Overview
This article can be used a a glossary for common terms when speaking about AI. Definitions and terms are subject to change.
- Agent - an AI that can perceive its environment, make decisions, and take actions to achieve specific goals.
- AI washing - the practice of exaggerating or misrepresenting AI functionality.
- Algorithm - a process or set of rules to be followed in calculations or other problem-solving operations.
- Artificial Intelligence (AI) - the ability of a system to perform tasks that typically require human intelligence, such as learning, problem-solving, and decision-making.
- Cloud model / Local model - these describe where a model runs (deployment location), independent of whether it's proprietary or public. A model can be proprietary and cloud-hosted, proprietary and local, public and cloud-hosted, or public and local — the two categories aren't mutually exclusive with proprietary/public.
- Context window - the amount of text (measured in tokens) a model can process or "remember" at once during a conversation.
- Fine-tuning - additional training applied to a pre-trained model to specialize its behavior for a specific task or domain.
- Foundation model - a large model trained on broad data that can be adapted (via fine-tuning or prompting) to many downstream tasks; frontier models are typically the newest, most capable foundation models.
- Generative AI (Gen AI) - a category of AI models that create new content (text, images, audio, video, code) rather than just classifying or predicting from existing data. LLMs are one type of generative AI; image generators (e.g., DALL-E, Midjourney) are another.
- GPT - stands for "Generative Pre-trained Transformer". It is a type of large language model, a neural network trained on a massive amount of text to produce human-like language outputs.
- Grounding - techniques (such as RAG) that tie a model's output to verified, external, or current data sources rather than relying solely on its training data.
- Guardrails - technical or policy-based controls that constrain what an AI system can do or say, used to reduce misuse or unsafe output.
- Hallucination - confident but factually incorrect or fabricated output generated by an AI model.
- Large Language Model (LLM) - AI model trained on massive datasets of text and computer code.
- Local model - an AI model that stores and processes data locally.
- Frontier model - highly advanced, large scale AI model.
- Machine Learning (ML)- enables computers to learn from data without explicit programming, allowing them to make predictions and decisions.
- Model - an AI that has been trained on a set of data to recognize certain patterns or make certain decisions.
- Multimodal - a type of artificial intelligence that can process and understand multiple types of data, such as text, images, audio, and video.
- Natural Language Processing (NLP) - an algorithm that enables computers to understand, interpret, and generate human language.
- Open-weight model - a model whose trained parameters ("weights") are publicly released for download and self-hosting, though the training data or code may remain undisclosed. Distinct from fully open-source models.
- Prompt injection - a security risk where malicious input is crafted to override or manipulate a model's intended instructions or behavior.
- Proprietary model - Model controlled by an organization that restricts access.
- Public model - a model whose weights are openly published for download, modification, and self-hosted deployment (also called "open-weight" or "open-source," depending on license terms). Not to be confused with a model that is merely publicly accessible as a hosted service (e.g., ChatGPT) — that's a proprietary model with public access, not a public model.
- Retrieval Augmented Generation (RAG) - a technique that allows LLMs retrieve and incorporate new information from external sources.
- Shadow AI - unsanctioned or unmanaged use of AI tools within an organization, outside of approved channels or oversight.
- System prompt - background instructions given to a model (usually invisible to the end user) that shape its behavior, tone, or constraints for a session.
- Training data - the dataset used to teach machine learning and artificial intelligence models how to perform specific tasks or make predictions.
- Token - a unit of text (roughly a word or word-fragment) that a model processes; used to measure input/output size and cost.
- Wrapper - a software layer that enhances interactions between users and underlying AI models, including accessibility and usability.
- Zero Data Retention (ZDR) - a contractual or configuration commitment where a vendor does not store or retain user inputs/outputs beyond the immediate processing needed to generate a response.
Through these definitions, we can say that ChatGPT is a proprietary frontier multimodal cloud AI LLM that heavily utilizes NLP, trained on large volumes of text and other data from the public internet and licensed sources. It has RAG functionality called Search, AI agent functionality called Agent Mode, and is the underlying model for numerous wrappers including Microsoft Copilot. As you continue to interact with ChatGPT, it may use your input to improve future models unless you opt out.