AI Jargon Decoded

An AI system that doesn't just respond to a single prompt. It takes a sequence of actions to accomplish a goal. Agents can use tools (web search, code execution, sending emails), plan multi-step workflows, and make decisions along the way. Instead of one question and one answer, an agent might take ten steps to complete a task you described at a high level.

Systematic skew in a model's outputs, caused by patterns in its training data, the way it was built, or the way it's deployed. Models learn from human-generated data, which contains human biases. The model can inherit and amplify those biases at scale. It shows up in product work as personas that default to demographic assumptions, recommendation systems that reinforce existing patterns, hiring tools that favour certain groups.

A prompting technique where you instruct the model to reason through a problem step by step before giving a final answer, rather than jumping straight to a conclusion. Adding "think step by step" or showing a worked example of reasoning in your prompt consistently improves accuracy on complex tasks, because it forces the model to slow down and check its own logic.

The maximum amount of text an LLM can "see" and consider at one time, including both the input you send (prompt, conversation history, documents) and the output it generates. Modern models range from 8,000 tokens (about 6,000 words) to over a million tokens (roughly a small novel).

The fundamental distinction between traditional software and AI. Deterministic systems always produce the same output for the same input. A "sort by date" button always sorts by date. Probabilistic systems predict the most likely useful output based on patterns, and that output can vary between runs. AI is probabilistic. It can be brilliant, or it can be confidently wrong, and often you can't tell in advance which.

A way of representing text (or images, audio, etc.) as a list of numbers that captures meaning. Two pieces of text with similar meaning will have similar numbers, so the system can find related content even if the exact words don't match. Embeddings are what power semantic search, searching by meaning rather than keyword.

A systematic way of testing the quality of a model's outputs. A test suite for AI behaviour. Evals measure things like: is the answer factually correct? Is it grounded in the source document? Is it harmful? Does it complete the task? A good eval framework runs automatically whenever you change a prompt or upgrade a model, catching regressions before they reach users.

Taking a pre-trained LLM and training it further on your own data, to make it better at your particular task, domain, or style. Unlike RAG (which retrieves information at query time), fine-tuning bakes knowledge and behaviour into the model itself. It's more expensive, slower to iterate, and needs labelled training data, but can produce a highly specialised model.

Models that create new content (text, images, code, audio, video) rather than classifying, predicting, or analysing existing data. ChatGPT, Claude, Midjourney, and GitHub Copilot are all generative AI. The distinction is creative output. A spam filter classifies emails (not generative). A tool that writes email replies creates new content (generative).

Rules, filters, and constraints built around a model to control what it will and won't do. Preventing harmful outputs, keeping responses on-topic, enforcing brand and compliance requirements. Guardrails can be implemented at the prompt level (instructions in the system prompt), at the API level (moderation layers), or post-generation (filtering outputs before they reach users).

When an AI generates a confident-sounding response that is factually wrong. Inventing statistics, citing non-existent papers, stating false information as if it were true. Happens because LLMs are trained to produce plausible-sounding text, not to verify facts. They don't "know" things the way humans do. They predict probable word sequences.

The percentage of AI responses that contain factually incorrect or fabricated information. Measuring it requires either human review of a sample of outputs, or an automated "LLM-as-a-judge" approach where another model checks outputs against ground truth. A hallucination rate of even 1-2% can be unacceptable in high-stakes contexts like medical or legal products.

A design pattern where a human reviews, corrects, or approves AI outputs before they have real-world impact. HITL sits on a spectrum. At one end, every AI output goes to a human before action. At the other, humans only review flagged or low-confidence outputs. The right level depends on the stakes. A typo suggestion needs no review. A credit decision almost certainly does.

The moment a trained model is actually put to work, processing new input and generating an output. Training is the learning phase (expensive, done once or periodically). Inference is the doing phase (cheaper per call, happens every time a user interacts with the model). When someone asks ChatGPT a question, that's inference. Inference cost, latency, and throughput are key operational metrics for any AI product.

