Where AI Gets Numbers Wrong

Everything in this course works only if you can catch the errors. Analysis errors are the worst kind to miss, because the output is a number and numbers are believed — a wrong paragraph gets argued with, a wrong figure gets put in a slide.

Arithmetic without code is unreliable

A model producing a number in prose is predicting text, not calculating. It is usually right on simple sums and unreliable on long ones, on percentage changes, and on anything requiring several chained steps.

The fix is complete and simple: make it write and run code for every number. Then the arithmetic is done by a machine that does arithmetic, and you can read the code.

Correlation described as cause

Ask why something rose and you will get a fluent causal story, because that is what the question invited. The data almost never supports it.

• 'Revenue rose after the campaign' — and December, and the new product, and a competitor's outage
• Selection effects: customers who use the feature buy more, because customers who buy more use the feature
• Survivor bias: analysing only current customers tells you nothing about why the others left
• Regression to the mean: an unusually bad month is usually followed by a better one, with or without your intervention

The habit that catches all four: for any explanation offered, ask for three alternative explanations and what data would distinguish between them.

Invented context

Ask for an industry benchmark and you will often receive one — a specific percentage, attributed to a plausible-sounding source, that does not exist. This is the most damaging failure in this whole course, because a fabricated benchmark in a board deck looks exactly like a real one.

Any external figure needs a source you opened yourself. If a benchmark cannot be traced to a document you can read, it does not go in the deck.

The checks worth running every time

1. Sanity-check magnitude. Does the total roughly match what you already know about the business? Being off by a factor of ten is common and obvious once looked for
2. Verify the row count used matches the row count you expect
3. Recalculate one number by hand or in a spreadsheet. One is enough to catch a systematic error
4. Check the denominator on every percentage. Percentage of what, over which period?
5. Ask for the same analysis in a fresh conversation and see whether the number matches
6. Look at the group sizes behind any dramatic finding

Confidence is not calibrated

The tone of an answer carries no information about its reliability. A model states a correct total and an invented benchmark in exactly the same register, and there is no hedging to look for.

Asking 'are you sure?' is not a check — it usually produces either agreement or a reversal, depending on how the question was phrased, and neither tells you anything about the underlying number.

Where this leaves you

Used with code, checked row counts and traced sources, these tools do in twenty minutes what used to take a day, and they make analysis available to people who were never going to learn pandas.

The judgement that decides whether the analysis is honest — what the question is, what missing data means, whether a difference matters — was never the part that took the time. It is still yours, and it is the part that makes the numbers worth anything.