You will be able to verify the parts of an answer most likely to be wrong.
Ravi's monthly report to his manager had four parts an assistant had helped with: a summary of customer complaints with a percentage change, a quote from a food magazine review of one of the cafes, a link to a government page on food hygiene grading, and a spreadsheet formula to total weekly sales across the six outlets. Every part looked right. Two were not.
Lesson 7.1, Match the check to what is at stake, set out how hard to check. This lesson is about how, for the four kinds of content most likely to be wrong: numbers, quotes, links and code. Each has a quick, specific check that works better than reading the answer again and hoping.
Language models are not calculators. As AI fundamentals lesson 3.2, It writes by predicting the next token, again and again, explained, they predict text, and a figure in an answer is predicted like any other word. Many assistants can run code to do a calculation, which is far more reliable, but you often cannot tell whether that happened.
So recalculate any figure that matters yourself, on a calculator or in a spreadsheet. Ravi's summary said complaints "fell about 40 percent, from 25 last month to 18 this month". The drop is 7. Seven out of the starting 25 is 28 percent. The 40 percent figure came from dividing 7 by the new figure, 18, which gives about 39 percent, a classic slip. These figures are an example, but the mistake is a common one, and it would have overstated the improvement to his manager by more than ten percentage points.
If you cannot easily redo a calculation yourself, ask the assistant to show it step by step, with each figure it used and where it came from. A written-out calculation is much easier to check than a bare answer, and errors often become obvious once the steps are visible. Then check the steps, as lesson 3.4, When to use a reasoning mode, warned: a long trail is not proof.
Quotes are easy for a model to invent. A model asked for a quote that supports a point will often produce one that sounds exactly right, attributed to a real publication or person, that was never said.
The check is simple. Search for the quote word for word, using a search engine with the quote in quotation marks, or the search function in the original document if you have it. If you find it, read the surrounding text to make sure the meaning has not changed. If you cannot find it, treat the quote as invented and do not use it. A quote you cannot find is not "probably fine". It is a sentence written by a model and attributed to someone who may never have said it.
Ravi's magazine quote could not be found anywhere. The magazine had reviewed the cafe, and the review was positive, but the sentence in his report did not appear in it. He replaced it with a real line from the review.
When an answer cites a source or gives a link, open it. Confirm three things: the page exists, it is the page the answer says it is, and it says what the answer claims it says. Links fail in each of these ways. Some lead nowhere. Some lead to a real page on a different topic. Some lead to the right page, which says something different or more qualified than the answer suggested.
Ravi's link went to the right agency's website, but to a page about licensing, not grading. He found the right page himself through the agency's own site. When a source matters, it is often quicker to go to the official site and find the page directly than to trust a link from an assistant.
For a spreadsheet formula or a piece of code, reading it is not enough, even if you understand it. Run it on a small sample where you already know the right answer.
Ravi's formula was meant to total weekly sales across six outlets. He set up a tiny test: six rows with made-up sales of S$100 each, so the right total was S$600. The formula returned S$500. The range in it stopped one row short, so it missed the last outlet. On his real data, with hundreds of rows and uneven figures, he would never have spotted that by eye.
The principle is the same for any code: build a small case where you know the answer, run it, and compare. Include an awkward case or two, such as an empty cell or a zero, because that is where formulas tend to break.
None of these checks takes long. Recalculating a percentage takes a minute, searching a quote takes seconds, and testing a formula on six rows takes two minutes. What they need is the habit of doing them. To build it, find a recent answer from an assistant that contains at least one figure and one source, the kind of answer you might have passed on without a second look.
Take an answer containing figures and a source, recalculate the figures, open the source, and write down every discrepancy.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).