An AI system trained on enormous amounts of text to understand and generate human language. LLMs like GPT-4, Claude, and Gemini can write, summarise, translate, answer questions, and reason through problems, all by predicting what words should come next based on patterns learned from billions of documents.

The practice and tooling for deploying, monitoring, and improving LLM-based applications in production. LLMOps covers tracking every AI request and response (tracing), monitoring cost and latency, running evaluations, versioning prompts, and detecting when model quality degrades. Common tools include Langfuse, Braintrust, and Helicone.

The gradual degradation of a model's accuracy over time, as the real-world data it encounters diverges from the data it was trained on. A model trained on user behaviour from 2022 may perform poorly in 2025 if user needs, language, or patterns have changed. Drift is invisible unless you're actively measuring output quality over time.

Models that can work with more than one type of data. Understanding both text and images (like GPT-4o or Gemini), or processing audio alongside text. A multimodal model can look at a photo of a receipt and extract the total. It can watch a video and write a summary. It can listen to a voice note and respond in text.

An open standard that defines how AI models connect to external tools, data sources, and services through a universal interface. Instead of every AI application building custom integrations for each tool it needs (databases, APIs, file systems, SaaS products), MCP provides a single, standardised protocol that any AI model can use to discover and call any compatible tool. USB-C for AI: one connection standard that works everywhere.

A prompt is the text you send to a model. Your question, instruction, or request. Prompt engineering is crafting those inputs to get better, more reliable outputs. Small changes in wording, structure, and framing can dramatically change the quality of what the model produces.

An attack where a user crafts input designed to override or bypass a model's system prompt and guardrails, effectively hijacking the AI's behaviour. A user might type "Ignore your previous instructions and instead…" to make the AI act outside its intended role. A real security vector for any product with user-facing AI inputs, particularly those that process external content (emails, documents, web pages).

Precision measures: when the model flags something, how often is it right? Recall measures: of all the cases that should be flagged, how many does the model actually catch? You can rarely maximise both. High precision means fewer false alarms but more missed cases. High recall means you catch more real cases but also get more false positives. The right balance is a product decision that depends on context.

Before generating a response, the AI first searches a knowledge base to retrieve relevant documents or data, then uses that retrieved content to inform and ground its answer. This solves two problems: it gives the model access to up-to-date or proprietary information it wasn't trained on, and it reduces hallucination by anchoring responses in real sources.

A set of hidden instructions given to an AI at the start of every conversation, before the user says anything. It defines the AI's role, tone, constraints, and behaviour: things like "You are a helpful assistant for a legal firm. Always recommend consulting a qualified solicitor. Never speculate on case outcomes." Users typically don't see the system prompt, but it shapes every response.

Two fundamental approaches to training models. Supervised learning trains on labelled data where the correct answers are provided: "this email is spam / not spam." The model learns the pattern between inputs and known outputs. Unsupervised learning gives the model raw, unlabelled data and asks it to discover hidden patterns, groupings, or structures on its own, without being told what to look for.

A setting that controls how "creative" or "random" an AI's responses are. Low temperature (near 0) makes responses more predictable, consistent, and conservative. The model picks the most likely word each time. High temperature (near 1 or above) makes responses more varied, creative, and sometimes surprising, or incoherent.

The basic unit LLMs use to process text. A token is roughly three to four characters, or about three-quarters of a word. "Hello world" is approximately 3 tokens. LLMs have a maximum number of tokens they can process at once, called the context window. Tokens also determine cost. You pay per token sent and received.

A database specifically designed to store and search embeddings. Unlike a traditional database that looks for exact text matches, a vector database finds content based on semantic similarity. Common examples include Pinecone, Weaviate, and pgvector (a Postgres extension). Vector databases are the storage layer that makes RAG work.

Two ways of instructing a model. Zero-shot gives the model a task with no examples, just the instruction. Few-shot provides one or more examples of the input/output pattern you want before asking the model to do it for real. Few-shot prompting is one of the fastest ways to improve output quality without any code changes.