Using AI Without Getting Burned

Part 1

What it actually does, in one page

You do not need the mathematics, but you do need the shape of it, because every strength and every failure follows directly from this.

1. 1

   It predicts likely text

   It was trained on an enormous amount of writing and learned which words tend to follow which. When you ask a question it produces a plausible continuation. That is the whole mechanism. It is astonishingly effective and it explains everything below.

2. 2

   It is not looking anything up

   Unless it explicitly searches the web and shows you links, it is not consulting a database. It is generating text that resembles a correct answer. Most of the time resemblance and correctness coincide. Sometimes they do not, and nothing about the output changes when they do not.

3. 3

   It does not know when it is wrong

   There is no internal signal that separates a fact it reproduced accurately from one it assembled. This is why the confident tone is dangerous: the confidence is a feature of the writing style, not a measure of reliability.

4. 4

   It has a training cutoff

   It was trained up to a certain date and knows little after it. Ask about a recent event and it may answer from before the cutoff without saying so. Ask it directly what its cutoff is.

5. 5

   It agrees with you more than it should

   Push back on a correct answer and it will often fold and apologize. That is a trained politeness, not a reconsideration. Never treat agreement as confirmation — if anything, test it by pushing back on something you know is right.

Part 2

The three kinds of question

Almost all of the risk is decided before you type anything, by which kind of question you are asking. Sorting them takes a second and saves the trouble entirely.

1. 1

   Shaping — very reliable

   Rewrite this politely. Summarize this. Make this shorter. Turn these notes into a letter. Explain this like I am twelve. Here you supply the facts and the tool supplies the language, so it has nothing to invent. This is where nearly all the everyday value is.

2. 2

   Thinking — useful with judgment

   Give me ten ideas. What am I forgetting. What would someone argue against this. Wrong answers cost nothing because you are filtering them anyway. Excellent for getting unstuck and for pre-empting objections.

3. 3

   Recalling — treat as unverified

   What year did this happen. What is the dose. What does the law say. Which study showed this. What is this company's phone number. Every one of these has a real answer that exists somewhere, which means there is something for the tool to get wrong. Verify all of it.

4. 4

   The quick test

   Ask yourself: am I giving it the facts, or asking it for facts? Giving is safe. Asking needs checking. That single question catches most of the trouble.

Part 3

Getting better answers

Most disappointing results come from a thin question. These adjustments take seconds and change the output substantially.

• Say who it is for. 'Explain this to my 80-year-old mother' produces a different and better answer than 'explain this'.
• Give it the raw material. Paste your actual notes, the actual email, the actual job description. It cannot use what it has not been shown.
• Say how long. 'In three sentences' or 'one page' stops the padding.
• Tell it the tone: firm but polite, warm, plain, formal. Otherwise you get a default that suits nothing in particular.
• Ask for options rather than an answer. 'Three versions, different tones' beats one take you then have to argue with.
• Say what you do not want. 'No bullet points', 'do not use the word leverage', 'no introduction'.
• Follow up rather than restarting. 'Shorter, and warmer' works. Retyping the whole question loses the context.
• For anything factual, add: 'If you are not sure, say so.' It does not make it reliable, but it does surface some uncertainty that would otherwise be hidden.

Part 4

Checking an answer in under a minute

Verification does not have to be laborious. Scale the effort to what being wrong would cost.

1. 1

   Ask where it came from

   'What is your source for that?' If it produces a link, open the link. Do not accept the existence of a citation as evidence — invented sources are common, and a fabricated link often looks entirely normal until clicked.

2. 2

   Search one specific claim

   Take the most load-bearing fact — the number, the name, the date — and search it directly. If it is real, it appears immediately on a page that is not an AI answer.

3. 3

   Ask it again in a new conversation

   Start fresh and ask the same question differently. Consistent answers are weak evidence of correctness; different answers are strong evidence of a problem.

4. 4

   Check anything with a number twice

   Prices, doses, dimensions, interest, deadlines, phone numbers. Arithmetic and digits are a known weak spot, and a digit is exactly the kind of thing that is easy to get wrong and expensive to get wrong.

5. 5

   Match the check to the stakes

   A birthday message needs no checking. A quote to a customer, a medication name, or anything a court will see needs a real source. The judgment is yours and it is the actual skill.

Part 5

Everyday jobs it genuinely does well

Concrete uses that hold up, drawn from what people actually come to us with.

• Turning a rambling voice note or scribbled list into a clear email or letter.
• Explaining a letter you do not understand — insurance, medical, legal — so you know what to ask about. Understanding it, not acting on it.
• Drafting the awkward message: chasing an unpaid invoice, declining an invitation, complaining effectively.
• Summarizing a long document down to what it asks of you, then reading the original for anything that matters.
• Practicing a difficult conversation, or an interview, by having it play the other person.
• Translating, and explaining what a phrase implies rather than just what it says.
• Helping with homework as a tutor rather than an answer key: 'Do not tell me the answer, ask me questions until I find it.'
• Making a first draft you can react to. Editing something mediocre is far easier than facing a blank page.

Part 6

The honest limits

Things people are regularly surprised by, listed plainly so you are not the one finding out the hard way.

• It cannot browse the internet unless it says it is doing so and shows you links.
• It cannot see your files, your email, or your calendar unless you connected it deliberately.
• It does not remember previous conversations unless the product says it does.
• It is poor at counting things and at precise arithmetic. Use a calculator.
• It will invent citations, statistics, and quotations that look completely legitimate.
• It reproduces the biases in what it was trained on, which shows up in advice about people.
• It cannot tell you what it does not know, because it has no representation of the boundary.
• Two different tools will give you two different answers, and neither will mention that.