AI Glossary: Vocabulary You Actually Need

ai-glossary-vocabulary-you-actually-need

[[p]]It&rsquo;s&nbsp;called the &lsquo;AI Boom&rsquo; for a reason &ndash; [[a href="https://digitalmarketinginstitute.com/blog/ai-in-digital-marketing-the-ultimate-guide" target="_blank"]]artificial intelligence[[/a]]&nbsp;(AI)&nbsp;has arrived&nbsp;into&nbsp;many of our lives like an explosion. As with many&nbsp;new trends, there is a high degree of risk that it passed many people by, and now&nbsp;they&rsquo;re&nbsp;very confused by what everyone is talking about. And now it seems too late to ask.&nbsp;&nbsp;[[/p]]

[[p]]The modern workplace is awash in jargon of all kinds, which we accept without questioning, or try to learn from context.&nbsp;That&rsquo;s&nbsp;why, to make your life easier,&nbsp;we&rsquo;ve&nbsp;prepared this quick glossary of AI terms. The terms contained in this glossary are some of the ones&nbsp;you&rsquo;re&nbsp;most likely to&nbsp;encounter&nbsp;at work.&nbsp;&nbsp;[[/p]]

[[p]]The vocabulary of AI is still evolving, but these terms cover both the core concepts of AI, to more technical things a non-expert is likely to&nbsp;encounter. Definitions are given in plain language, with practical examples to help you to grasp the technology.&nbsp;&nbsp;[[/p]]

[[p]]Terms in this glossary are divided by:&nbsp;[[/p]]

[[ul]]
	[[li]][[a href="#heading_85872"]]Core concepts of AI&nbsp;[[/a]][[/li]]
	[[li]][[a href="#heading_85888"]]How AI works&nbsp;[[/a]][[/li]]
	[[li]][[a href="#heading_85906"]]Practical use of AI&nbsp;[[/a]][[/li]]
	[[li]][[a href="#heading_85926"]]AI risks and ethics&nbsp;[[/a]][[/li]]
[[/ul]]

[[p]]These are the&nbsp;very basic, foundational ideas that everyone who will be interacting with AI needs to understand.&nbsp;[[/p]]

[[p]]A broad term for computer systems that perform tasks which would typically require human intelligence, such as recognizing patterns, understanding language, or making decisions.&nbsp;[[/p]]

[[p]]Example:&nbsp;When your email account automatically sorts your inbox into Primary and Promotions, that&rsquo;s AI.&nbsp;[[/p]]

[[p]]A form of AI in which computer systems learn from data and adapt, rather than following explicit instructions. The more data it processes, the better it performs at its task.&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]A rule-based system for detecting credit card fraud might block transactions over a certain amount or flag purchases in another country. A machine learning system learns from millions of past transactions to understand what &ldquo;normal&rdquo; looks like for&nbsp;each individual, flagging unusual&nbsp;behaviour, adapting over time.&nbsp;[[/p]]

[[p]]A type of AI system trained on vast quantities of text, which can generate, summarize, translate, and reason about language. LLMs are behind most modern AI writing and research&nbsp;tools and&nbsp;is some of the most&nbsp;widely used&nbsp;AI technology.&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]ChatGPT, Claude, and Gemini are all LLMs. When you ask one of these to&nbsp;write&nbsp;an email for you, they are powered by LLMs.&nbsp;[[/p]]

[[p]]AI systems that produce&nbsp;new content&nbsp;like text, images, audio, or code in response to a request. Also known as&nbsp;GenAl, it is a subset of AI that goes beyond classifying or predicting to&nbsp;actually creating.&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]Asking an AI to write a job description, generate a product image, or compose a summary of a meeting recording are all uses of generative AI.&nbsp;[[/p]]

[[p]]A model is the &lsquo;engine&rsquo; behind an AI tool or application, or the trained system that produces AI outputs. They are trained on large datasets and learn to recognize patterns, make predictions, and generate new outputs.&nbsp;&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]When you switch between AI tools,&nbsp;you&#39;re&nbsp;often switching between different underlying models. If you ask ChatGPT and Claude the same question, they might give vastly different answers and in vastly&nbsp;different ways&nbsp;because they use different models.&nbsp;&nbsp;[[/p]]

