You will be able to rank requests by hallucination risk and adjust how much you check.
Two requests land on your desk on the same morning. A colleague wants a friendlier version of a reminder email about expense claims. Your manager wants to know which clause of a supplier contract allows early termination, and the exact notice period. You could hand both to an assistant in a minute. If you checked them with the same care, you would either waste time on the email or take a real risk on the contract.
Lesson 6.1 explained why hallucination is built into how a model writes. The good news is that the risk is not spread evenly. It rises and falls in fairly predictable ways, and once you know the pattern you can spend your checking time where it counts.
The first factor is how specific the answer has to be. Rewording an email has no single right answer, so there is little to get wrong. A request for an exact figure, date, name, quote, URL, clause number or reference has exactly one right answer, and many plausible wrong ones that look just like it.
Lesson 3.2 showed that the model picks likely tokens. For a general explanation, many sequences are reasonable. For "the notice period in clause 14.2 is", the right number competes with every other number that commonly follows those words. When the model has not seen the real answer often, a wrong one can easily win, and it will be written with the same confidence.
So treat precise details as the parts of an answer most likely to be invented, even when the surrounding explanation is sound. A summary of how CPF housing withdrawals work may be broadly right while the specific limit it quotes is wrong.
The second factor is how often the topic appeared in the text the model learned from. Lesson 2.3 made the point that a model does badly on what was rare in its data. Famous facts appear thousands of times in training text, while a detail about a small Singapore scheme, a local company, a town council's rules or a niche part of your industry may have appeared a handful of times, or never.
Many Singapore specifics fall into this group. The model has seen countless explanations of compound interest, but far fewer mentions of the eligibility conditions for a particular local grant or the name of whoever heads a mid-sized Singapore firm. When it has little to go on, it rarely says so, and instead fills the gap with something shaped like the answer.
Recent topics add a further problem. Lesson 4.4 showed that anything after the training cutoff is unknown to the model unless it searches, and policies in Singapore change every year at Budget time and in between.
The third factor is how you ask. Lesson 4.3 described the agreeable streak that human feedback leaves behind. A question that carries a premise, such as "Why did the company drop its four-day week policy?", invites the model to explain the premise rather than question it. If the company never had such a policy, you may still get a fluent explanation of why it was dropped.
The same happens with "Which section of the act says I can do this?" when no section says so, or "Summarise the three main findings of this study" when the study had two. Leading questions are often written without noticing, because you ask about what you already believe. Asking neutrally, for example "Did the company ever have a four-day week policy, and if so, what happened to it?", gives the model room to say no.
The risk falls when the answer is grounded in a source you can see. If you paste in the contract and ask about the termination clause, or the assistant searches and cites the official page, it is drawing on text in its context window, as lesson 5.3 explained, and invented details become less likely.
Less likely is the right phrase. The model can still misread a clause, mix up two sections or cite a page that does not say what it claims, and lesson 5.4 showed what happens with long files. A grounded answer is lower risk, and it is also much quicker to check, because the source is one click away.
Put the factors together and you can rank any request. Here is one example, from lowest risk to highest:
Rewriting a reminder email in a friendlier tone, where nothing has to be exactly right. A summary of a two-page memo you pasted in, which is grounded and short enough to check in a minute. A general explanation of how a fixed deposit differs from a savings account, a common topic with no exact figures. The notice period in a pasted contract, quoting the clause, which is grounded but hinges on one exact number. The current eligibility age for a specific Singapore grant with search off, which is specific, local and possibly recent all at once.
The first needs a quick read. The last needs checking against the agency's own website before anyone acts on it. Your own work will have its own versions of each, so start noticing which factor applies the next time you open an assistant.
Write five questions you might ask at work and rank them from lowest to highest hallucination risk, noting which factor drives each ranking.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).