Why the AI's Answer Changed [True Story Vol. 7]

Former Hospital Clerk'sAISide Hustle True Story — The Day the Mystery Was Solved

(Click here for the previous article)

Something like this happened in Vol. 2.

The AI kept saying, "Create a listing page on CrowdWorks." Even though no such page existed anywhere.

No matter how many times I told it, "I can't find it," the answer didn't change.

But the moment I said, "Check the user manual," the answer changed.

And when I pressed it and provided clear conditions, the AI's tone changed.

I had been wondering about this for a long time. Why did the answer change? Today, I found out the reason.

(Vol. 2 is here↓)

(If you want to read from Vol. 1, click here here)

First, understand how AI works in a super simple way

Setting aside the difficult stuff, I'll tell you only the most important thing.


AI is a tool that generates a "plausible continuation."


You throw a question at it. The AI predicts and returns "the answer that should come next" based on that question.

What's important here is that the AI is not looking for the "correct answer." It is generating a "plausible answer to this question."

Therefore, if the question changes, the answer changes. If the premise changes, the answer changes. If the conditions change, the answer changes.

This is the fundamental reason why the AI's answer changes.

Dissecting what happened in Vol. 2

Using this mechanism, let's look back at the events of Vol. 2.

① Why did it suggest CrowdWorks?

The first question was this:

"Where should I register?"

In response to this question, the AI generated a "plausible answer." The largest in the country, with many projects, a place you should register first. That was CrowdWorks.

The condition "I want to list my own service" was not included in the question. So the AI did not consider that condition.

It wasn't wrong. It was just misaligned with the answer I was truly looking for.

② Why didn't the answer change when I said "I can't find it"?

When the AI receives the words "I can't find it," the AI generates a "plausible answer for the situation where the listing page cannot be found."

That was, "Please look elsewhere."

The AI was already speaking under the premise that "there is a listing page on CrowdWorks." The information "I can't find it" alone was insufficient to break that premise. If the premise doesn't change, the answer doesn't change.

③ Why did the answer change when I said "Check it"?

The words "Check the user manual" were not a report of the situation, but an instruction for action.

Upon receiving this instruction, the AI tried to generate a "plausible answer as a result of checking the user manual." As a result of actually checking, it turned out that the premise was wrong.

Because the premise collapsed, the answer changed.

④ Why did the tone change when I pressed it?

The words "From now on, please obtain and organize reliable information before speaking" were a clear setting of conditions.

The AI received this condition and began to generate a "plausible answer under these conditions." Because the conditions changed, the quality of the answer changed.

In other words, it's like this

----------

What determines the AI's answer is
the three things: "question," "premise," and "conditions."

Question is vague → Answer is vague
Premise is wrong → Answer is wrong
No conditions → AI answers without constraints

Question is clear → Answer is clear
Correct the premise → Answer is corrected
Give conditions → AI follows those conditions

----------


It's not about whether the AI is smart or not.

Depending on how you communicate, the AI's answer changes completely.

How to actually use it

With this mechanism in mind, here is how you should think when asking an AI something.


Clarify what you really want to know before asking.


If you had asked 'What platforms allow me to list my own services?' instead of 'Where should I register?', Coconala would have appeared from the start.

If you feel the answer is strange, 'question the premise'.

Instead of just reporting that you 'can't find it', shake the premise by saying 'please verify if that information is correct'.

Give the AI clear conditions.

'Please answer after checking for reliable information', 'Please summarize into three bullet points', 'Please consider this in the context of a medical setting'. When there are conditions, the accuracy of the answer increases.

What has been learned from Vol. 2 to here

----------

Vol. 2: AI made a mistake
→ Actually, the question was ambiguous

Vol. 3: I pressed for answers
→ The answer changed because I broke the premise

Vol. 4: Correct information came out
→ Because the premise was corrected

Vol. 5: I was the one who was wrong too
→ I was acting without verifying

Vol. 6: I took the lead in using it
→ I improved accuracy through back-and-forth

Vol. 7: I understood why the answer changed
→ Questions, premises, and conditions determine the answer

----------

The true nature of the event in Vol. 2 where I thought "the AI made a mistake" is finally clear here.

The AI didn't make a mistake; it was just answering my questions honestly.

My Honest Current Status

----------

Earnings: Almost zero
CrowdWorks: Registration complete, waiting for projects
Coconala: Listing page published
X (Twitter) activity: Followers gradually increasing
Relationship with AI: I've started to understand how it works

----------


I haven't earned money yet.

But once you understand how AI works, the way you use it changes. When the way you use it changes, the results should change too.

Next Preview

I understand the mechanism now.

Next, I will write specifically about "how to actually communicate to change the quality of the answers."

(To be continued in Vol. 8.)

👉Past articles in this series are compiled in a magazine.

*For those thinking about a "side hustle with AI," let's go through trial and error together. Send your thoughts and questions to my X DM.*