What AI is, and what it is not | Ekalavya Academy Short Book 1

A short book for owners and teams who teach themselves. Read it, try the exercise in your own tools, and check what you learned.

Ekalavya Academy by Almost Magic
Short Book 1 of 10 | Free to read and share

How to read this book

4. Check it
Three questions. Guidance is at the back, not next to the question.

Two levels in every chapter. The main slides are written so a first-time reader can follow them. A "Going deeper" slide gives the same idea at the level an experienced reader would argue about.

This book works with any AI tool you already have. It never asks you to buy or sign up for a particular product.

Why we are called Ekalavya Academy

In the Mahabharata, Ekalavya was a young archer who wanted to learn from a famous teacher and was turned away.

He did not give up. He made a clay figure of the teacher in the forest and practised in front of it every day, until he became an exceptional archer.

This academy is named for that habit. You can teach yourself, on your real work, with steady practice and honest checking.

THE IDEA

Learn by doing the real work, and keep checking yourself.

Where this book sits in the series

In this book

What AI is, and what it is not

AI finds patterns, then makes a good guess

The AI tools most people use today were trained on very large amounts of text, images or other data. During training, the system learns patterns: which words tend to follow which, what an invoice usually looks like, how a polite refusal is normally phrased.

When you ask a question, the tool uses those patterns to build the most likely useful reply. It is not looking the answer up in a book of facts. It is putting a reply together, one piece at a time.

That is why the replies often sound natural and helpful. It is also why they can be wrong.

IN PLAIN WORDS

AI is a very well-read guesser.

It can be wrong and still sound sure

Because an AI tool builds its answer from patterns, it can write a sentence that reads perfectly and is false. It may give you a name, a date or a figure that does not exist.

The writing style does not change when the content is wrong. A calm, well-organised answer is not evidence that the answer is true.

So the tone of an answer tells you nothing about its accuracy. The only test is to check it against something you already trust: the original document, your own records, or a person who knows.

REMEMBER

Confident is not the same as correct.

Where AI helps most, and what people use it for

Drafting
A first version of an email, notice or outline. You read it and edit it before it goes out.

Summarising
Long documents, meeting notes and email threads, shortened. You check the summary against the original.

Research and explaining
A hard idea in simpler words, or a first scan of a topic. You check the facts that matter.

Sorting and tidying
Labelling, grouping and cleaning lists, comments and spreadsheet columns.

Planning and productivity
Task lists, agendas, checklists and schedules you then adjust.

Repeating a routine
Turning a job you do every week into saved steps, which later books show how to automate safely.

Where you must check first

The more a job depends on a fact you cannot see, or the costlier a mistake would be, the more checking it needs.

A tool may give a neat total that is simply wrong. Check it with a calculator or your records.

A tool only knows what it was trained on and what you give it. It may not know last month, or your customer.

Hiring, pricing, credit and discipline need a person who can explain and own the decision.

From a one-off request to a workflow

Most of the talk about AI at work now is about productivity and automation. It helps to see them as steps, each with its own check. Book 3 covers steps 3 and 4.

Parts of the job run on their own, such as sorting an inbox or filling a template.

A tool takes several steps for you. A named person approves anything that matters.

Why checking is the scarce skill

Fluent is not calibrated
A language model is trained to produce likely text, not to report how sure it should be. Fluency and accuracy come apart, and the tool gives few reliable signals when they do.

Errors cluster at the edges
Mistakes concentrate in rare facts, recent events, exact figures and your own private data, which are the places a general pattern has least to say.

Drafts are cheap, checking is not
AI cuts the cost of producing a first version to near zero. The cost of checking stays with you, so design the work around where the check happens.

Treat output as a sample
Ask the same question twice and you may get different answers. A reply is one draw from a range, not a lookup, so record the version you checked.

A short example

Sam runs a small plumbing business. A supplier sends a long email about delivery changes. Sam asks an AI tool to summarise it and list the key dates.

The tool returns a tidy summary with three dates. Sam almost pastes it into the team calendar, but reads the email first.

One date is wrong: the email says the delivery moves to the following month, and the summary put it in the current one. A missed delivery would have held up a job.

The tool did not behave strangely. It did what it does: it produced a likely-looking list. Sam's check is what made it safe.

THE LESSON

Check the facts that matter before they leave your hands.

