You will be able to verify citations and decide whether a source is good enough to rely on.
Open an AI answer you've saved that comes with sources, choose a sentence that ends in a citation, and follow that citation to the page it names. Then try to find the sentence's point written on that page. The citation tells you which source the assistant claims to have used. Whether the source actually contains the claim is something you can only confirm by reading it.
Kavitha came close to finding this out the hard way. In lesson 4.1 she produced a research-mode report, and one sentence in it pleased her a great deal: it gave the number of companies of her target size operating in Singapore, followed by a tidy citation. She was about to copy it onto a slide for her director when she clicked the link instead. The page was on SingStat, but it covered a different year and counted businesses in a different way. Her figure did not appear anywhere on it.
For each claim you plan to use, do three things: go to the source it cites, locate the passage that backs it up, and read that passage together with the text around it.
The first of these is the one most often left out. A link sitting after a sentence gives the impression the checking has been done, but you have no evidence until the page is in front of you. Some links are dead. Some land on an organisation's home page instead of the article that matters. Others reach a page on the right subject that makes a different point from the one in the answer.
When the page loads, use your browser's find function and type a word or number from the claim that is unlikely to appear by chance. If the answer mentions S$2.4 million, for example (an invented figure, as are the two that follow), search for "2.4". When the search turns up nothing, treat the claim as unsupported by that source, no matter how closely the page matches the topic.
If the search does find it, read the whole paragraph it sits in, because small errors often show up in that wider text. The source might apply the number only to a single sector, a single year or a single kind of company, and the AI answer might leave that limit out. A figure given as "up to 30 per cent" in the original can appear as "30 per cent" in the summary. A survey of 200 firms can be reported as if it described "Singapore firms" as a whole.
Sources differ in how much weight they can bear. A primary source is the body that creates the information in the first place. On Singapore work topics this usually means a government agency, or, for a company's results, its own annual report and announcements. Employment rules come from the Ministry of Manpower (MOM), tax from IRAS, official statistics from SingStat and company filings from ACRA.
A secondary source passes on what primary sources say. Blogs, news summaries, articles by consultancies and aggregator sites fall into this group, and a few of them are very good. Even so, every retelling gives a number another chance to be rounded off, a condition another chance to fall away and a rule another chance to be misunderstood. When the AI answer points to a blog explaining an MOM rule, open MOM's own page and confirm the rule there. Where the primary source is available to you, rely on it and put it in your citation in place of the blog.
With search answers you'll often have to do this tracing yourself. Search engines order results on many factors, so the agency's page may sit well below other sites in the list.
A page can be correct about the past and still mislead you today. Rules get changed, schemes get revised and many figures are refreshed each year. A tidy, readable article written several years back may set out a requirement that no longer applies.
Look for the published or last-updated date on each source. On government sites it's often near the foot of the page. Once you have it, consider whether things are likely to have moved since. Treat rates, thresholds, limits and anything linked to an annual budget or review as possibly changed, and go to the official site for the current version.
After checking a claim, label it with one of these:
Verified: the source states it, read in context, and the source is recent enough and reliable enough. Wrong: the source states something different. Outdated: the source backs it up but is too old, or a newer source has replaced it. Unsupported: the claim isn't in the cited source, or the source doesn't exist.
Any claim you couldn't trace comes out of your draft, even if it sounds right. Plausible claims are often the very ones you were hoping to use, which makes cutting them difficult. If you can't show where a claim came from, you have nothing to say when someone asks about it.
Kavitha later worked through her report one claim after another. Among the first eight, four came out verified, one wrong, one outdated and two unsupported. The company count she had almost put on the slide was one of the two unsupported claims, and she took it out. She then located the actual number on SingStat, together with its definition and the year it applied to.
Now pick a sourced answer of your own containing at least five claims. For each one, open the cited link, search the page for a distinctive word or number, read the paragraph around it, check whether the source is primary and find its date. Then mark the claim Verified, Wrong, Outdated or Unsupported, and delete any claim you could not trace.
Take a sourced AI answer with at least five claims, check each against its citation, and mark it verified, wrong, outdated or unsupported.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).