AI Terminology, Decoded

Artificial Intelligence (AI)

A computer program trained to recognize patterns in huge amounts of data, so it can produce useful output — text, images, decisions — without being explicitly programmed for each case.

Machine Learning (ML)

The technique behind most modern AI: instead of writing rules by hand, you show the program thousands or millions of examples and let it learn the pattern itself.

Neural Network

The structure most AI models are built from — layers of simple math functions loosely inspired by neurons, stacked together so the whole thing can learn complex patterns.

Large Language Model (LLM)

A neural network trained on massive amounts of text to predict the next word in a sequence. ChatGPT, Claude, and Gemini are all products built around an LLM.

Generative AI

Any AI that creates new content — text, images, audio, code — rather than just classifying or scoring existing content.

Prompt

The text you send the AI. Everything you type — your question, instructions, and any pasted content — is the prompt.

Prompt Engineering

Deliberately wording your prompt to get a better, more specific, or more reliable answer — the AI equivalent of asking a good question.

System Prompt

Hidden instructions set by the app (not you) that shape how the AI behaves before your first message even arrives — its personality, rules, and boundaries.

Context Window

The maximum amount of text an AI can "see" at once — your current message plus recent conversation history and any pasted documents. Go past it and the oldest content gets dropped.

Token

The unit AI models actually process — roughly a chunk of a word. "Unbelievable" might be three tokens. Pricing, limits, and context windows are usually measured in tokens.

Training Data

The enormous collection of text, images, or code an AI model learned from before it was ever released to the public.

Parameters

The internal numeric values a model adjusts during training. "A 70-billion-parameter model" is a rough proxy for its size and, loosely, its capability.

Fine-Tuning

Taking an already-trained model and training it further on a narrower, specific dataset — e.g. turning a general model into a customer-support specialist.

Inference

The moment the trained model is actually used to generate a response to your prompt — as opposed to training, which happens beforehand.

Hallucination

When an AI states something false with total confidence, because it's predicting plausible-sounding text, not retrieving verified facts.

Multimodal

A model that can handle more than one type of input or output — text and images together, for example, instead of text only.

RAG (Retrieval-Augmented Generation)

A technique where the AI first searches a set of documents or the web for relevant facts, then writes its answer using what it found — instead of relying only on what it memorized during training.

Agentic / AI Agent

An AI that doesn't just reply — it takes multi-step actions on your behalf: browsing, clicking, running code, or using tools, then deciding what to do next based on the result.

Reasoning Model

A model built to work through a problem in explicit steps before answering, rather than producing the first plausible response — generally slower, but more reliable on math, logic, and multi-step tasks.

Embedding

A way of converting text into a list of numbers that captures its meaning, so a computer can measure how similar two pieces of text are. The backbone of search and RAG.

Foundation Model

A large, general-purpose model (like GPT or Claude) trained once on broad data, then adapted or fine-tuned into many different specific products.

Open-Weight / Open-Source Model

A model whose internal parameters are published publicly, so anyone can download and run it on their own hardware — Llama and Mistral are examples.

Closed / Proprietary Model

A model only accessible through the maker's app or API — you never get the underlying weights, only the ability to send it prompts and get responses back.

API (Application Programming Interface)

The way developers plug an AI model directly into their own apps and products, instead of using the company's own chat interface.

Alignment

The ongoing effort to make an AI's behavior match what its developers (and users) actually intend — helpful, honest, and non-harmful — rather than just technically capable.

Guardrails

Rules and filters built into an AI product to stop it from producing certain content or taking certain actions, regardless of how it's asked.

Jailbreak

A prompt specifically crafted to trick an AI into ignoring its guardrails — a reminder that safety filters are software, not certainty.

PII (Personally Identifiable Information)

Any detail that can be tied back to a specific real person — name, email, phone number, SSN, home address, and similar identifiers.

→ This is exactly what specVIO scans for and strips out of your prompts before they leave your device.

Data Retention Policy

A provider's stated rules for how long they keep your conversations and whether those conversations are used to train future models.

→ Retention terms vary a lot by provider and by pricing tier — see the privacy notes in the tool breakdown above.

Data Leakage

Sensitive information ending up somewhere it shouldn't — like an API key or a colleague's email address getting pasted into a chatbot prompt without you noticing.

→ The core problem specVIO exists to prevent, one prompt at a time.