[[p]]A large AI model trained on broad, diverse data that can be adapted for many different tasks. Most commercial AI tools are built on top of, or derived from, a foundation model.&nbsp;[[/p]]

[[p]][[strong]]Example:[[/strong]]&nbsp;When a company builds a specialized AI product, like a legal research tool or a healthcare documentation assistant, they are typically building on top of a foundation model rather than training one from scratch. GPT-4, Claude, and Gemini are all foundation models.&nbsp;[[/p]]

[[p]]An AI system that can process and generate multiple types of content like text, images, audio, and video rather than just text alone.&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]Image-to-text tools, where you describe the&nbsp;image&nbsp;you want to&nbsp;create to&nbsp;an AI and then it generates it, is an example of multimodal AI. Similarly, an AI tool that can analyse a photograph of a document and extract the text from it is multimodal.&nbsp;&nbsp;[[/p]]

[[p]]Now that you understand the most basic elements of AI, these are terms that explain how AI works. While you might not&nbsp;encounter&nbsp;them on a regular basis, and they may not be important to your role, having a grasp of them will make you better at using AI.&nbsp;&nbsp;[[/p]]

[[p]]The process by which an AI model learns patterns from a large dataset. Training happens before the model is deployed and shapes everything it knows and how it behaves.&nbsp;[[/p]]

[[p]][[strong]]Example:[[/strong]]&nbsp;An LLM trained primarily on English-language text will perform better in English than in less-represented languages because&nbsp;that&#39;s&nbsp;what the training data reflected.&nbsp;[[/p]]

[[p]]The body of text, images, or other content used to train an AI model. The composition of training data&nbsp;is responsible for&nbsp;the model&#39;s knowledge and potential biases.&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]If an AI HR tool for candidate screening is trained on historical hiring data that reflect gender bias, the model may replicate bias in the candidates it recommends.&nbsp;&nbsp;[[/p]]

[[p]]A token is the name for the basic unit that language models process. It is usually a full word or a part of a word. Models read input and generate output token by token, not word by word or sentence by sentence like a human.&nbsp;[[/p]]

[[p]][[strong]]Example:[[/strong]]&nbsp;The limit on how long a document you can paste into an AI tool is usually expressed in tokens, not words.1,000 tokens&nbsp;is&nbsp;usually equivalent to&nbsp;roughly 750&nbsp;words.&nbsp;&nbsp;[[/p]]

[[p]]This is the amount of text (your messages, documents you shared, and an AI&rsquo;s own responses) that a model can hold in attention at any one time.&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]If an AI&nbsp;seems to forget&nbsp;something you mentioned earlier in&nbsp;a long conversation, it may have dropped out of the context window.&nbsp;In essence it&nbsp;can &lsquo;forget&rsquo; things the longer your conversation goes on.&nbsp;[[/p]]

[[p]]A setting that controls how varied or predictable an AI&#39;s outputs are. Lower temperature produces more consistent, conservative responses&nbsp;whereas&nbsp;higher temperature produces more varied, creative ones.&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]A customer-service chatbot set to low temperature will give consistent, predictable answers. A creative writing tool set higher will produce more varied, surprising suggestions, but also a higher chance of something odd.&nbsp;[[/p]]

[[p]]RAG is a technique where an AI model retrieves relevant information from external sources, such as databases or documents, before generating a response. This allows the model to go beyond its original training data and knowledge cutoff, allowing for more up-to-date answers, or proprietary information, and often enabling more reliable responses.&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]A legal research tool that searches a firm&#39;s own case database, rather than relying solely on what it was trained on before drafting a research summary is using RAG.&nbsp;&nbsp;[[/p]]

[[p]]Inference is how an AI model uses its knowledge to respond to you. AI cannot think, so it makes predictions based on patterns in its large set of training data. Every time you&nbsp;submit&nbsp;a prompt and receive a response, that&#39;s inference.&nbsp;[[/p]]

