Not every question should get an AI verdict

The last two posts covered two things:

First, AI’s answers don’t come out of a “database of truth.” They’re the most likely response generated within the current context.

Second, AI doesn’t need to truly understand like a person to generate a fluent, complete, even useful answer, because it’s extremely good at handling relationships within language and context.

So here’s the question:

if AI is this powerful and yet not always reliable, how should ordinary people actually use it?

My judgment now is:

it isn’t that you can’t ask AI. It’s that you first need to be clear about:

what kind of question you’re actually asking.

Different questions call for AI to play different roles.

1. Factual questions: AI can explain, but can’t be used directly as a source

For example:

• What is this policy now?
• Is this job still open?
• Who is the CEO of this company now?
• What exactly does this regulation say?
• Does this product have this feature right now?

These questions have external facts.

The key isn’t:

which answer sounds more reasonable?

It’s:

what is reality, actually?

So for factual questions you can’t just go by what AI says. Even if AI can search the web, that doesn’t mean it’s necessarily right.

Because it may cite outdated pages. It may be reading a secondhand summary rather than the official original. It may confuse regions, status, dates and conditions of eligibility. It may also apply a seemingly relevant rule to your specific situation.

So the right way to use it for factual questions is:

have AI help you explain, organize and compare sources. But in the end go back to the official documents, the original text, the dates and the specific conditions that apply.

In a sentence:

for factual questions, AI is an explainer, not the final source.

2. Questions of probability: AI is good at laying out possibilities

For example:

• How good are my chances for this job?
• Is this product direction worth trying?
• What might it mean that this person replied this way?
• Will readers be interested in this post?
• What are the main risks of this choice?

These questions don’t have one definite answer.

You can’t know in advance whether a role will definitely give you an interview. You can’t be sure from one reply what someone is really thinking. Still less can you know whether a product definitely has a market before it’s validated.

So these questions suit probabilistic thinking better.

But “probability” here doesn’t mean having AI give you a number like 73.6%. Numbers like that are usually meaningless.

What’s more useful is:

Which explanations are more likely? Which risks are bigger? Which judgment does the current evidence support more? What should be validated next?

So for questions of probability, AI is very valuable. It can help you step out of a single idea and see more possibilities.

Say someone replies coldly. Your instinct might be:

does he not care about me?

But AI can help you break it into several possibilities:

He might be busy. He might just not say much in general. He might be avoiding something. His investment in the relationship might have dropped. Or you might be magnifying one signal because you’re anxious.

Here AI isn’t telling you the truth. It’s helping you avoid jumping to conclusions too early.

In a sentence:

for questions of probability, AI is a generator of hypotheses, not a fortune-teller.

3. Questions of value: AI can analyze pros and cons, but can’t rank them for you

For example:

• Should I quit my job?
• Should I leave a city?
• Should I stay in a relationship?
• Should I give up creativity for stability?
• Should I accept a direction I don’t like for higher pay?

Of course these questions involve probability too.

For example:

staying in a city may be more stable. Leaving may bring new opportunities. Staying in a relationship may keep a connection. Ending it may reduce the drain.

But probability can’t decide for you.

Because the real questions are:

What do you value more? What are you willing to bear? What price is unacceptable to you?

AI can tell you:

what the benefits of this choice are. What the risks are. What you might be overlooking. What the trade-offs of different paths are.

But it can’t decide for you:

whether freedom or stability matters more. Whether security or growth matters more. Whether emotional connection or self-protection matters more.

Those aren’t AI’s questions. They’re your ranking of values.

In a sentence:

for questions of value, AI is a mirror, not the master.

4. Questions of responsibility: AI can help you prepare, but can’t sign for you

For example:

• Should I take this medication?
• Should I sign this contract?
• Should I take legal action?
• Should I put money into this?
• Should I deal with a visa, banking, tax or medical issue?

These questions especially can’t be handed straight to AI.

Not because AI is useless. On the contrary, AI is great at helping you prepare:

organizing materials. Listing questions. Explaining technical terms. Rehearsing conversations with a lawyer, doctor, bank or HR. Helping you spot risks in documents that need confirming.

But it can’t make the final judgment for you.

Because if it’s wrong, you bear the consequences. Not AI.

And high-stakes questions usually hinge on many key details:

your status. Your documents. Local regulations. Deadlines. Contract terms. A doctor’s examination. Records in official systems.

No “reasonable-sounding” answer can replace these.

In a sentence:

for questions of responsibility, AI is a preparation tool, not the signatory.

So it isn’t “should I ask AI or not”

The more accurate question is:

for the question I’m asking right now, what role should AI play?

If it’s a factual question: let it explain, but check the sources.

If it’s a question of probability: let it lay out possibilities, but don’t treat it as prophecy.

If it’s a question of value: let it help you see the trade-offs, but don’t let it rank them for you.

If it’s a question of responsibility: let it help you prepare, but in the end go back to professionals, official channels and real-world action.

That’s more useful than simply saying “trust AI / don’t trust AI.”

A prompt I use often

When I’m not sure how to put a question to AI, I write:

First judge whether my question is a factual question, a question of probability, a question of value, or a question of responsibility. If it’s factual, tell me which sources need verifying. If it’s about probability, list several possible explanations and ways to check them. If it’s about value, help me break down the trade-offs, but don’t decide for me. If it’s about responsibility, help me put together a list of questions and risks, but don’t give a final conclusion.

This avoids a very dangerous situation:

I clearly only wanted a bit of help judging, and ended up handing the conclusion to AI.

AI’s best position isn’t being your master. It’s helping you break the problem apart, lay out the possibilities and shine a light on blind spots.

One last line:

AI can help you judge better. But the power to judge is best kept in your own hands.

This is part 3 of the series “how ordinary people can trust AI the right way.” Next I want to talk about whether, after using AI a lot, people slowly stop trusting themselves.