networks with many layers.
• Supervised learning: learning from labelled examples.
• Unsupervised learning: finding patterns in unlabelled data.
• Reinforcement learning: learning by trial and reward.
• Synthetic data: training data made by another AI.
• Distillation: training a smaller model to imitate a bigger one.

If you got here, you now recognise more AI vocabulary than most people applying for AI-adjacent work. The next step is picking one thing to build or one platform to try — not learning more terms.

If you're here because you're deciding whether to bother

Fair question. Here's the honest version.

Paid tasks reviewing AI output. DataAnnotation is commonly reported at around $20/hr base, with specialist tasks reported higher (Legit Reviews Lab). Its parent, Surge AI, is reported at around $14-$20/hr for verified general tasks (Thrifty Hustle Hub). Both are gated by an unpaid assessment, which is where most people fall out. Work availability is inconsistent — treat it as variable, not a salary.

Marketplace for domain experts. Mercor is reported from roughly $25/hr up to $200/hr, with the top rates going to existing professional expertise like medicine, law or finance (AI Miracle). If you have no domain yet, this is a month-three target.

Freelance platforms. Upwork's own guidance puts entry-level work at $10-$25/hr, intermediate at $25-$75/hr, and specialised AI/consulting at $75-$150+/hr (Upwork). Fiverr has no gate to list, but early orders are usually priced low to build ranking (Legit Reviews Lab).

Most people earn little or nothing at first. Four hours at $20/hr is about $80 — whether you get four hours in a given week is not something you control. That's the real shape of this work.

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