Ten minutes, your own tools

1. Pick a short document that is safe to share. It must hold no customer, staff or private business information.

3. Check each bullet against the document. Mark it: correct, partly correct, or not in the document.

YOUR TOOLS

Use any tool you already have. If you are not sure a document is safe to share, choose a different one.

GUIDANCE

What to look for is in the Guidance section at the end.

Check yourself

Question 1
A tool gives you a clear, well-written summary. Does that tell you the summary is accurate? Why or why not?

Question 2
Name two jobs where AI helps and the result is easy to check.

Question 3
Which of these needs a person to decide: choosing wording for a newsletter, or choosing who to shortlist for a job? Why?

Chapter 1 in four sentences

AI tools find patterns in large amounts of data and use them to make a good guess. A good guess can still be wrong, and it will sound just as sure. Use AI where the result is easy to check, and check the facts that matter before anyone acts on them. You own the result.

How AI learns, and why it gets things wrong

Training: where the patterns come from

Before you use an AI tool, its maker trains it on a huge collection of examples, such as text from books, websites and documents. Training lets the system pick up patterns in language and in other kinds of data.

It does not keep a neat library of facts to look up. It adjusts itself so that, given some words, it can predict what is likely to come next.

IN PLAIN WORDS

Training is practice at guessing what comes next.

Why it makes things up

When a tool has no good pattern for your question, it still produces something that fits the shape of an answer. That is how you get a made-up quote, a book that does not exist or a rule that was never written.

People call this "hallucination". It is not lying and it is not a rare glitch. It is the same guessing that gives you good answers, applied where the pattern is thin.

Be most careful where the answer must be exact: names, figures, dates, references and rules.

REMEMBER

A made-up answer looks just like a true one.

Four kinds of error

Wrong fact
The tool states something that is simply untrue, such as a figure, a name or a quote.

Out of date
The tool only knows what it was trained on, up to a cut-off date. Newer prices, rules or events may be missing.

Missing your context
It does not know your customers, contracts or prices unless you tell it.

Skewed by its data
Patterns in the training data can reflect the people and views most often written about, so answers can lean one way.

What the tool does and does not know

Usually good at
General patterns in language and common topics that many people have written about.

Often thin on
Rare subjects, local detail, and anything that happened after its cut-off date.

Cannot see
Your private records, emails and files, unless you give them to the tool.

Cannot judge
How sure it should be. It has no reliable way to tell you when it is guessing.

Training data is a sample, not the world

A sample with gaps
Training data is a sample of what people chose to write down and publish. Gaps and skews in the sample carry through into answers.

Tools that can look things up
Some tools can search the web or read your documents. The question then becomes whether the source is right, and the tool can still misread a good source.

Same idea, different tools
Models differ in their training and settings. Two tools may disagree on the same question, and neither is the authority.

Plausible is the target
A text model is built to produce plausible text. Plausibility is therefore the one thing you cannot treat as evidence.

A short example

Ayesha runs a small bookkeeping firm. She asks an AI tool for the current rule on a deduction and for a link to the official page.

The tool gives a clear answer and a link. The answer is close to the rule as it was some years ago. The link leads nowhere.

Ayesha checks the official site and finds the rule has changed. She keeps the tool's plain-English explanation of the idea, and takes the rule itself from the official source.

The tool was helpful for the explanation. It was the wrong place to read the rule.

THE LESSON

Use the tool to explain. Use the source for the rule.

Find the errors

1. Ask your tool a question about something you know well, such as your own trade or town.

YOUR TOOLS

Use any tool you already have. Do not paste private or customer information into the question.

GUIDANCE

What to look for is in the Guidance section at the end.

Check yourself

Question 1
Why might a tool give you a book title that does not exist?

Question 2
What does the training cut-off date mean for a question about last month's prices?

Question 3
A tool gives you a link as its source. What do you do before you rely on it?

Chapter 2 in four sentences

AI tools learn patterns from large collections of examples, then predict what is likely to come next. Where the pattern is thin, they can invent details that look real. A tool can be out of date and can miss your own context. Use it to explain and draft, and use a trusted source for the facts.

Asking well: ask, check, correct

A good request has four parts

A clear request says what the job is, who the result is for, what material the tool should use and what a good answer looks like.

