You will be able to explain why an assistant can describe a calculation correctly and still get the answer wrong.
Joanne asked an assistant a simple question: if she puts S$300 a month into an account paying 4% a year, compounded monthly, how much will she have after twenty years? The answer explained compound interest clearly and correctly, mentioned that regular saving adds up, and gave a figure of about S$157,800. It even showed that she would have put in S$72,000 of her own money.
The explanation was right. The figure was wrong by nearly S$48,000.
You already know the reason from lesson 1.1, What an assistant is actually doing when you ask about money. A language model writes digits the same way it writes words, by predicting what should come next. Unless it actually runs code to do the sum, the number at the end of a paragraph is a prediction of what a number in that spot usually looks like.
The explanation and the figure come from different kinds of skill. Describing how compounding works is a language task, and the model has seen thousands of good explanations. Producing the exact result of 240 monthly deposits, each growing for a different length of time, is arithmetic, and prediction is a poor substitute for it.
Here are the figures for Joanne's question, which are an example. Using the FV function in a spreadsheet, with a monthly rate of 4% divided by 12, 240 months and a payment of S$300, the result is S$110,032.39. That assumes each deposit goes in at the end of the month; if they go in at the start, it's S$110,399.16. Neither is close to S$157,800.
The assistant's figure is what you get if all S$72,000 had been deposited on day one and left to grow for the full twenty years. But the first S$300 grows for twenty years and the last S$300 for a month. On average, the money is invested for about half the time. The assistant described the right method and then calculated a different one.
Simple sums usually survive. Ask for 15% of S$2,000 and you'll almost always get S$300. The trouble starts when a calculation has several steps, each depending on the last.
Compound growth is one: each period's interest depends on the balance after all the previous periods. Loan repayment schedules, or amortisation, are another: each month's interest depends on how much is still owed, which depends on every earlier payment. So are comparisons with fees taken off each year, inflation adjustments over many years, and anything involving a monthly rate derived from a yearly one.
In a multi-step sum, a small slip early carries through to the end. Each predicted step can be slightly off, and the model has no way to notice, because it isn't checking its working. It's writing the next plausible line.
Some assistants can write and run code, and show you both the code and its output. That changes things. A figure produced by actually running a calculation is far more reliable than one predicted in prose, because the computer is doing the arithmetic.
The difficulty is telling which happened. Some assistants show a visible block of code and a result. Some do it behind the scenes and only mention it. Some describe a calculation in prose that looks like working but was never run. As Working with AI assistants lesson 7.3, Checking numbers, quotes, sources and code, puts it, you often can't tell.
So the rule is simple. A figure with no visible calculation behind it is a guess. Treat it the way you'd treat a number scribbled by a stranger on a napkin: possibly right, not usable until checked.
The fix isn't to stop using assistants for money calculations. It's to change where the calculation happens. Instead of asking for the answer, ask for the formula, put it in a spreadsheet, and run it yourself. Every input sits in a cell you can see. Every step can be checked. When something looks wrong, you can change one input and watch what moves.
How money works lesson 2.4, Build a compounding table in a spreadsheet, showed this for a lump sum. Lesson 5.2, Ask for the formula, then run it yourself, extends it to loans and regular savings. Lesson 5.3, Sanity-check results with a rough estimate, adds a quick check that would have caught Joanne's figure in ten seconds: S$72,000 invested for about ten years on average, at 4%, lands somewhere around S$105,000 to S$110,000, nowhere near S$158,000.
Pick a savings question of your own, with your own amount, rate and number of years, and keep a blank spreadsheet open beside the chat before you ask it.
Ask an assistant a compound interest question in prose, then compute it yourself in a spreadsheet and compare the two answers.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).