Figures the model makes up

A large language model works by predicting the next word. To it, numbers are words too. When you ask it to "add up this column", the model does not run a calculator the way a spreadsheet does; it writes the figure most likely to appear in that position. For short sums it is often right. For long tables, percentages of percentages, or numbers with many digits, it can get a digit wrong, swap the order, or round silently. How this works is explained in the text prediction machine lesson.

There is a mistake more dangerous than a wrong sum: an invented figure. This is one form of hallucination. Ask a model to write monthly report commentary without giving it the figures, and it may write "revenue rose 8.4% on last month" with great confidence. That figure comes from no data at all. It appears because report commentary usually contains percentages like that. The same happens when a model is asked to "fill in" a table with gaps, estimate a balance, or state a rate.

In finance work, a wrong figure rarely looks wrong. It is written neatly, with correct thousands separators and two decimals. People read financial reports with trust because the figures there have usually been checked. That is why figures made by a model must be treated differently from figures that come out of the accounting system.

The answer is a figure verification routine you run every time, not only when suspicious. Its four steps:

• Recompute by another method. Total it with a spreadsheet formula, a calculator, or from the accounting system's report. Do not ask the same model to recompute; the same method repeats the same mistake.
• Cross-foot. The total of all rows must equal the total of all columns. Totals by branch and totals by product must arrive at the same figure. If they differ, a row was missed or counted twice.
• Compare with the source. Every figure in a report must have an origin you can point to: a system report, a bank statement, an invoice or a working file. A figure with no origin is removed or flagged.
• Test reasonableness. Compare with last month and with a rough estimate. An electricity cost that suddenly rises tenfold may be right, but it needs an explanation.

How to recompute and test spreadsheet formulas is taught in the checking figures and spreadsheets lesson. In this track you apply it to finance work: reconciliations, reports, invoices.

One more habit, the simplest: flag. When you receive a draft from AI, such as report commentary, an invoice summary or an explanation of differences, colour or mark every figure in it. The mark comes off only after the figure is matched with its source. A draft that still has flagged figures is not sent to anyone. Better still, ask the model from the start to write a marker such as ‹figure› instead of a number, then fill in the figures yourself from the system. That way no invented figure can slip through, because every figure in the final document passes through your hands and has a source you can point to when someone asks.