For example: "Write a short reminder to customers whose invoices are 30 days overdue. Friendly, three sentences, no threats. Use only the details below."

A vague request, such as "write a reminder", leaves the tool to guess all of this.

FOUR PARTS

Job, reader, material, shape.

Four things that make a request better

Give context
Tell the tool who the result is for and why. Paste in only material that is safe to share.

Set the format
Ask for a length, a layout or a list, so the result is easier to check and use.

Say what to leave out
Ask it not to invent figures or names, and to mark anything it is unsure of. This helps but does not replace checking.

Ask what it assumed
Ask the tool to list the assumptions it made. Wrong assumptions are easier to spot when they are written down.

A checking routine

1. Mark what matters
Highlight the names, numbers, dates and claims that someone will act on.

2. Compare with a source
Check each one against the original document, your records or an official page.

3. Test the logic
Read the reasoning. Does each step follow? Is anything missing?

4. Decide and record
Accept, fix or reject. Save the version you checked.

Correcting the tool

If an answer is wrong, say what is wrong and what you want instead. Short, specific feedback works better than "try again".

A tool may agree with your correction even when you are the one who is mistaken. If it matters, check your correction as well.

Start a fresh request when the conversation has drifted. A clean request with the right context often beats a long argument.

REMEMBER

You are the editor. The tool gives you the first draft.

A prompt is a specification

A request is a spec
Treat your request like a short brief to a capable but literal colleague who cannot ask questions. What you leave out, the tool fills with the most likely guess.

Longer is not better
Extra detail helps only when it is relevant. Pasting in everything can bury the point and share information you did not need to share.

Agreeable by habit
Tools often go along with the person asking. A leading question tends to get a leading answer, so ask neutral questions when you want a real check.

A second tool is a weak check
Asking another tool to review the first can catch slips, but both may share blind spots. A source or a person who knows is stronger.

A short example

Tom manages a small removalist company. He asks a tool to "write a reply to a complaint".

The reply is polite but promises a refund that Tom never offered. He rewrites the request: the job, the facts of the complaint, what he is willing to offer, and "do not promise anything else".

The new reply fits. Tom checks the offer against what he meant to give before he sends it.

The tool did not get smarter. Tom gave it something it could get right and he could check.

THE LESSON

A clear request makes an answer you can check.

Ask, check, correct

1. Pick a small writing job from your week, such as a customer email. Do not use private details.

3. Rewrite it with the job, reader, material and shape, and compare the two results.

YOUR TOOLS

Use any tool you already have. Keep the vague and the clear versions so you can compare them.

GUIDANCE

What to look for is in the Guidance section at the end.

Check yourself

Question 2
A tool agrees at once when you say its answer is wrong. Does that prove you are right?

Question 3
Why is asking a second tool to check the first not enough for something important?

Chapter 3 in four sentences

A clear request names the job, the reader, the material and the shape of the answer. A short checking routine is: mark what matters, compare it with a source, test the logic, then decide and record. When the tool is wrong, say exactly what is wrong. You are the editor, and the tool provides the first draft.

Putting it to work on a real task

Choosing the job

Safe to share
The material holds no private or confidential information.

Five steps from start to finish

Who signs off

Every result that leaves your business needs a person who owns it. For small jobs that is you. For anything that affects a customer, an employee or money, name the person.

Write the name next to the task. "The tool did it" is not an answer if something goes wrong.

REMEMBER

A named person owns each result.

Keep, change or stop

After a few runs, look back. Keep what saved time and held up under checking. Change what needed too much fixing. Stop what created risk or extra work.

Note how long the task took with and without the tool, in your own records. Do not rely on a vendor's claims of time saved.

REMEMBER

Measure in your own work, not in advertising.

Where the check sits in a workflow

Place the check on purpose
Decide where in the process a person looks at the output, and what they are looking for. A check with no defined target tends to turn into a glance.

The hidden cost
Time spent checking and fixing is part of the cost. Some jobs end up slower with a tool, and the only way to know is to measure.

Drafts before decisions
Start with jobs where the tool drafts and a person decides. Move toward more automation only after the checks have held up for a while.

Keep a log
A simple log of what you asked, what you checked and what you changed lets you repeat what works and explain what happened to someone else.

