Artificial Intelligence: Glossary

Source: Caleb Sponheim; Nielsen Norman Group glossary · 2026-08-21

Agent: A system that can use a model and tools to carry out multiple steps toward a goal, evaluate its progress, and decide what to do next. The product around it controls its tools, permissions, stopping rules, and opportunities for human review.

Agentic AI: A description of how much latitude a system has between prompts. An agentic system can choose several steps and use tools toward a goal without being told each move. Autonomy is a matter of degree, so calling a product an agent says little by itself about how much autonomy it actually has.

AI Slop: Low-quality, mass-produced AI-generated content: text, images, or code that is plausible on the surface but careless, generic, or unverified. Guarding against it with quality bars and human review is a design and governance concern.

AI Washing: Making false, misleading, or exaggerated claims about whether a product uses AI, how it uses it, or what it can do. It can erode user trust when the capability does not match the claim.

Alignment: The effort to make AI systems behave in accordance with human values, intentions, and ethical guidelines. A central design challenge is making systems that do what users want while avoiding harmful or unintended outcomes.

Anthropomorphism: Attributing human qualities such as feeling, intent, or humanlike understanding to an AI system because of the way it communicates. It can make a product feel approachable, and it can just as easily lead users to trust the system further than its reliability supports. First-person language, apologies, and other humanlike cues may strengthen the effect, often called the “ELIZA effect” after a 1966 chatbot.

Automation Bias: The tendency to favor an automated system’s output and to scrutinize it less than you would a person’s, even when it is wrong. Automation bias is one form of overreliance; overreliance is broader and can arise without automation-specific bias. Introducing friction at high-stakes moments and providing easy ways for users to verify or override a result help prevent automation bias.

Chain-of-Thought Prompting: A prompting technique that uses an instruction or worked examples to prompt a sequence of intermediate reasoning steps before a final answer. It can improve performance on many multistep tasks, but the generated steps are not guaranteed to faithfully explain how the model arrived at its answer.

Context: Everything supplied or made available to a model for a single request: the prompt, the system prompt, selected conversation history, retrieved documents, stored memory, and the results of any tools it has called. In most current assistant products, the surrounding application or service stores conversational state and supplies or references it when relevant. Context is what is available for one request; memory is retained information that may later become part of that context. Choosing deliberately what goes into the context is sometimes called context engineering.

Explainability: How well a system can give a person a meaningful and accurate account of how or why it produced a result. A generated explanation may not accurately reflect the system’s actual process.

Generative AI (GenAI): AI systems that generate new content (text, images, code, audio) from patterns learned during training, rather than only retrieving or classifying what already exists.

Generative UI (GenUI): Interfaces whose content, layout, or controls are generated or selected at run time by an AI rather than being entirely fixed in advance. Instead of navigating to a fixed screen, the user describes a need and the system composes a response, often from approved components and rules. GenUI shifts the designer’s work from specifying screens to specifying the components and rules a system may assemble them from.

Grounding: Connecting an AI’s output to selected source data or documents, so the system answers from that material rather than from training data alone. Grounding can improve factual accuracy, but it does not guarantee that the source is correct or that the model uses it correctly. Showing users those sources is a separate decision: a grounded answer is not automatically a transparent one.

Guardrails: Checks and constraints built into an AI system to reduce harmful, off-topic, or unauthorized output. Guardrails lower risk, but they can also fail or be circumvented.

Hallucination: Plausible but factually incorrect or nonsensical AI output. Language models can hallucinate because they generate statistically likely continuations rather than verify each claim; errors in training or retrieved data can also produce false outputs. Grounding and RAG help reduce them.

Human-in-the-Loop: A design pattern that puts a person in position to review, approve, correct, or monitor an AI’s output. It is one common control for high-stakes automation, and it counts as oversight only when the reviewer has the information, time, expertise, and authority to intervene. A confirmation step users click through without reading adds delay instead.

Knowledge Cutoff: A provider-reported date indicating the approximate recency of a model’s training data. It is not a hard boundary: coverage before the date can be uneven, and some later information may appear. More-recent information generally has to be supplied in the prompt or retrieved when the model runs.

Latency: The delay between when a user submits an input and when they receive a response. Lower latency makes an experience feel more responsive. Latency depends on the model and hardware, network and queueing delays, system load, input and output length, and whether it is measured to the first output or the completed response.

Memory: Information an AI product carries forward in subsequent interactions with a user. Conversation, or short-term, memory retains recent interaction within a session; long-term memory consists of information stored across sessions and retrieved when relevant. In most current assistants, the surrounding product or service stores this information and supplies it when relevant.

Multishot Prompting (Few-Shot Prompting): Prompts that provide several worked examples, usually input-and-output pairs, to guide a model in producing outputs with a particular structure. Well-chosen examples improve performance, but more is not reliably better: irrelevant, unbalanced, or badly ordered examples can make results worse. These are contrasted with one-shot prompting (a prompt with exactly one example) or zero-shot prompts.

Overreliance: A pattern of use in which someone gives an AI more authority than its reliability or the stakes justify. It can follow from automation bias, from weak support for verification, from time pressure, or from simple misplaced trust.

Prompt: The input (usually text) supplied to an AI system to get a response. It usually comes from the user, but application-level instructions such as a system prompt are prompts, too.

Prompt Engineering: The practice of designing and refining prompts to get more reliable, accurate, or useful output from a model. Common techniques include supplying relevant context and examples, stating constraints, and specifying the output format. Asking a model to review its own result is unreliable without external feedback and can sometimes make reasoning performance worse.

Prompt Injection: An attack in which instructions to an LLM are inserted into content that the system reads (e.g., a user prompt, a web page, a document). This type of attack attempts to make an AI system break its own rules, leak information, or misuse its tools. It is a leading security concern for any system that uses tools or reads outside data.

Sycophancy: Instances in which an AI model adapts responses to align with the user’s views, even when doing so makes the answer less accurate or useful. This behavior is generally undesirable.

System Prompt: Background instructions that shape how an AI model behaves and responds. A system prompt might define the model’s role, tone, or constraints. Users see the model’s responses but typically not the system prompt directing them.

Tool Use (Function Calling): The ability of an AI system to act outside the conversation. When given tools, a model can request an action such as retrieving data, running a calculation, or sending a message; the surrounding application decides whether to carry the request out, then hands back the result.

Vibe Coding: A method of building software in which the user describes what they want in natural language and lets an AI generate and iterate on the code, often without reading closely everything it produces. It shortens the path from idea to working prototype and lets people build software with less direct coding.

Zero-Shot Prompting: Using a prompt that does not contain examples of the desired answer. The prompt can still carry instructions, background, and source material; what is missing is a worked example. For instance, “Summarize this article” with no sample summary attached.