Lessons0 of 20 done

Module 3: Using AI well. Lesson 12 of 20, 8 minutes.

Trust but check.

How much checking an answer needs, and five concrete ways to do it: sources, code, reasoning, a second opinion and red flags.

A colleague forwards you a summary of a forty-page contract. A chatbot produced it in twelve seconds. It reads well. Three of its five bullet points are right. One is half right. One describes a clause that does not exist in the contract.

Nothing in the text tells you which is which. The wrong bullet is written in exactly the same confident tone as the right ones. This lesson is about what to do with that fact.

Match the checking to the stakes

You cannot verify everything, and you do not need to. The question to ask of every answer is: what happens if this is wrong?

  • Low stakes. A synonym, a recipe idea, a first draft you will rewrite anyway. Read it critically and move on. If it is wrong, you lose a minute.
  • Medium stakes. An email to a client, a summary you will forward, a figure in a slide. Check the claims that carry weight. Skip the rest.
  • High stakes. Anything you will act on, publish, send to many people, or that touches money, health, law or someone's reputation. Verify every factual claim before it leaves your hands.

Most people check too little at the top of this list and too much at the bottom. Decide the level first, then pick the methods below that fit.

Open the sources

An earlier lesson covered why a model can produce a reference that looks real and is not. The remedy is mechanical: ask for sources, then open them. Not does a source exist but does this source say what the model says it says. Models summarize loosely. A paper that says in some conditions becomes studies show.

For each factual claim, give the source and quote the exact sentence it comes from.

If you pasted a document, point it there: Quote the passage from the document that supports this. If there is none, say so. If the tool gives links, click them. If the model cannot give you anything you can open, treat the claim as unverified, however confident it sounds.

Make it compute with code, not in its head

Models produce numbers the way they produce words: by predicting what looks right. For a sum, a percentage, a date difference or a unit conversion, looking right is not enough.

Many chat apps can write and run a small program (often labeled a code interpreter or an analysis feature). The program does the arithmetic properly. Ask for it directly:

Work this out using code. Show me the code and the result, and state any assumptions you made.

Then read the assumptions. Did it treat the dates the way you meant? Did it include tax? If the tool cannot run code, do the arithmetic in a spreadsheet or a calculator yourself. Never let a model do mental math on something that matters.

Ask for the reasoning and find the weak step

Ask it to show its work: Walk me through how you reached this, step by step. Then do not read the steps for fluency. Read them for the one step that does the real work. There is usually one: an assumption, a number pulled from nowhere, a therefore that moves faster than the evidence. Check that step and ignore the rest.

One caution. An explanation written after the answer is a reconstruction, not a recording of what happened inside the model. It can be a perfectly plausible story for a wrong answer. Use it to find what to check, never as proof that the answer is right. A useful follow-up: What would have to be true for this answer to be wrong?

Get a second opinion

Paste the same question into a different model, or run a plain web search. If two independent systems agree on a specific fact, your confidence should go up (though not to certainty; they learned from much of the same text). If they disagree, you have found exactly where to look.

Do not ask the same model are you sure? and accept yes. It will often simply agree with its own earlier words, or flip for no reason, and neither tells you anything. A better version: open a fresh chat, paste the answer, and ask it to find the errors as if a stranger had written it. Removing the model's own history from the picture makes it a far more honest critic.

Red flags, and whose name is on it

Some patterns should make you slow down on their own:

  • A very specific number with no source. A 37 percent increase sounds measured. Ask who measured it.
  • A tidy answer to a messy question. Real questions about law, health, money and people usually come with it depends. A clean answer may mean the mess was skipped.
  • A citation you cannot find in two minutes of searching.
  • A quotation attributed to a famous person. These are invented often.
  • An answer that agrees with everything you hoped.

And the last thing, which is not a technique. When you send, publish or act on an answer, the words become yours. The AI said so does not work with a client, a reader or a court. In 2023 a US federal judge fined two lawyers who had filed a brief citing court cases that did not exist; a chatbot had produced them, and nobody had opened them. The tool drafted. The people signed. That order of responsibility has not changed.

Try it

Ask any chatbot: Give me three statistics about how much time office workers spend in meetings, with a source for each. Now try to find each source. Does it exist? Does it contain that exact number? Give yourself two minutes per source and keep score. Then, in a fresh chat, paste one claim and its source and ask: Does this source support this claim exactly as stated? Quote the relevant sentence. Finally, test the arithmetic habit: What is 17.5 percent of 2,384, and what does that amount become after growing 4 percent a year for six years? Use code if you can and show it. Check the result with a calculator.

Key terms

Stakes
What happens if an answer turns out to be wrong; the thing that decides how much checking it deserves.
Code interpreter
A feature in some chat apps that lets the model write and run a small program, so arithmetic is calculated rather than guessed.
Weak step
The one step in a chain of reasoning that carries the real weight, and the one most worth checking.
Second opinion
Asking a different model, or a search engine, the same question to see whether independent sources agree.
Red flag
A pattern in an answer, such as a precise number with no source, that signals it needs checking before use.

Check yourself

1.Which of these most needs verifying before you use it?
2.You ask the model whether it is sure, and it says yes. What have you learned?
3.The reply gives a precise figure, 41.3 percent, with no source. What is the best move?

Scale your checking to what happens if the answer is wrong, and remember that the moment you send it, the answer is yours.