AI in Plain English: What the Words Mean—and Why You Should Care.

You do not need to build an AI system to ask sensible questions about it. These words help you understand what a tool does, where it can fail, and what remains yours to decide. The examples are illustrations, not reports of particular incidents.

The basic building blocks

Artificial intelligence or AI

The broad field of computing concerned with tasks such as recognizing speech, making predictions, and generating answers. A phone transcribing your voice uses AI. The label alone tells you little about its reliability.

Machine learning or ML

A way to build systems that learn patterns from data. A spam filter can learn from examples of unwanted email. The examples influence what it catches and what it misses.

Neural network

A mathematical model made of connected processing units with adjustable numerical settings. It might recognize handwriting. The brain-inspired name does not establish human understanding or consciousness.

Deep learning

Machine learning using neural networks with many layers. It helps systems recognize complicated patterns, such as objects in photographs. More layers do not guarantee better judgment.

Transformer

A neural-network design that uses a mechanism called attention to weigh relationships within information. It can connect words across a sentence. Attention here is a calculation, not conscious concentration.

Generative AI

AI that produces content, including text, pictures, audio, or code. It can draft a customer letter. Newly generated content still needs checking for accuracy and suitability.

Models and the products we use

AI model

The trained mathematical system used to produce predictions or other outputs. Think of it as one component inside a service. The surrounding software, information, tools, and permissions also shape the result.

Large language model or LLM

A large neural-network model trained on extensive language data. It can summarize a report or draft an explanation. Fluent language can include mistaken claims.

GPT

Short for generative pre-trained transformer. The name describes a type of model and is used for OpenAI model families. It identifies technology, not a guarantee that an answer is correct.

ChatGPT

OpenAI’s conversational product, which brings AI models and tools together for users. You interact with the product; a model supplies part of its capability. Product features and available tools can change.

Training

The process of adjusting a model using data and feedback. Learning from many example messages can improve drafting. Training does not mean every fact is stored accurately or kept current.

Inference

Using a trained model to produce an output from an input. Asking for a summary triggers inference. A new answer does not necessarily mean the underlying model has been retrained.

What happens during a conversation

Prompt

The request or instruction given to an AI system. “Explain this bill in everyday language” is a prompt. Clear instructions help, but cannot make missing information appear.

Token

A unit a model processes, such as a word, part of a word, or punctuation. Tokens help measure input and output size. They are not interchangeable with a product’s credits or message allowance.

Context window

The amount of information a model can work with at once, measured in tokens. Think of limited desk space. A long conversation may need summarizing; important details can become unavailable.

Hallucination

An AI output that invents or misstates information while presenting it plausibly. A nonexistent research citation is an example. Ask to see the actual source before relying on the claim.

Retrieval augmented generation or RAG

A method that retrieves outside information for a model to use when answering. A workplace assistant might consult a policy manual. It can still retrieve the wrong passage or interpret it incorrectly.

Multimodal AI

AI that handles more than one kind of information, such as text and images. You might ask about a photographed notice. Reading part of an image successfully does not establish that every detail was captured.

From useful answers to accountable action

Automation

Having a process run with reduced human effort, whether or not it uses AI. A scheduled payment is automation. Decide what exceptions should interrupt the process.

AI agent

A system that can choose steps and use tools to pursue a goal. It might compare travel options and make a booking if permitted. Its access and spending limits matter.

Human oversight

People checking and directing a system’s work. A reviewer needs information, time, skill, and authority to intervene. Clicking approval without those conditions provides little protection.

Delegation and abdication

In this book, delegation hands over work while retaining its purpose, boundaries, and accountability. Abdication leaves nobody meaningfully answerable. Having AI draft a reply differs from allowing unchecked decisions about customers.

Permission debt

This book’s term for the gap between a system’s capabilities and deliberately considered permission. A drafting tool gains power to send messages, but its approval remains unchanged. Review authority when capabilities or uses change.

Training debt

This book’s term for future loss of human competence when learning opportunities disappear without replacement. AI takes beginner assignments; beginners lose practice. Preserve ways to develop and demonstrate independent judgment.

When a convincing message asks for your trust

A supposed recruiter can know your résumé, discuss a plausible job, and answer careful questions. None of those details proves that the person or the job is real. AI can help produce a polished, personal exchange, but this warning also applies when no AI is involved.

When AI can imitate credibility, credibility is no longer verification. Before you send sensitive details or do consequential work, locate the organization through its own website or another trusted channel and confirm the opening and the person. Do not rely on a phone number or link in the original message. Chapter Nineteen explains the wider verification problem and the limits of the LinkedIn anecdote.

Sources and Notes

Technical reference notes checked September 25, 2026. Oversight and the final three entries follow the existing Living Book. OpenAI tools assisted preparation of this draft; provider explanations support terminology, not independent evidence of product reliability.

Basic technologies and models. IBM, What is artificial intelligence?; What is machine learning?; What is deep learning?.

https://www.ibm.com/think/topics/artificial-intelligence

https://www.ibm.com/think/topics/machine-learning

https://www.ibm.com/think/topics/deep-learning

Conversation and inference. IBM, What is a context window?; What are AI hallucinations?; What is AI inference?.

https://www.ibm.com/think/topics/context-window

https://www.ibm.com/think/topics/ai-hallucinations

https://www.ibm.com/think/topics/ai-inference

Retrieval and multiple input types. IBM, Retrieval augmented generation; Multimodal AI.

https://www.ibm.com/think/topics/retrieval-augmented-generation

https://www.ibm.com/think/topics/multimodal-ai

Products and agents. OpenAI, Key concepts and ChatGPT overview; Anthropic, Building effective agents.

https://developers.openai.com/api/docs/concepts

https://openai.com/chatgpt/overview/

https://www.anthropic.com/engineering/building-effective-agents