AssociationAI / AI Literacy
AssociationAI team Published

Article

Beginner

Why AI Tools Get Numbers, Counts and Arithmetic Wrong in the Summaries You Reuse

An AI summary of a dues report or salary table mixes copied, rounded, calculated and invented numbers, all in the same confident voice. A study of two 2023 chat models shows why the arithmetic deserves the closest look.

AI Literacy Data Finance Board Reporting

The finance committee meets Thursday, and someone pastes the quarterly dues report into an AI tool and asks for three sentences for the board memo. The summary comes back clean, with a member count, a revenue figure and a total. Every one of those numbers reads as fact, and nothing in the wording tells you which ones were copied from the report, which were rounded, which the tool worked out on its own, and which came from nowhere.

That is the problem with numbers in an AI summary. They arrive in one voice. A sentence with a correct figure and a sentence with a wrong one look identical, and a board memo built from them carries both.

Why a number in a summary is not a lookup

It helps to be clear about what the tool is doing. A chat tool writing a summary is producing text, and a number in that text is produced the same way as the words around it. In our reading, that is the root of the trouble: the tool is not reaching into your table and copying a cell unless something in the setup makes it do so. Sometimes it gets the figure right because the figure is right there in what you pasted. Sometimes it rewrites the figure, rounds it, or combines two figures into a new one, and it does all of this without saying so.

The shortcut most people take is to read the summary and ask whether the numbers seem about right. That check fails in exactly the cases that matter. A total that is off by a few hundred members, or a revenue figure with two digits swapped, still seems about right. The numbers that look wrong are the easy ones; the dangerous ones look plausible.

What a careful study of chat models found

The most detailed evidence we found comes from mathematicians who tested chat models directly. Frieder and colleagues, in Mathematical Capabilities of ChatGPT, published in the NeurIPS 2023 Datasets and Benchmarks track, tested the January 9 and January 30, 2023 releases of ChatGPT, along with GPT-4, on sets of mathematics questions graded by mathematicians. The findings are about those 2023 models; newer tools were not tested, and they may behave differently.

Their finding on arithmetic is the part that matters for a dues report. Of the January 2023 ChatGPT, they wrote that “the ability of ChatGPT to execute algebraic manipulations is surprisingly inconsistent. In some cases, ChatGPT executes complicated symbolic tasks with ease; in other cases, it fails on simple arithmetic operations or rearranging terms.” They added that “the mistakes do not seem to correlate with the complexity of the algebraic expression.”

Read that from the board memo side. A tool that handled a hard step well can still drop a simple sum, and you cannot predict which step will go wrong from how hard it looks. Getting the difficult part right is no evidence about the easy part.

The better model did not remove the problem. The same study reports that “despite its better overall performance, GPT-4 still seems to be vulnerable to mistakes in seemingly simple calculations.” Their conclusion was that the models tested were “not yet ready to deliver high-quality proofs or calculations consistently.”

One more observation from the paper explains why a single slip spreads. When the January 2023 ChatGPT made an algebraic mistake, the authors found that “it mostly carries over this mistake reliably to the rest of the computation.” In a summary, that looks like a wrong subtotal feeding a wrong total and then a wrong comparison, each one consistent with the last and all of them wrong.

We would not turn this into “AI cannot do math.” The study is narrower than that, and the authors also found the answers could be useful for some tasks. What it supports is a working rule: arithmetic done inside a chat tool needs to be redone by a person before anyone repeats it.

An example of the five ways a number slips

This is a teaching example, not a case study. The organization and every figure below are invented to show the pattern.

Suppose the dues report for an invented Lakeshore Society of Planners lists 1,846 renewing members and 312 new members, with dues revenue of $428,750 across two membership tiers. The report gives no overall total. An AI summary of that report might say:

“Lakeshore renewed 1,864 members and added 312 new joiners, for 2,176 members in all, bringing in about $430,000 in dues across three membership tiers.”

Every number in that summary is doing something different. The 312 is copied exactly. The 1,864 is the renewal count with two digits swapped, so it is simply wrong, and nothing in the sentence warns you. The 2,176 was never in the report: the tool added its own wrong renewal count to the new members, so the sum is correct arithmetic built on a bad input, the carried mistake the study describes. The “about $430,000” is a rounded version of $428,750, and the word about is the only signal. The “three membership tiers” is a count, and the report names two.

Those are the five slips we look for: copied wrong, rounded, derived, counted, and a changed unit or definition. The taxonomy is ours, not the study’s, but each type follows from what the study describes. Copying can go wrong because the tool is producing text, not lifting a cell. Rounding happens because round numbers read naturally. Derived numbers, such as totals, averages and changes from last quarter, may never have been in the source at all, which makes them the ones most exposed to the arithmetic errors the study found. Counts slip because “three tiers” is a number the reader rarely thinks to verify. And units drift when an hourly figure becomes a yearly one or a share of one group turns into a share of another, with the digits left untouched.

What to do before a number goes in the memo

The fix is not to stop using summaries. It is to stop treating their numbers as finished. Before any figure leaves the summary, find the cell it came from. If it matches exactly, it can stay. If it is rounded, decide whether rounding is acceptable and say about. If it is a total, an average or a change, recompute it yourself from the source cells and note how you did it, because that is the arithmetic the study found least reliable. If you cannot find it at all, take it out. Then check the counts and the units, and carry the source’s own footnotes along with the figures so the definition travels with the number.

Our tutorial on rechecking every number in an AI summary against the source table walks through that process step by step, with a working lab built on a real federal salary table.

The time this takes is small compared with the cost of a wrong figure in front of a board. And the staff member who can point to the cell behind every number in a report is the one leaders come back to when the next set of numbers matters.

Sources (1)