AI in Plain English

Free intro · AI in Plain English

1. Concept: It predicts words

   It writes by guessing the next word over and over, from patterns in a mountain of writing, instead of looking facts up in a database.

   1. Term: Autocomplete, grown up

      It's the same trick as your phone suggesting the next word in a text, scaled up until it can write a whole letter.

      1. In plain words: Your phone keyboard's big brother

         When you text, your phone guesses the next word. ChatGPT, Claude and Gemini do the same thing, one piece of a word at a time, until a whole answer comes out. That one idea explains both the magic and the mistakes.

         See also Hallucinations

      2. How it works: Guess, add, repeat

         It reads the whole conversation so far, picks a likely next piece of a word, tacks it on and goes again, very fast. Techies call those pieces tokens. Nobody wrote the answer ahead of time; it's built fresh every time you ask.

      3. Try it: Ask twice

         Ask the same question in two new chats. The answers will be close but not word-for-word the same, because it's predicting, not reading from a script.

   2. Term: Trained on a mountain of writing

      Before you ever typed a word, it studied a huge pile of text and learned how words tend to go together.

      See also Stale answers

      1. In plain words: It read the library, then stopped

         Before a model is released, it studies a huge pile of writing: books, websites, code and more. Then the studying stops. The date it stopped is called its knowledge cutoff.

      2. How it works: Patterns, not a filing cabinet

         It mostly keeps patterns, not pages it can pull up like files, the way you can hum a song without the sheet music. That's why it can write about almost anything, and why details sometimes come out scrambled.

      3. Watch out: Last week isn't in there

         Anything after the cutoff is news to it unless the tool searches the web, and cutoffs often run months behind the calendar. xAI, for one, lists its Grok 4.7 model's cutoff as May 2026 (checked Oct 23, 2026).

         See also Not a search engine, unless it searches

   3. Term: Not a search engine, unless it searches

      A plain chatbot answers from memory, but many can search the web first, and that changes how far you can lean on the answer.

      See also Stale answers

      1. In plain words: Memory versus a phone call

         Without search, it answers from memory, like a well-read friend at the counter. With web search on, it makes a quick phone call to the internet first and can show you where it looked.

      2. How it works: Look for the links

         ChatGPT, Claude, Gemini and Copilot can all search the web, and when they do, the answer usually comes with links or footnotes. No links usually means no search, so treat that answer as memory.

      3. Watch out: A link still needs a look

         A link can point to a page that says something different, or one that's out of date. Open one or two and find the sentence that backs up the claim.

         See also Hallucinations

2. Concept: It can be confidently wrong

   Because it predicts what sounds right, it says wrong things in the same calm voice it uses for right ones.

   1. Term: Hallucinations

      That's the industry's word for a confident answer that's simply made up.

      See also The traffic light

      1. In plain words: Confident wrongness

         A hallucination is the AI stating something false like it's gospel: a fake quote, a made-up number, a court case that never happened. It isn't lying. It's predicting words that sound right.

      2. How it works: Sounds right beats is right

         Most of the time, sounding right and being right line up. When the model doesn't know, it tends to fill the gap with something that sounds about right instead of saying "I don't know." Math can slip the same way: unless the tool runs a calculator in the background, it's predicting the digits rather than adding them up.

         See also Autocomplete, grown up

      3. Try it: The about-me test

         Ask a chatbot to write a short history of your business or your hometown. Circle what's wrong, and notice the wrong parts sound every bit as sure as the right ones.

      4. Watch out: It happened in an Iowa court

         In 2025, the Iowa Court of Appeals caught a lawyer's AI-assisted brief citing two court cases that never existed, and the lawyer hadn't checked them before filing. The court ordered a $150 penalty, and the state's Attorney Disciplinary Board followed with a public reprimand, reported in May 2026.

   2. Term: Stale answers

      Its memory stops at a cutoff date, so prices, rules and deadlines can be out of date with no warning label.

      See also Trained on a mountain of writing, Not a search engine, unless it searches

      1. In plain words: Last year's newspaper

         Asking a model about something recent, with no web search, is like reading last year's paper: well written and confidently out of date.

      2. Try it: Ask what changed

         Ask about a deadline, price or rule you know changed this year. If the answer has no link and no date, check the official page yourself before you act on it.

      3. Watch out: A real Iowa example

         USDA's REAP energy grants, for farms and rural small businesses, changed on Oct 16, 2026. Now you build the project first and apply only after it has 12 months of energy data. An AI trained before that can walk you through the old process without blinking.

