Opens in a new tab
needsahuman.

What is generative AI? AI in plain English

What is generative AI, and why does it sometimes get things wrong? This lesson explains what today's AI assistants are and how they work, in plain words, with no maths and no jargon left unexplained. In this lesson What an AI model is, and the difference between training it and using it. Why today's assistants…

What is generative AI, and why does it sometimes get things wrong? This lesson explains what today’s AI assistants are and how they work, in plain words, with no maths and no jargon left unexplained.

In this lesson

  • What an AI model is, and the difference between training it and using it.
  • Why today’s assistants are fluent and fast, and why they can be confidently wrong.
  • The three habits that make wrong answers much less likely.
  • How the new “generative” AI differs from the older AI already inside much of your software.
  • What a prompt, a model and a context window are.

A model is trained once, then used

When people say “AI” today, they usually mean an assistant such as ChatGPT, Claude, Gemini, Copilot or Grok. Underneath each one is a model: a very large computer program that has learned patterns from huge amounts of text, and often images, sound and code too.

There are two stages in a model’s life.

  1. Training. The maker feeds the model a vast amount of material and adjusts it, over weeks or months, until it is good at predicting what comes next. This is expensive and done by a small number of companies. The material it learned from is its training data.
  2. Use. Once trained, the same model answers millions of questions. When you type into an assistant, you are using the model, not training it. It does not learn from your conversation in the way a new employee would, unless the product is built to remember things or the maker uses conversations to train later versions. (Lesson 2 covers how business plans handle that.)

That split matters for a business owner. The model knows what was in its training data, up to a cut-off date. It does not know your prices, your customers or yesterday’s news unless you, or the product, give them to it.

Prediction, not lookup

The single most useful idea in this course is this: a language model predicts likely text. It does not look facts up.

When you ask a question, the model produces an answer one small piece at a time, each piece chosen because it is a likely next step given everything before it. Those pieces are called tokens; a token is roughly a short word or part of a word. A model that has read a great deal of good writing about, say, sales tax or employment contracts produces fluent, sensible text about them, because that is what likely text on the subject looks like.

This is why these models are called large language models (LLMs), and why they are so good at:

  • Language: drafting, rewriting, translating, changing the tone of a message.
  • Summaries: turning a long document, call or email thread into the main points.
  • Patterns: sorting, tagging and extracting details from messy text, such as pulling the invoice number, date and amount out of a supplier email.
  • Breadth: being reasonably useful on almost any subject, at any hour, in seconds.

It is also why they go wrong in a particular way.

Why it makes things up: AI hallucination

Because the model is producing likely text rather than checking a record, it can produce an answer that sounds right and is not. It might give a plausible but wrong date, invent a policy, cite a source that does not exist, or get a sum slightly off. People call this hallucination. The model is not lying, and it does not know it is wrong: a confident, well-formed answer is simply what likely text looks like.

Wrong answers are more likely when:

  • the question is about something specific to you that the model was never given, such as your refund policy;
  • the answer depends on recent events after its training cut-off;
  • the task needs exact figures or careful arithmetic over many steps;
  • the honest answer is “it depends” or “nobody knows”, and the question invites a definite one.

Three habits that reduce it

You cannot switch hallucination off, but three habits make it much rarer and much easier to catch.

  1. Give it your documents. Paste in, attach or connect the material the answer should come from: your price list, your policy, the contract. The model then works from your text instead of from memory. When a product does this automatically, searching your files and passing the relevant parts to the model, it is called retrieval, or “RAG” for short.
  2. Ask for its sources. Ask the assistant to quote the part of your document, or link the web page, that supports each claim. A claim with no source is a claim to check.
  3. Check before it leaves the building. Treat every AI answer as a draft. The person who would have written it checks it before it goes to a customer, a supplier or the IRS.

Flow: your words and your documents go into the model. The model writes a draft. You check the draft. Only then is it sent.

Your wordsthe question or instruction
Your documentsprices, policies, notes
Modelpredicts likely text from its training and what you gave it
Draftfluent, fast, and sometimes wrong
You checkfacts, figures, tone
Sent
A model predicts; it does not look things up. Giving it your own documents makes the draft better. Checking it before it goes out is what makes it safe.

That last habit is the thread running through this whole course. You will see a small amber box marked “you check” in almost every diagram. Amber, on this site, always means a person.

Your words, your documents, the context window

Three more terms come up constantly.

