An agent can do the arithmetic perfectly on the wrong data. Imagine you ask for a summary of your team’s expenses in September, and the agent goes by the date each booking was made when the card statement goes by the date it was charged. Every sum is correct, and the total still doesn’t match the statement. Of the four questions from When the Result Is Wrong: Four Questions, this is a what was it given problem.
Now imagine you ask whether your team is keeping up with its weekly sales target, and the agent compares a whole month of sales against that weekly target. The figures are real, but the comparison is off, so the team looks as if it’s selling about four times its target. That’s a what was asked problem, because the agent answered a question nobody meant to ask.
If you checked a single number against a source you trust, you’d catch both mistakes.
Confident and Wrong
A wrong answer reads just as smoothly as a right one. Because a model writes what’s likely, it produces clear, sure-sounding text whether or not anything was checked, so a confident tone tells you nothing about whether the answer is correct. A reply can name a web page that doesn’t say what the reply claims, or report a total that was never added up, in the same calm voice it uses for everything else.
