Yesterday I gave ChatGPT in an anonymous browser window (not logged in) two columns of TAB separated numbers, about 40 rows. I asked it to give me the weighted average of the numbers in the second column, using the first one (which were integer, "quantity", numbers) as the weight.

It retuned formulas and executed them and presented a final result. It looked good.

Too bad Excel and then Claude, that I decided to ask too, had a different result. 3.4-something vs. 3.8-something.

ChatGPT, when asked:

> You are absolutely right to question it — and thank you for providing the intermediate totals. My previous calculation was incorrect. I mis-summed the data. With a dataset this long, a manual aggregation can easily go wrong.

(Less than 40 small integer values is "this long"? Why did you not tell me?)

and

> Why my earlier result was wrong

> I incorrectly summed:

> The weights (reported 487 instead of 580)

> The weighted products (reported 1801.16 instead of 1977.83)

> That propagated into the wrong final value.

Now, if they implemented restrictions because math wastes too many resources when doing it via AI I would understand.

BUT, there was zero indication! It presented the result as final and correct.

That has happened to me quite a few times, results being presented as final and correct, and then I find they are wrong and only then does the AI "admit" it use da heuristic.

On the other hand, I still let it produce a complicated Excel formula involving lookups and averaging over three columns. That part works perfectly, as always. So it's not like I'll stop using the AI, but somethings work well, others will fail - WITHOUT WARNING OR INDICATION, and that is the worst part.