Another key term is “inference,” a best guess based on new inputs. In AI, those inputs are the data and training, and an inference is the logical conclusion. It’s what powers the output.

A prompt is what you put into the AI. These can be simple (“proof this email”) or they can be complex and prescriptive (“proof this email for an audience of donors who know little about the news business”). Generally, the more specific the prompt, the better the result is going to be. Think of it as a set of instructions that you give the machine.

Outside of their training, AI has a limit to what it can remember and apply to a set of outputs. This is what’s known as a “context window.” You can increase a context window, but there’s a downside — both in speed (it has to scroll through a longer working memory before replying) and cost.

Retrieval-Augmented Generation, or a “RAG,” improves Large Language Model responses by adding another layer of data or context. For example: A publisher could connect an AI search tool to its archive. When a reporter asks what the city council previously said about a development project, the system retrieves relevant articles and supplies them to the model before it answers. RAGs also greatly reduce hallucinations (see below), because the machine is relying on data sets more specific to your task.

Non-Deterministic outputs are what make AI different from so many other technological processes. If you plug two numbers into two cells on a spreadsheet, and then add them together in a third cell, you will always get the same answer. These are deterministic results, because when the ingredients are the same, the outputs will be the same. By contrast, AI produces non-deterministic results because they can be different every time — even if you’ve given it the same input or prompt. This is because the outputs are probabilistic.

A hallucination refers to an incorrect, ungrounded, or otherwise false output. Given AI outputs are generated by probabilities, some might say that every output is a hallucination. But colloquially, this usually refers to incorrect outputs.

It’s part of the reason why, when people talk about the infrastructure around AI, they often also mention the importance of a “Human-in-the-loop” (or HITL). The high-level concept is that some processes that rely on accuracy (such as journalism) need to have a human who can greenlight or adjust outputs. However, in practice, this can mean a lot of different things: Does the human fully edit the output, or give a high-level judgment on it? The practical application can vary widely by organization.

A “loop” also describes the multi-step cycle through which the model thinks, generates an output, takes in feedback, and adjusts. Sometimes it’s a human that provides the feedback, and other times an AI agent can be instructed to manage this process on its own. (You may have heard of the “AI Doom Loop,” in which you give feedback to the LLM, and the results progressively get worse instead of improving.)

Lastly, there’s Agentic AI — a term that is often thrown around to describe a myriad of use cases but actually means something quite specific. It’s a type of AI (like Generative AI) that can work autonomously to act on a human’s behalf and make decisions. Because agents can take actions — not merely suggest them — they need clear limits, permissions, checkpoints, and human review. An error in a draft is one thing; an agent publishing, emailing, or modifying records based on that error is another.