You will be able to run a short, repeatable check on any AI answer before you use it.
Wei Jie does freelance design work on the side and wonders whether he needs to register for GST. He asks an assistant and gets a clear, friendly answer of about two hundred words. It mentions a turnover figure, a time period and a rule about when registration becomes compulsory, with a link to an IRAS page at the bottom, and it reads so smoothly that he is about to forward it to his accountant friend with "looks like I'm fine".
At this point most people either trust the answer because it sounds right, or feel vaguely uneasy and forward it anyway. What Wei Jie needs instead is a short routine he can run every time, one that takes a few minutes and catches the errors that would cost him. It has four moves, and none of them needs special software.
Go through the answer and underline every claim that could be shown right or wrong. In practice that means numbers, names, dates, quotes and rules. "You must register if your turnover exceeds" is a rule with a number attached. "This applies from the start of the following quarter" is a date rule. "IRAS recommends" is a claim about what a named body says.
Leave the general explanation alone for now. Lesson 6.2 showed that the specific details are where invented content is most likely, and that is where your time goes. Wei Jie's answer had five underlined claims. Most answers of that length have between three and ten.
This step alone changes how you read. A paragraph that felt like one solid answer turns out to be a handful of separate claims, each of which could be right or wrong on its own.
You do not need to check every underlined claim with the same care. Ask which ones you will act on, and which would cost you if they were wrong. For Wei Jie, the turnover figure and the rule about when registration becomes compulsory matter most. A wrong detail about how to fill in the form matters less, because he will see the real form anyway.
Check the ones that matter against a primary source: the organisation that makes the rule, or the original document itself. In Singapore that usually means the agency's own website, such as IRAS, the CPF Board or the Ministry of Manpower, or the text of the law on Singapore Statutes Online. For a figure from a study, go to the study itself rather than an article about it. For a company's results, it means the company's own report.
A news article, a blog, a forum post or another AI answer is a secondary source. It might be right, but it is one more step away from the truth, and it may be out of date.
If the answer cites sources or gives links, open every one. Look for three things. First, does the page exist at all? Models can produce URLs that look exactly like a real agency address but lead nowhere. Second, is it the right kind of page: official, current, and about Singapore if your question was? Third, does it actually say what the answer claims? Use your browser's find function to search the page for the figure or phrase in question.
Wei Jie's link went to a real IRAS page. It covered GST registration, but the turnover rule it described had a condition the assistant had left out. That is the error a quick glance at a citation would have missed.
The last move costs ten seconds. Ask the assistant: "Which parts of your answer are you least sure about, and why?" It will often point to something useful, such as a figure that changes yearly or a rule with exceptions.
Treat that reply as a lead to follow up. The model's view of its own uncertainty is produced the same way as everything else it writes, by prediction, as lesson 6.1 explained. It can flag a part that was fine and stay quiet about the part that was invented. So use its answer to add claims to your checking list, never to remove them.
For a low-stakes request, such as rewording an email, you can skip most of this. For anything you will act on, sign, send to a client or pass to someone who trusts you, the routine is usually quicker than fixing the mistake later. The two New York lawyers in lesson 6.1 skipped the step of opening their sources, and that was the whole difference.
The routine only works if you use it without having to think about it. The four moves above are one version. Yours should be in your own words and short enough to remember, written where you will actually see it the next time an answer looks good enough to forward.
Write the routine as a five-line checklist in your own words and save it where you will see it when you next use an assistant.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).