Tables and figures: a minefield

This is where trust is most dangerous, because the result has every appearance of being solid.

Why the calculation remains fragile

A language model predicts text. A number is just text to it. It has seen so much arithmetic that it reproduces most of it correctly, but it does not calculate in the way a calculator does: it produces the most plausible result.

For a simple addition, it works. For a thirty-line sum, a compound percentage, a weighted average, errors become likely, and they are presented exactly like everything else.

What has changed, and what hasn't

Many tools now execute code to actually calculate. When you see an analysis note or a code snippet appear, the calculation is done by a program and becomes reliable.

But reading the table remains a weak point: misaligned columns, merged cells, two-line headers, thousands separators. The error happens before the calculation, and a correct calculation on misread numbers gives a wrong result with a nice decimal point.

The protocol

1. Explicitly ask it to calculate with code, not from memory.
2. First have it copy out the table as it understood it, and check two or three rows.
3. Re-add one random total yourself.

Step 2 is the one people skip, and it's the one that catches the most errors.