You will apply the checking routine to a real AI answer on a Singapore topic and record every error you find.
You now know why assistants invent things, where they are most likely to, and a routine for catching it. This exercise puts all three to work on one real answer, from the question to the last checked claim. It is slower than you will usually be, on purpose. Doing it thoroughly once shows you what kinds of errors turn up in your own topics and how long a proper check takes.
You need about thirty minutes, an assistant, a browser, and the annotation sheet below this lesson. Use the checking routine you wrote at the end of lesson 6.3.
Pick a Singapore topic where the official rules are public and where you, or someone you know, might actually need the answer. Good choices include CPF withdrawal rules, a question about a tenancy, sick leave or notice periods under employment rules, or eligibility for a government grant or payout.
Ask a detailed question, the kind that needs a paragraph or two to answer properly. "What are the CPF rules?" is too broad. "I'm 54 and turn 55 next year. What can I withdraw from my CPF at 55, and what has to stay in?" is about right. Ask in a new chat, and note whether the assistant searched the web or answered from what it already knew. Copy the full answer into your sheet before you read it closely.
Go through the answer and number every factual claim: each figure, date, age, name, condition and rule. Split sentences that carry more than one claim. "You can withdraw from age 55, after setting aside your retirement sum" is two claims.
Then check each one against the official source, usually the CPF Board, the Ministry of Manpower, IRAS, HDB or the relevant law on Singapore Statutes Online. Record the page you used beside each claim. Do not accept a news article or a blog as the check, even a good one, unless you cannot find the official page, and if that happens, say so on the sheet.
For each claim, write what the official source says in your own words next to what the answer said. If they match exactly, mark it correct.
Every claim that is not fully correct goes into one of four types. Each type tells you something different about why the assistant went wrong.
Invented: the claim matches no version of the official source, old or new. A figure, scheme name or condition that seems to have been made up. Outdated: the claim was true at some point but the rule has since changed. This is common for anything set yearly, and lesson 4.4 showed why. Misquoted: the answer cites a source that exists but says something different from what the answer claims. Missing a condition: the claim is correct as far as it goes but leaves out an exception or qualification that changes how you would act.
The fourth type is easy to underrate. An answer can be accurate sentence by sentence and still lead you wrong because it left out the part that applies to you.
Note the time when you start step 2 and when you finish step 3, and write the total on the sheet. Be honest, including the time you spent finding the right official page.
That number belongs in any estimate of what AI saves you. If the assistant wrote the answer in ten seconds and checking it took twenty minutes, the real cost of the answer is twenty minutes, and you can compare that with how long it would have taken to read the official page yourself.
Nadia, an HR assistant at a retail chain, ran this on a question a new part-timer had asked her: how many days of paid sick leave they would get, and after how long. Her results are an example of the pattern, and yours will differ.
The answer had eight claims. Four were correct. One, the length of service before sick leave applies, matched neither the current MOM page nor any older version she could find, so she marked it invented. One gave a detail as it stood in an older version of the guidance, which she marked outdated. One cited an MOM page that turned out to be about annual leave, not sick leave, so that was misquoted. The last was correct but did not mention that the entitlement for part-timers is worked out differently from full-timers, which she marked as missing a condition.
The check took her 24 minutes, most of it spent finding the right part of the MOM site. Her note at the bottom of the sheet said she would have sent the part-timer the wrong answer if she had trusted the reply, and that next time she would go to the official page first and use the assistant to explain it.
Your own sheet will look different depending on your topic and your assistant. Choose your question now, and resist the urge to read the official page first, so you are checking the answer rather than confirming what you already know.
Produce an annotated copy of the answer with each claim marked correct or wrong, the source used, the error type and the total checking time.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).