Recompute every number

Paste a table, ask what it shows, and you will get a fluent paragraph: the total, the average, which region grew, where the outlier is.

Some of those numbers will be right. Some will not. And nothing in how they are presented distinguishes the two — which makes this the single most dangerous output in the module.

Why prose arithmetic fails

E1 covered the mechanism: a model produces likely continuations. When it writes “total revenue was £847,200”, it is generating text that fits the pattern of a total, not performing addition and reporting the result.

Sometimes that lands exactly. Sometimes it is close — which is worse than being wildly wrong, because a number that is nearly right survives a sanity check.

Some tools now run real code to compute results, and where they do, this is genuinely better. But you have to know that happened for your specific answer, and the prose reads identically either way. If you cannot see the computation, assume there wasn’t one.

Method from the assistant, arithmetic from the sheet

The safe division:

The assistant — which comparison answers my question, what would make this misleading, what is the formula.

The spreadsheet — the actual number.

That keeps the model where it is strong, which is knowing how to analyse something, and keeps computation where results are reproducible and auditable. It also means the number in your document traces to a cell somebody can open, which is what anyone challenging it will ask for.

Ask about shape, not values

Framed well, this is genuinely useful:

Below are the column headers and 10 of 400 rows. Do not calculate anything.

What comparisons would answer “is the North region underperforming”? For each, give the formula and say what would make it misleading.

Do not calculate anything is the key instruction. You are asking for analytical judgement — which comparison, which confound, which denominator — and that is a question about method, where a model’s breadth genuinely helps.

Then you compute it, and you have both the reasoning and a number you can defend.

Percentages and the base rate

One specific trap. Percentage changes are where confident errors cluster, because they need a base you may not have stated.

“Up 40%” against what — last month, last year, budget, the same month last year? A model asked for growth will pick something reasonable-sounding and not tell you which. Always specify the base yourself, and always state it in the output. A percentage without a stated base is not a fact.

Try it now (6 minutes)

Take a table you have and ask for a summary of what it shows, arithmetic included.

Then check every number against the sheet. Note not just whether each is right, but whether a wrong one would have caught your eye. That second answer is why this lesson exists.

Check your understanding

1. Why is a total produced in prose unreliable? Models round aggressively It is generated as likely text rather than computed Tables lose precision when pasted

2. The safe division of labour with data is: Assistant for the method, spreadsheet for the arithmetic Assistant for both, spreadsheet as a backup Spreadsheet for the method, assistant for the arithmetic

3. A percentage change quoted without a stated base is: Fine if the direction is right Not a fact — the base determines what it means Acceptable in summaries but not in reports

Recap

Numbers in prose are generated, not computed, and a nearly-right figure is more dangerous than an absurd one. Ask the assistant which comparison and which formula, do the arithmetic in the sheet, and never quote a percentage without stating its base.