   3. Term: The traffic light

      Sort the work three ways: green for drafts and ideas, yellow for facts you'll check, and red for calls a person makes.

      See also A tool, not a teammate

      1. In plain words: Green, yellow, red

         Green: drafts, ideas and summaries of things you handed it. Yellow: names, numbers, dates, links and anything recent, which you check. Red: final calls on money, health, law or people, which a person makes.

      2. Try it: Color your week

         List five tasks from your week and give each one a color. The green ones are where AI can start helping tomorrow.

      3. Watch out: Green still gets read

         Even a green-light draft can carry a wrong fact or a tone you'd never use. Read it before it goes out with your name on it.

         See also Swinging it: what you type

3. Concept: It's a tool, with several makers

   ChatGPT, Claude, Gemini and Copilot are rival products from different companies, and a smaller cousin can even run on your own laptop.

   1. Term: A tool, not a teammate

      It isn't alive, plotting or thinking about you; it's a power tool that does its best work with a steady hand on it.

      See also The traffic light

      1. In plain words: Hammers and windows

         AI is a tool. Hammers build houses and break windows. We teach swinging. It won't care how the job turns out, so that part stays yours.

      2. How it works: Good with words, shaky on facts

         Because it's a word predictor, it shines when the job is mostly words: drafting, summarizing, rewording, brainstorming, and turning messy notes into neat lists. It slips when the job needs today's facts, careful math, or details about your business nobody gave it.

      3. Try it: Messy to neat

         Snap a photo of a handwritten list with made-up names and ask for a table of names, phones and emails. Then check every digit. Alex built a phone app that does this with open-house sign-in sheets for a real estate agent, and it still has the agent check each contact before it sends.

      4. Watch out: More rope, more checking

         Newer "agent" features let AI click, type and send things for you. Same word predictor, longer leash, so keep a person approving anything that leaves the building.

   2. Term: Swinging it: what you type

      What you type is called a prompt, and clear instructions get you better guesses.

      1. In plain words: Order like it's the deli

         "Give me a sandwich" gets you a sandwich. "Turkey on rye, toasted, no onions" gets you lunch. Tell it who the writing is for, what you want, and how long it should be.

      2. How it works: Show it your voice

         Paste two or three emails you wrote, with names taken out, and say "match this voice." Copying patterns is what it does best. Alex has built a one-page voice guide on the same idea, so a client's AI-drafted emails sound like the client.

      3. Try it: Shorter. Friendlier. Again.

         Treat the first answer as a rough draft. Reply with "shorter," "friendlier" or "for a customer who's been with us 20 years" and watch it change.

      4. Watch out: It's never met your business

         It doesn't know your prices, hours or tone, so it will guess them. Give it the facts it needs, and only the ones that are safe to share.

         See also The coffee-shop rule

   3. Term: The big names

      It's the same basic idea from different companies, with different strengths, prices and rules for your data.

      See also Business plans, AI on your own laptop

      1. In plain words: Who makes what

         ChatGPT is made by OpenAI, Claude by Anthropic, Gemini by Google and Copilot by Microsoft. They all predict words, and they differ in style, features, price and what they do with your data. Even the independent scoreboards disagree on the best one: in October 2026, a crowd-vote ranking put a Google model first while a test-based index put an Anthropic model first.

      2. What it costs: Free to start, about $20 to go big

         Every one of them has a free tier, which is plenty for kicking the tires. Checked Oct 23, 2026: ChatGPT Plus is $20 a month (ChatGPT Go is $8), Claude Pro is $20 and Google AI Pro is $19.99. Paying mostly buys more use and stronger models.

      3. Try it: Same question, three answers

         Paste the same request into two or three free assistants and compare. You may like one for writing and another for research, and that's normal. No AI company pays Iowa AI, so the only vote that counts here is yours.

      4. Watch out: Free has a price tag too

         Free ChatGPT may show ads in the US (testing began in February 2026), and several free plans may use your chats for training unless you switch that off. Plans and names also change every few months, so check the maker's own pricing page before you buy.

         See also The training switch

   4. Term: AI on your own laptop

      Smaller free models can run right on a laptop, even with the Wi-Fi off, so nothing you type leaves the machine.

