Glossary
AI, in plain English.
The tech industry loves jargon. Your team doesn't have time for it. Here's the shortlist that actually matters — with the honest version of each term.
- PromptThe instruction you give an AI.
- A prompt is anything you type into an AI tool — a question, a task, a template. Better prompts get better output. That's most of the skill.
- ModelThe specific AI you're talking to.
- Different models (GPT-4, Claude, Gemini, etc.) have different strengths. For most YMCA work, any modern general-purpose model is fine.
- TokenHow AI counts words.
- Roughly ¾ of a word. Long prompts cost more tokens, but for everyday work you don't need to think about them.
- HallucinationWhen AI confidently makes things up.
- AI generates the most plausible-sounding continuation of text — which is not the same as truth. Always verify facts, quotes, statistics, and names.
- Context windowHow much AI can hold in its head at once.
- Everything in the conversation — your instructions and its replies — competes for space. Very long chats forget the beginning.
- System promptThe instructions that shape the AI's role.
- A hidden or up-front instruction like 'You are a safeguarding lead writing for volunteers'. It steers tone and priorities before the user's question.
- RAG (retrieval-augmented generation)Grounding AI in your documents.
- The AI is given specific source material (a policy, transcript, funder brief) and told to answer only from it. Reduces hallucination.
- Chain-of-thoughtAsking AI to think in visible steps.
- Instead of jumping to an answer, the AI shows its working. Better on analysis; worse on creative writing.
- Few-shotShowing AI examples of what 'good' looks like.
- Paste 2–5 examples of the input→output pattern you want, then ask for a new one. The AI mimics the pattern.
- Fine-tuningTraining an AI on your own data.
- Rarely needed for YMCA work. Good prompting and a few examples usually get you the same result faster and cheaper.
- EmbeddingA number-fingerprint of a piece of text.
- Used behind the scenes to find similar documents. You won't touch this directly, but tools like 'search my documents with AI' rely on it.
- AgentAn AI that takes actions, not just gives answers.
- For example: sends an email, updates a spreadsheet. Powerful but risky. Only use for low-stakes, reversible tasks until you've built trust.
- GuardrailsRules that keep AI in bounds.
- Instructions or systems that block AI from producing certain kinds of output (unsafe, off-topic, off-brand).
- AI drafts, humans decideThe rule.
- Everything AI produces is a draft. A named person reads, edits, and signs off before it's used. Non-negotiable for anything that touches young people, safeguarding, funding, or the public.