You will be able to pick the right kind of AI research for a question and know what each kind can get wrong.
Before you put a research question to an AI assistant, choose which of the three kinds of research you want: plain recall, search or research mode. The kind you choose decides whether the answer is likely to be right. Most people don't realise there is a choice, and a polished answer can still be wrong.
Kavitha is in business development at a firm that runs corporate wellness programmes. Her director usually brings her questions on Friday afternoons. One week it is what competitors charge at the moment, another week it is the number of companies of the target size in Singapore, or whether workplace health rules have changed. She used to give up Monday morning to searches and phone calls to answer them. Now she types the question into an assistant and has a well-written reply within seconds. Whether that reply is accurate depends on the kind of research the assistant did.
To see the difference, she put one question about workplace health rules to the assistant in all three modes. In plain recall she got a tidy answer with no sources, and one detail in it later proved to be out of date. In search mode she got three links. Two were from a consultancy's blog. The third was from the Ministry of Manpower (MOM) website, which is the Singapore government's source for workplace rules. In research mode she got a long report citing more than twenty sources, among them the MOM page and several news articles. It included one figure that she couldn't find in any of those sources.
With plain recall, the assistant doesn't search anything. It writes its answer from patterns it picked up in training. AI fundamentals lesson 4.4, "Find the edges of what a model knows", explains how this works. Two of its limits matter when you are researching for work.
The first limit is the training cutoff date: the point where the model's training data ends. Anything that happened or changed after that date is invisible to the model, and it won't always mention this. Suppose a rule changed last year. If you ask about it, you may get a confident account of the old version. Kavitha's outdated detail is this kind of mistake.
The second limit is that you get no sources to check. The model may write something that looks like a reference, such as the title of a report or a statistic with a year next to it. It produces that text the same way it produces every other sentence, so none of it is a real source. AI fundamentals lesson 6.2, "Where invented answers are most likely", explains why invention clusters in exact figures, names and references.
Plain recall works well when you want a concept explained, ideas on what to look into, or a rough first list of questions. Don't rely on it for any fact you plan to pass on to your director.
Most assistants can search the web now. Some only search once you turn search on. Others decide for themselves when a question calls for it. With search, the assistant runs web searches, pulls up a number of pages, writes its reply from what they say, and usually gives you links. AI fundamentals lesson 5.3, "Retrieval: how an assistant looks things up", walks through how this works.
Search deals with both weaknesses of plain recall: the information can be up to date, and you have links to open. The reply can only be as good as the pages the search turned up, though. Say the top results were a two-year-old blog post and a vendor's marketing page. The summary would pass on the old claims and the sales pitch exactly as written. The model can also misread a page, blend two sources together, or put a link next to a claim the linked page never makes.
So before you rely on a search answer, look at the pages it used. In Kavitha's case, two of the three links pointed to a consultancy's blog. For a question about workplace rules, the MOM page is the one with official standing.
Many assistants also have a longer research mode, and each one calls it something different. A research mode draws up a list of questions, reads a large number of sources, and writes a structured report with citations. Expect it to take minutes, sometimes more, where the other modes take seconds.
It helps in two ways. You can get your bearings fast on a subject you don't know, and it often turns up sources you wouldn't have found on your own. The catch is that a long, well-presented report is more convincing to read, while its length and presentation tell you nothing about whether it is accurate. Don't take them as proof. A report with forty citations takes much more work to check than a paragraph with three. A mistake buried in a report that looks thorough is also harder to spot, since everything around it looks careful. Kavitha's report had more than twenty sources and still contained a figure none of them backed up.
Pick the mode based on what you need. Plain recall is for concepts, pointers and draft questions. Search is for current information you want links for. A research mode is for getting your bearings on an unfamiliar topic or finding more sources.
Whichever mode you use, you still need to check. Any of the three can misquote a source, present an outdated rule as current, or make up a detail that sounds believable. The difference between them is in how you check and how far you go.
Start by deciding what the answer is for, then check to match. If Kavitha only wants a rough idea of what wellness programmes usually include before a brainstorm, it's enough to skim a search answer and see whether it seems plausible. If her director is going to quote a figure to a client, she has to follow every number back to the page it came from and confirm it there. Lesson 4.2 sets out how to do that.
Try this with a question of your own. Pick something about a Singapore market or regulation that matters in your job, and ask it in each mode your assistant offers so you can compare where the answers came from.
Ask one question about a Singapore market or regulation three ways: plain, with search, and in research mode if you have it. Compare the sources each one used.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).