[[p]][[strong]]Example:[[/strong]]&nbsp;A chef&nbsp;doesn&#39;t&nbsp;relearn how to make pasta every time&nbsp;it&#39;s&nbsp;ordered. Thus, an AI model&nbsp;doesn&#39;t&nbsp;retrain itself each time you&nbsp;submit&nbsp;a&nbsp;prompt but&nbsp;uses what it already knows to produce your response.&nbsp;[[/p]]

[[p]]A set of instructions given to an AI tool before the user interacts with it, typically written by the organization that has deployed the tool. The system prompt shapes the AI&#39;s behavior, tone, and boundaries within that product.&nbsp;[[/p]]

[[p]][[strong]]Example:&nbsp;[[/strong]]When a company&#39;s customer service chatbot refuses to discuss competitor products or always responds in a specific tone, that behavior is usually defined in a system prompt, rather than the underlying model.&nbsp;&nbsp;[[/p]]

[[p]]Now we&rsquo;ve established what AI is, and how it works. This section deals with how you interact with AI, and what work it can do for you. These terms appear in the day-to-day use of AI tools across professional contexts.[[/p]]

[[p]]A prompt is an instruction, question, or input you give to an AI system. The quality and specificity of the response an AI gives you significantly depends on the quality of your prompt.[[/p]]

[[p]][[strong]]Example:[[/strong]][[em]] "Summarize this report in three bullet points for a non-specialist audience"[[/em]] is a more effective prompt than [[em]]"summarize this," [[/em]]as it gives the AI the context it needs to produce a useful output.[[/p]]

[[p]]The practice of crafting prompts carefully to get better outputs from AI systems. It is a valued skill with established techniques for different tasks and contexts.[[/p]]

[[p]][[strong]]Example:[[/strong]] Specifying the audience, format, tone, and length in a prompt is basic prompt engineering. More advanced techniques include role prompting (telling the AI to act as a specific expert), chain-of-thought prompting (asking it to reason step by step), and adversarial prompting (asking it to argue against its own conclusions).[[/p]]

[[p]]The output is the thing an AI creates; the response or content in reply to your prompt. Outputs can be text, images, code, audio, or other formats depending on the model.[[/p]]

[[p]][[strong]]Example:[[/strong]] The draft email an AI writes in response to your instructions is its output.[[/p]]

[[p]]Fine-tuning is teaching an existing AI model to be better at a specific task by training it on relevant examples. The model keeps all its general knowledge and gains specialist knowledge on top.[[/p]]

[[p]][[strong]]Example:[[/strong]] A law firm might fine-tune a model on thousands of legal documents, so it produces more accurate, appropriately worded legal analysis. The underlying model hasn&#39;t changed; it&#39;s just become much better at one specific job.[[/p]]

[[p]]This is the point in time after which an AI model has no knowledge. It is usually because a model was trained on data up to that point, but nothing after. Everything that happened before that date, the model may know, but everything after it, the model is entirely unaware of. This is why AI tools can give out of date answers.[[/p]]

[[p]][[strong]]Example:[[/strong]] Asking an AI about legislation passed after its knowledge cutoff will produce unreliable results. A model with a 2024 cutoff has no knowledge of a 2025 court ruling or a 2025 budget announcement.[[/p]]

[[p]]This is a term likely familiar to anyone using Microsoft products, but it has a broader meaning within AI. It describes AI tools that work alongside a professional within existing software &ndash; any AI assistant embedded in a workplace tool. It might make suggestions or complete tasks without taking them over entirely.[[/p]]

[[p]][[strong]]Example:[[/strong]] Microsoft Copilot in Word can draft documents from a brief, summarize long reports, and suggest edits as you write.[[/p]]

[[p]]The process of anchoring AI outputs to specific, verified sources of information rather than relying solely on the model&#39;s training data. Grounding reduces hallucination risk and is especially important in regulated professional contexts where accuracy is paramount. Grounded outputs are easier to verify because they come with source references.[[/p]]

[[p]][[strong]]Example:[[/strong]] A tax research tool that is grounded in current UK guidance retrieves and cites specific HMRC pages before generating its response, rather than producing plausible-sounding tax rules from training data alone.[[/p]]

[[p]]An AI system that can take actions, use tools, and complete multi-step tasks autonomously, beyond just responding to a single prompt to pursue a goal over a sequence of steps.[[/p]]

