You will be able to predict the three most common failure types in a money answer before you read it.
After the Sunday he wasted, Darren tried again with a narrower question. He asked an assistant what share of young Singaporeans have an emergency fund, because he wanted to know if he was behind. The answer gave a percentage, named a survey by a local bank with a year attached, and added a link. It read like a newspaper paragraph. The link led to a page that didn't exist, and when he searched for the survey title, nothing with that name turned up.
He hadn't done anything wrong. He'd met the most common failure in money answers, and once you know the four types, you can expect them before you read a word.
The first type is a figure that looks specific and comes from nowhere. A fund's expense ratio given to two decimal places. A savings account rate. An average return over ten years. The precision is what makes it convincing, and the precision means nothing.
Recall from lesson 1.1, What an assistant is actually doing when you ask about money, that the model writes by predicting what usually comes next. When the next thing in a sentence is a number, it predicts a number that fits: the right size, the right format, in the range such figures usually fall. Fitting is not the same as being true. A made-up fee of 0.45% looks exactly like a real one.
Expect invented numbers whenever you ask for a fee, a rate, a return, a premium or any figure about one particular product. The more specific the product, the less text the model has seen about it, and the more it fills the gap.
The second type is a figure that was true once. Every model is trained on text up to a cutoff date, and money rules in Singapore keep moving after it. CPF interest rates, contribution caps, income tax reliefs, Singapore Savings Bonds rates and housing grant amounts all change on schedules set by the agencies that run them.
So an assistant can tell you a rule that was correct, quote it accurately from its training text, and still be wrong for you today. This is harder to spot than an invented number, because if you search, you may find the same figure in an old article and think you've confirmed it.
Some assistants can search the web, which helps, but it doesn't solve the problem. A search can land on a blog post from three years ago as easily as on the official page. Module 6 covers this in detail. For now, the habit is simple: any rate, cap, limit or threshold is a look-it-up item, and you look it up on the site of whoever sets it.
The third type comes from what the model has read most. There is far more text on the internet about American personal finance than Singaporean, so answers drift towards US rules unless you stop them. Ask about saving for retirement and you may get 401(k) plans and Roth IRAs. Ask about improving a credit score and you may get advice built on how US scores work.
The fix starts with saying where you are. "I live in Singapore" at the top of a question changes a lot. It doesn't change everything, because the model can still mix in a foreign rule it has seen thousands of times. Lesson 6.2, Answers that quietly assume you live in the US, gives you a checklist for catching what slips through.
The fourth type is the one Darren hit, and it makes the other three worse. When you ask where a claim comes from, the assistant may produce a report title, an author, a publisher, a year and a link. Some will be real. Some will be stitched together from real-sounding parts: a genuine organisation, a plausible title, a date that fits, and a web address in the right format that leads nowhere.
This happens for the same reason as invented numbers. A citation is text with a familiar shape, and the model is good at producing familiar shapes. That's why asking "are you sure?" rarely helps. The assistant may apologise and give you a different source that is just as made up, or insist on the first one.
The only test is to open it. Click the link. Search for the exact title in quotation marks. If it's a government source, go to the agency's own site and look for it there. A source you haven't opened is not a source yet.
Put the four together and you get a quick way to read any money answer. As you go, sort each sentence into one of two piles. Explanations of how something works go in the first pile, and you can mostly judge them by whether they make sense. Specific figures, dates, rules and sources go in the second pile, and every one of them is unconfirmed until you have checked it somewhere else.
Darren's emergency fund answer had a sensible paragraph on why such a fund matters. That part was fine. The percentage, the survey and the link were all in the second pile, and none of them survived a check.
Try the same thing with a statistic of your own. Ask for a money figure, then ask where it comes from, and see what happens when you go looking for each source it gives you.
Ask an assistant for the source of a money statistic, open every link or title it gives, and record which ones exist and say what it claimed.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).