A short example

Priya runs a small cleaning business. Every Friday she writes a message to the team about next week's jobs. She asks a tool to turn her rough list into a clear message.

She checks the addresses and times against her roster, fixes one wrong time, and sends the message under her own name.

After three weeks she keeps the routine and adds "check addresses against the roster" to her notes. She stops asking the tool to guess job lengths, because it kept getting them wrong.

THE LESSON

Keep what holds up. Change or stop what does not.

A real task, start to finish

1. Pick one job from your week that is frequent, checkable and safe to share.

YOUR TOOLS

Use any tool you already have. Keep the note: Book 2 builds on it.

GUIDANCE

What to look for is in the Guidance section at the end.

Check yourself

Chapter 4 in four sentences

Pick a job that is frequent, checkable, low in stakes and safe to share. Run it end to end: ask, check, correct, decide and record. Name the person who signs off. Look back and decide what to keep, change or stop.

Guidance and answers

Questions and exercises stay in the chapters, on their own slides. Guidance lives here, so you can try first and look afterwards.

Try the task or question in your own words, then compare. Where your answer differs, that is worth a note, not a correction.

WHERE TO FIND IT

The Guidance section below: answers and exercise guidance, one part for each chapter.

Chapter 1: answers and guidance

Answer 1
No. A tool can write a calm, well-organised summary that contains wrong facts. Style does not tell you about accuracy. Compare it with the original.

Answer 2
Examples: drafting an email you will read before sending, and shortening a document you can compare with the original. In both, you can see and fix the result.

Answer 3
Choosing who to shortlist. It affects a person, so someone must be able to explain and own the decision. Newsletter wording is low-risk and easy to review.

Exercise
Most summaries are mostly right and wrong in small ways: a number changed, a point added that is not in the document, or a caveat dropped. If you found no errors, check once more for what the summary left out.

Chapter 2: answers and guidance

Answer 1
It builds answers from patterns, not from a list of real books. Where it has no good pattern, it still produces something that fits the shape of an answer.

Answer 2
The tool may not know about recent changes. Check current prices with a current source, or give the tool the new information yourself.

Answer 3
Open the link and read the page. Check that it exists, that it says what the tool claims and that it is a source you trust.

Exercise
Most people find at least one source that does not exist or does not say what the tool claimed. Notice which kind of fact failed. Names, figures and references fail most often.

Chapter 3: answers and guidance

Answer 1
The job, the reader, the material to use and the shape of the result, such as length or layout.

Answer 2
No. Tools often go along with a correction. If it matters, check your own correction against a source.

Answer 3
Both tools can share the same blind spots. A trusted source or a person who knows is a stronger check.

Exercise
The clearer request usually gives something closer to what you wanted and easier to check, because you named the material and the shape. Notice which part of the request made the biggest difference.

Chapter 4: answers and guidance

Answer 1
It happens often, you can check the result quickly, a mistake costs little and the material is safe to share.

Answer 2
Someone must own the result. Naming the person stops checking from being skipped and gives you an answer if something goes wrong.

Answer 3
Claims of time saved come from other people's work. Your own numbers show whether it helps in your jobs, including the time spent checking.

Exercise
A good note names the job, what you asked, what you changed, how long it took and what you will do next time. If the tool saved no time, that is a useful result too.

About Ekalavya Academy

Ekalavya Academy is a learning series from Almost Magic Tech Lab. The books are written for people who teach themselves, using the tools they already have.

Almost Magic builds tools that help people check and govern their AI use. The books never depend on those tools.

LINKS

Almost Magic: almostmagic.net.au

Case studies: ai-casestudies.almostmagic.net.au

Page for this book on the Academy site: link to be added when the page exists.

Book 1	What AI is, and what it is not	Foundations
Book 2	Productivity with AI: writing, summaries, research, meetings	Practical use
Book 3	From one task to a repeatable workflow, and AI deputies	Workflow automation, agents
Book 4	Handling data safely	Data handling
Book 5	When a data shortcut becomes the rule	Data quality and governance
Book 6	Staying secure with AI	Security
Book 7	Rules, risk and what you must know	Risk and compliance awareness
Book 8	Your AI rules on one page	Policy and accountability
Book 9	When AI gets it wrong	Recovering from AI mistakes
Book 10	When not to use AI, and who stays in charge	Judgement and oversight