Reference
S2 AI Glossary
Plain-language definitions for the words that come up at AI Studio. Used in the Summer Kickoff and pinned in the back of every classroom routine.
| LLM | Large Language Model. The kind of AI behind ChatGPT, Claude, and Gemini. Trained on enormous amounts of text; predicts what comes next. |
| Token | A small chunk of text the model thinks in. About three or four characters on average. When a model has a "context window of 200,000 tokens," that is roughly 150,000 words. |
| Context window | How much text the model can read at once. Bigger window = the model can hold a longer document, a longer conversation, more attached files. |
| System prompt | Instructions given to the model up front that shape every response. The "always answer in plain language; be concise; if you're unsure, say so" layer. |
| Hallucination | The model confidently states something that isn't true. Not a bug — a property of how prediction works. Always check facts that matter. |
| Prompt injection | An attack where text inside a document or webpage tells the model to ignore your instructions and do something else instead. Real, growing, especially at Levels 4 and 5. |
| RLHF | Reinforcement Learning from Human Feedback. The training step where humans rate model responses and the model gets nudged toward the patterns humans preferred. |
| Custom GPT | A configured version of ChatGPT tuned for one specific job. System prompt + reference files + sometimes uploaded data. |
| Claude Project | Anthropic's equivalent of a Custom GPT. Same idea: a configured assistant with attached context for a specific use case. |
| Skill | A capability the AI can invoke as part of a larger conversation. A specialized move the model knows how to do when the conversation calls for it. |
| Memory file (CLAUDE.md) | A standing context file the AI reads at the start of every conversation. Your personal or classroom context, persistent. |
| Agent | An AI that can take multi-step action on its own — calling tools, reading files, sending messages, updating systems — while you supervise the results. |
| Training opt-out | The setting that tells the AI provider they may not use your prompts and outputs to train future versions of the model. Required for any S2-confidential use. |
| DPA | Data Processing Agreement. The contract that governs how a vendor handles your data, including FERPA and NY Ed Law 2-d obligations. |
| FERPA | Family Educational Rights and Privacy Act. The federal law that protects the privacy of student education records. |
| NY Ed Law 2-d | New York's state-level student data protection law. Adds requirements beyond FERPA for vendors handling NY student data. |
| Disclosure | Saying — in writing, when expected — that AI was used and how. The S2 default is "disclose when in doubt; especially when the audience could feel misled." |
| Validator | A human in the loop whose job is to check that AI output is correct, appropriate, and aligned with what's intended. Every AI workflow at S2 has a named human validator. |
Glossary · v1 · Last updated June 2026