PrivateBox · Learn
The Course Book. Every lesson from the interactive course in one readable document, written to teach, not just to remind.
Ask an AI who you are and it will answer warmly, confidently, and probably invent half of it. That single strange fact is the key to everything else in this course, because it shows what the machine really is.
AI is the world's best autocomplete. Your phone suggests the next word based on your texting history. A language model does the same thing, except it learned from a huge slice of everything humans have ever written. Scale changes what is possible. It does not change what the machine fundamentally is.
When it answers you, it writes one word at a time, and each choice narrows what can plausibly come next. It never plans the whole answer, never checks a database, and never recalls a fact. It learned the patterns in a billion books, not the pages. Two consequences follow directly:
Models read tokens, chunks of roughly three-quarters of a word. Training happened once, in the past; when you chat, the patterns are frozen and the model learns nothing from your conversation. And because patterns do not keep receipts, the model genuinely cannot tell you where it learned something.
When you use ChatGPT, Claude or Gemini, you are talking to a product built around a model. The company adds hidden instructions, fixes the creativity dial, and bolts on abilities like search and file reading. The machine underneath is still a prediction engine, and when the polish slips, it fails in exactly the ways this book describes.
AI is finishing your sentence, not searching a database. Fluency is how it sounds, not how it knows.
In 2023 a lawyer submitted a legal brief written with AI help. It cited six court cases. Judges checked, and none of them existed. The AI had not lied. It had done exactly what Lesson 1 says it does: it produced the most plausible-sounding text, and plausible legal citations look exactly like real ones.
This is called hallucination. There is no intent behind it, only probabilities, which is why researchers prefer the word confabulation: sincerely filling gaps with invention. The model fills three kinds of gap most confidently:
The crucial rule: confidence is not accuracy. Tone never signals truth, because the tone is itself just predicted text. What does signal risk? Two smells: suspiciously specific numbers, and very smooth answers about very obscure things.
Ask your AI these two questions, in this order:
What do you know about [your name / your small company]?
List everything in your last answer that you would need to verify.
Your AI may answer the first question with real facts about you. That is not the model's training. Most big AI apps now have a memory feature that saves details from your earlier chats, and many can search the web to cover recent events. These features help, but memory only knows what you told it, and a search result can be summarised wrongly with the same confident tone. Verify anything important. The industry has reduced hallucinations. Nobody has eliminated them.
Fluent does not mean true. The more obscure the topic, the more beautiful the guess. Verify anything that matters.
Two hours into a chat, your AI gets dumber. Instructions you gave at the start quietly stop applying. You are not imagining it, and it is not the model that changed. Three failures explain almost all of it.
Everything in a conversation is input the model re-reads on every reply: your messages, its answers, every tangent. A study across 18 leading models found that accuracy falls as input grows. Models did worse with a full conversation history than with only the relevant excerpt. Long chats bury the important instruction under noise.
Models are agreeable. Ask "why is X true?" and the model will cheerfully explain a falsehood. Frame a bad plan positively and it becomes your loudest cheerleader. The frame you choose dictates the answer you get, so never ask a question that contains its own answer.
One wrong "fact" early in a chat gets treated as established truth for everything after it. Correct it explicitly, or start clean.
The four habits that fix all three:
Modern AI apps know long chats degrade, so they compress older parts of the conversation and carry key facts between chats using memory. This raises the ceiling but does not remove it, and compression can drop the exact detail you cared about. A fresh chat with a clear restatement is still the sharpest tool you have.
Long chats rot. Leading questions get echoes. Fresh chat plus neutral framing gives you a smarter AI, instantly.
A chainsaw is neither good nor bad. It is the right tool for some jobs and a catastrophe for others. The whole framework for AI fits in two questions you can ask about any task:
Two yeses: green light. Drafts, rewrites, summaries of text you supply, brainstorming, explaining, categorising, adjusting tone. You are the quality control and mistakes cost seconds.
Any no: slow down. Two nos is the danger zone: legal, medical and financial answers you cannot verify, unsupervised decisions, anything with your name on the line. That is not where AI helps you. It is where it helps you fail fluently.
Do: give it source material to work from. Ask for three options, not one. Use it to critique your own work before others see it. Make it explain anything you do not understand.
Don't: paste secrets into free public tools. Outsource judgement or final decisions. Trust its memory over your documents. Skip verification on anything that matters.
Use AI where you can check it and mistakes are cheap. Keep judgement, and secrets, on your side of the desk.
"Write me a LinkedIn post" gets you generic mush. The same request, properly briefed, gets something you would actually publish. Same model, same minute. Only the brief changed. The framework is four slots, and you will not always need all four:
The golden rule: if a colleague reading only your message would be confused, so is the AI. Three habits compound with the brief:
Goal: [what you want + who it's for + why]
Context: [the details that change the answer]
Example: [paste a sample whose style you like]
Format: [length, tone, structure]
Say "I don't know" rather than guessing. Ask me anything unclear before answering.
Most AI apps let you save custom instructions: a standing brief that applies to every new chat. Who you are, how you like your answers, what to avoid. It is the same four-slot idea written once instead of every time. Five minutes setting it up pays off in every conversation after.
Goal, Context, Example, Format. Thirty extra seconds of briefing beats thirty minutes of fixing.
Every car has four wheels, and a hatchback and a freight truck are still very different purchases. Every AI model predicts text, and they are still very different tools. Four facts cover 90% of real decisions:
Many AI products now route your question automatically between fast and thinking modes. Handy, but not always right. Knowing the difference means you can switch modes yourself when a hard problem gets a shallow answer.
Everyday task, fast model. Hard problem, thinking model. Current events, web access. Disappointed? Try a bigger model before blaming AI.
Everything so far used AI's general knowledge. The real prize is asking questions of your contracts, notes, manuals and records. This is also where most people make their worst AI mistake: pasting confidential material into free public tools that nobody in their organisation ever approved.
The model has never read your files. Pasting them in every time is clumsy and invites context rot. The better pattern is document-grounded AI, which works in three steps:
Instead of asking the intern to remember, you hand them the right page. Grounded answers cite their source passages, so you can verify in one click. And something subtle becomes possible: "that isn't in the documents", an honest no. A pure prediction machine cannot say that. A grounded one can. The pattern is called RAG, Retrieval-Augmented Generation, and it also sidesteps context rot, because the model receives only the relevant excerpts rather than your whole archive.
Attaching a file to a chat is the everyday version of this idea: the app reads your document and answers from it, for that one conversation. Document-grounded systems take it further: a whole library, indexed once, shared with your team, with sources on every answer. This is the approach we build at PrivateBox, with your documents kept in a private workspace instead of a public tool.
AI plus your documents, privately, with sources. That is when it stops being a toy. Never feed confidential files to tools you haven't vetted.
Introduction to AI · Course Book v1.0 · Take the interactive course free at privatebox.co.za/resources/intro-to-ai · Built by PrivateBox, private AI for your organisation's documents.