      See also Where your words go

      1. In plain words: Your own little AI

         A local model is an AI you download and run on your own computer. Free programs like Ollama and LM Studio handle the setup, and once the model is downloaded it keeps working with the Wi-Fi off.

      2. What it costs: Free, but slower

         The software and many models cost nothing. A small model like Meta's Llama 3.2 is about a 2 GB download (checked Oct 23, 2026) and runs on an ordinary laptop with no fancy graphics card, just a good deal slower than the cloud.

      3. Try it: The Wi-Fi-off test

         On a class ThinkPad, turn off the Wi-Fi, open the local AI and ask it to draft a thank-you note to a customer. It works, and not one word left the room.

      4. Watch out: Smaller brain, same rules

         Laptop-sized models are well behind the big cloud ones, so expect plainer writing and more mistakes. Use them for private drafts and for learning, and still check the facts.

4. Concept: Your data, and what not to paste

   Whatever you type into a cloud AI goes to that company's computers, so decide what goes in before you hit Enter.

   1. Term: Where your words go

      Typing into a cloud AI sends your words to the company's servers, where they're answered and usually kept for a while.

      See also AI on your own laptop

      1. In plain words: It leaves the building

         With ChatGPT, Claude, Gemini or Copilot, what you type travels to the company's computers and gets answered there. That's normal for cloud apps, same as email, but it means a copy lives somewhere besides your screen.

      2. How it works: Kept for a while

         How long depends on the company and your settings. Checked Oct 23, 2026: Gemini auto-deletes your activity after 18 months by default, but chats picked for human review are kept up to three years. Claude removes a chat you delete from Anthropic's servers within 30 days.

      3. Watch out: Some go overseas

         DeepSeek's privacy policy says it stores personal data in China, and in February 2025 Iowa's governor banned DeepSeek from state government use, calling it a security risk. If you do business with the state, check the rules before you pick an AI app.

   2. Term: The coffee-shop rule

      If you wouldn't hand it to a stranger at a coffee shop, don't paste it into a chatbot.

      See also Business plans, AI on your own laptop

      1. In plain words: The stranger test

         Before you paste, picture sliding the paper across to a stranger at the next table. If that makes you wince, keep it out.

      2. Try it: Swap in fake names

         Need help with a real customer email? Swap the name, address and account number for made-up ones first. The AI writes the reply just as well for "Pat from Waverly."

      3. Watch out: The never list

         Keep these out of personal AI accounts: passwords, bank and card numbers, Social Security numbers, customer lists with personal details, employee and health records, and anything under an NDA. Passwords stay out on every plan, period.

   3. Term: The training switch

      Many personal plans may use your chats to improve their AI unless you turn a setting off.

      See also Business plans

      1. In plain words: Who else learns from this?

         Training means the company may use your conversations to improve future versions of its AI, and some have people read samples. On most personal plans, you can turn it off.

      2. How it works: Where the switch lives

         Checked Oct 23, 2026. ChatGPT: Settings → Data controls → "Improve the model for everyone" (on by default). Gemini: turn off "Keep Activity" on the Gemini Apps Activity page (on by default for adults). Claude: Privacy settings → the model-improvement setting.

      3. Try it: Two minutes tonight

         Open the assistant you use most, find the setting and switch it off. For one-off questions, start a temporary chat (Claude calls it incognito); ChatGPT, Gemini and Claude all say those aren't used to train their AI.

      4. Watch out: Off isn't a time machine

         Switching it off protects what you type from then on. Anything already used to train a model stays in that model, and thumbs-up or thumbs-down feedback can be kept for years.

   4. Term: Business plans

      Team and business plans generally don't train on your company's data by default, and that promise is written into the contract.

      See also AI on your own laptop, The coffee-shop rule

      1. In plain words: A contract, not a toggle

         On ChatGPT Business, Claude Team, Google Workspace and Microsoft 365 Copilot, the default is no training on your company's data (checked Oct 23, 2026). That promise lives in the business terms, not in a switch someone might forget.

      2. What it costs: Per person, per month

         Checked Oct 23, 2026: Claude Team and ChatGPT Business each run $25 per person a month, or $20 a month billed yearly, and both start at two seats. Google Workspace Business Standard, with Gemini built into Gmail and Docs, is $14 per person a month on a yearly plan.

      3. Watch out: Still not a vault

         No training doesn't mean nobody else holds your data; it still sits on their servers. If HIPAA covers your business, like a clinic or dental office, you need a signed business associate agreement with the AI company before patient information goes in.