[[p]][[strong]]Example:[[/strong]] An AI agent prompted to [[em]]"research the top five competitors and produce a briefing document" [[/em]]will independently conduct searches, retrieve and read relevant pages, synthesize the information, and produce the document without needing step-by-step human instruction for each stage.[[/p]]

[[p]]A DPA is a legal contract between an organization and an AI tool provider (or other data processor) that governs how personal and sensitive data is handled, stored, and used. A DPA is required under the General Data Protection Regulation (GDPR) whenever personal data is shared with a third-party processor.[[/p]]

[[p]][[strong]]Example:[[/strong]] Before using an AI tool to process employee or client data, your organization should have a DPA in place with the provider.[[/p]]

[[p]]Ethical concerns are paramount when using AI. These terms deal with risks and ethical issues that are of concern in the workplace.[[/p]]

[[p]]When an AI model generates content that is plausible sounding but factually incorrect, fabricated, or unsupported it is a hallucination. This is a structural consequence of how language models work, and not a bug that can eventually be patched. It is an inherent, persistent feature of probabilistic text generation.[[/p]]

[[p]][[strong]]Example:[[/strong]] An AI asked to provide case citations may generate references that sound entirely real, with a plausible case name, court, date, and holding, but do not exist.[[/p]]

[[p]]Because AI is built by humans, and trained on human-generated data, it is susceptible to systematic human and statistical skews. These typically reflect imbalances or historical inequalities affecting language, recommendations, assessments, and decisions in ways that disadvantage certain groups.[[/p]]

[[p]][[strong]]Example: [[/strong]]An AI tool trained on historical medical research may perform less accurately for certain demographic groups that were underrepresented in that research.[[/p]]

[[p]]The ability of AIs to describe in simple terms a human can understand how an AI system arrived at a specific output or decision. It is a legal and ethical requirement for AI systems used in high-stakes decisions in many jurisdictions, including under GDPR Article 22.[[/p]]

[[p]][[strong]]Example:[[/strong]] An AI system operating in a bank that scores loan applications should be able to explain how it reached a particular conclusion, not just produce a score. Without explainability, individuals cannot exercise their legal right to challenge automated decisions that significantly affect them.[[/p]]

[[p]]These are the policies, processes, and oversight structures an organization puts in place to ensure AI is used safely, ethically, and in compliance with applicable law and professional standards.[[/p]]

[[p]][[strong]]Example:[[/strong]] An AI governance framework might specify which tools can be used for client work, what data handling protocols apply, who reviews AI-assisted outputs before they are used professionally, and how the policy is updated as the technology and regulations evolve.[[/p]]

[[p]]This is the principle in which a human is involved in reviewing, approving, or overriding AI decisions at key points in a process, rather than allowing the AI to complete tasks entirely autonomously. Human-in-the-loop is both an ethical safeguard and, in regulated contexts, a legal requirement.[[/p]]

[[p]][[strong]]Example:[[/strong]] An AI agent that drafts a client communication is not "human-in-the-loop" if the communication is sent automatically. It is human-in-the-loop if the draft is presented to a practitioner for review and approval before sending.[[/p]]

[[p]]Guardrail is the name for the constraints built into an AI system, either technically or through policy, that define what an AI can and cannot do. Guardrails are what prevent an AI agent from taking actions that fall outside acceptable boundaries.[[/p]]

[[p]][[strong]]Example[[/strong]]: A HR AI agent might have guardrails that prevent it from accessing individual salary records without explicit authorization. These constraints may be built into the system prompt, enforced programmatically, or both.[[/p]]

[[p]]This list is by no means comprehensive. AI is an ever-evolving landscape, and there will be new developments and new terms. However, the list of terms contained here should help you to navigate the world of AI with confidence, as you begin to experiment with new tools and systems in your work life.[[/p]]

[[p]]Artificial intelligence is a game changer for marketers as it helps personalize, optimize, and automate many key marketing activities. Our short interactive AI course (with practical AI tool walkthroughs) will help you understand the fundamentals of AI and use key tools for research, content development, data-driven decision-making, and creative design. Get started today![[/p]]

AI Glossary: Vocabulary You Actually Need

Related Free Video Lessons