Set a checking level for any task by asking who will see the output and what an error would cost.
If you use an AI assistant often, you have to decide how much to trust each answer it gives you. If you check every answer as carefully as you can, you will have no time left for your other work. If you check none of them, sooner or later a mistake will reach someone else and embarrass you in public. The fix is to decide how hard to check before you read the output, and to base that decision on what is at stake. This module is about checking, and it starts with that decision.
Ravi does marketing for a group of six cafes. On a normal Tuesday he uses an assistant about a dozen times (also an example figure). He asks it to brainstorm names for a new seasonal drink, tidy up his notes from a supplier meeting, draft an Instagram caption, summarise customer reviews for his manager, and check the rules on advertising a promotion as "free".
Ravi had to make the same trade-off. If he checked everything thoroughly, he had no time for anything else. If he checked nothing, he risked a public mistake with his name on it. He solved it by giving each request one of three levels of checking. Each level includes everything in the level below it.
A light read is for drafts only you will see: brainstorming, personal notes, and first ideas you plan to develop further. Read the output to see whether it is useful and whether anything is obviously wrong, then move on. A light read is enough here because the output is never the final step. Any error gets caught later, before it reaches anyone. Ravi gives the drink names a light read. If one of them turns out to be a competitor's product name, he will catch it long before it gets onto a menu.
A full read with a fact check is for anything that leaves your hands, whether it goes to a client, your manager, a colleague in another team or the public. Read every sentence the way the recipient will read it. Then check every fact you could be asked about: names, dates, figures, claims about what someone said, and claims about what a product does. Ravi gives the review summary a full read before it goes to his manager, and checks each quoted complaint against the actual review.
A primary source check is for anything involving money, law, health or someone's job. Don't stop at the assistant's answer, or at a website that repeats it. Check the claim against the primary source, meaning the source that actually settles the matter: the law itself, the government agency's own page, the contract or the official policy. Ravi's question about advertising a promotion as "free" gets this level. A wrong answer could mislead customers and put the business in breach of the rules.
There are two reasons these four areas need the most careful check. The first is the cost of a mistake. A wrong figure in an Instagram caption is embarrassing. A wrong figure in someone's pay, a wrong clause in a contract, or a wrong statement about someone's medical entitlement harms a real person.
The second reason is that assistants are most likely to be wrong in these areas. AI fundamentals lesson 6.2, "Where invented answers are most likely", explained that the risk of an invented answer rises when you ask for specific details such as exact numbers, dates and legal clauses, and when the topic is niche, local or recent. Singapore's rules on employment, tax, CPF and consumer protection are specific and local, and they are updated regularly. An answer on these topics can be fluent and confident and still be a year out of date.
In these areas, use the assistant only to start your research. It can explain a concept, suggest what to look for, or help you read a document you have found. The answer you rely on must come from the primary source. In Singapore, the main ones are:
the Ministry of Manpower, for employment rules IRAS, for tax the CPF Board, for CPF Singapore Statutes Online, for the law itself
AI fundamentals lesson 6.3, "A five-minute routine for checking any answer", gives you a quick method for checking against sources like these.
Whatever level you choose, you are responsible for what you send. If a client finds a wrong figure in your proposal, they will hold you responsible. If an email to staff gets a policy wrong, it still carries your name. Nobody accepts "the assistant wrote it" as an excuse, just as nobody would accept "the spellchecker sent the wrong attachment". That doesn't mean you should avoid assistants for important work. It means you should match your checking to the stakes every time. That way the important things get proper attention and the trivial ones don't take up your whole day.
Decide the level before you read the output. If you leave it until afterwards, a fluent and confident answer will tend to persuade you to check it less carefully than it needs. Two questions set the level. First, who will see this? Second, what would a mistake cost, and who would pay for it?
Ravi now answers these questions in his head while he types the request, and it takes him seconds. The drink names are for him alone, and a wrong one costs nothing, so they get a light read. The review summary leaves his hands and could mislead his manager's decision, so it gets a full read. The promotion question involves law and money, so it goes to the primary source.
Your own list will look different from Ravi's, but the questions are the same. Think back over the past week and pick out the different things you used an assistant for, or could have, from the throwaway to the serious.
Write down ten things you might use an assistant for. Next to each, write its checking level: light, full or primary source.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).