  • Prompt. What you type: the question or instruction. A good prompt says what you want, for whom, in what format, and what to use. “Write a reply to this complaint” works; “Write a polite reply to this complaint, under 120 words, offering the refund our policy below allows, and no more” works better.
  • Context window. How much the model can take in at once: your prompt, any documents you attach, and the conversation so far. Modern models can hold a lot, often whole reports, but not unlimited amounts. If a conversation runs very long, early details can be dropped or given less weight. For an important task, start a fresh conversation and give it what it needs.
  • Model. The engine underneath. One product may offer several models, from fast and cheap to slower and more careful. Lesson 2 explains who makes them.

What is generative AI, and how is it different from the AI you already have?

Generative AI, the kind that writes, draws and talks, is the new part. But your business has probably used AI for years without calling it that. The older kind is usually called predictive AI: models that learn from past data to score, sort or forecast.

  • Your email’s spam filter predicts which messages are junk.
  • Your card processor’s fraud check predicts which payments look wrong.
  • Your accounting or stock software may forecast sales or flag unusual transactions.
  • Your phone’s voice typing turns speech into text.

Predictive AI gives a narrow answer to one fixed question (spam or not, fraud or not, how many units next month). Generative AI produces open-ended output (a paragraph, a picture, a spoken reply), which makes it far more flexible and far harder to check automatically. That is why generative AI needs a person checking its work in a way the spam filter never did.

Five types of AI. Predictive AI: spam filters, fraud checks, demand forecasts. Generative text: drafting emails, summaries, answers from your documents. Image: reading receipts, product photo clean-up, mock-ups. Voice: phone answering, call transcription. Video, the newest and least proven: captions, short training clips.

Predictive AI

Spots patterns in past data to sort or forecast. Already inside a lot of software.

  • Spam filter on your inbox
  • Fraud check on card payments
  • Demand forecast for stock

Generative text

Writes new text from a prompt. What most people mean by "AI" today.

  • Drafting emails and replies
  • Summarizing meetings
  • Answering from your documents

Image

Reads or makes pictures.

  • Reading scanned receipts
  • Cleaning up product photos
  • Mocking up a flyer

Voice

Hears and speaks.

  • Answering the phone
  • Transcribing calls
  • Reading text aloud

Video

Newest, least proven for business use.

  • Captions for your videos
  • Short training clips
Predictive AI has been in business software for years; generative AI is the new part. The dashed border marks the least proven type.

Text, images, voice and video

Generative models now come in several kinds, and many assistants combine them:

  • Text models draft, summarize and answer questions. Most business use starts here.
  • Image models create or edit pictures from a description, used for social posts, mock-ups and product shots.
  • Voice models transcribe calls, take notes in meetings, and power phone assistants that can hold a spoken conversation.
  • Video models create short clips from a description. They are the newest and the least used in ordinary businesses so far.

What changed, in one paragraph

Language models existed before, but they became a household tool after OpenAI released ChatGPT on November 30, 2022, saying “during the research preview, usage of ChatGPT is free” (OpenAI). Since then the models have become better at following instructions, working with longer documents, reading images and speaking aloud, and they have been built into software that businesses already use, such as email, word processors and accounting packages. The biggest recent change is that some can now take actions, not just write text: book the meeting, fill in the form, update the spreadsheet. Those are called agents, and Part 3 of this course is about them. What has not changed is the core idea: a model predicts likely output, so a person still needs to check what matters.

How we judge what AI can do

On NeedsAHuman, we rate every task in the official US task list for each job as AI does it, AI helps or needs a human. That rating is about whether today’s AI can do the task to the standard a job needs, not whether it can produce something that looks like it. How we do it, and its limits, is on our page about whether AI can do a job’s tasks. Lesson 4 shows how to use those ratings for the roles in your own business.

Key terms

Every term is explained in the glossary.

Try this

Pick one document your business relies on, such as your returns policy, your opening hours and services, or a standard quote. Ask any AI assistant three questions about your business without giving it the document, and note what it gets wrong or invents. Then ask the same three questions with the document pasted in, and ask it to quote the line it used for each answer. Compare the two sets. This ten-minute test shows hallucination, and the cure, better than any explanation.

(Before you paste anything in, check what the assistant’s terms say about your data. Lesson 2 explains what to look for. Do not paste in anything private about customers or staff on a free personal account.)

Sources

This lesson is conceptual and quotes no study figures. One date is sourced: