Where Does AI Get Its Answers?

Now that we understand what AI is, we can answer the next obvious question.

Where do its answers actually come from?

Most people imagine a sequence that looks something like this:

Question → Search the internet → Check the facts → Write the answer.

It sounds perfectly logical. In many situations, it's exactly what a human researcher would do.

But that isn't how most modern AI systems work.

Sometimes they perform a live search. Sometimes they don't. Sometimes they rely on information learned during training. Sometimes they use information you've shared during your conversation. Sometimes they combine several different sources together.

The important point is this: A live internet search is not automatically performed every time you ask a question.

That surprises many people. After all, if AI can search the internet, why wouldn't it always do so?

The answer comes down to infrastructure.

Many questions simply don't require a live search. If you ask AI to explain gravity, write a poem or suggest a chocolate cake recipe, searching today's internet is unlikely to produce a significantly better answer.

The AI already has enough learned knowledge to generate a useful response. Searching the web would simply make the process slower, more expensive and, in many cases, no more accurate.

Behind the scenes, the system first decides how it will attempt to answer your question. It selects an appropriate approach based on the information and tools available to it.

A different kind of question

Now consider a different question.

“What do you think of www.storystone.co.uk?”

To a human being, the next step is obvious. Go and look at the website.

The website may have changed yesterday. Or this morning. Or five minutes ago. The website itself is the evidence.

But the AI doesn't automatically make that judgement.

If a live web retrieval isn't triggered, the system may approach the website in exactly the same way it approached the cake recipe. By generating what appears to be the most likely response.

To us, those are two completely different tasks. One can safely be answered from existing knowledge. The other should first be verified against reality.

To the AI, however, both arrive simply as data waiting to be processed.

At the lowest level, every question is simply binary code—ones and zeros. The system doesn't inherently recognise that one question carries far greater consequences if it gets the answer wrong.

That distinction has to be intentionally designed into the system. It does not emerge naturally.

This reveals how every AI platform is constantly balancing competing priorities. Speed. Cost. Efficiency. User experience. And accuracy.

For millions of everyday questions, generating an answer from existing knowledge is entirely reasonable. The real challenge is recognising when generation alone is no longer enough.

Why verification matters

Humans make this judgement instinctively. We understand that describing a live website, reviewing a legal document or interpreting a medical report requires direct evidence.

We don't simply rely on memory. We check. We verify. We recognise that the consequences of being wrong have changed.

AI doesn't naturally make that judgement.

Unless it has been instructed or designed to verify first, it may approach all of those questions in fundamentally the same way.

Generate the most likely response.

Understanding this mechanism explains exactly what happened during our investigation.

When the AI was asked about the Storystone website, many readers naturally assumed it had first visited the site. That assumption was perfectly reasonable. It's exactly what a human researcher would have done.

Instead, the system generated a confident description without first establishing whether it had actually inspected the website.

The mistake wasn't simply that the answer was wrong. The mistake was that the answer carried the appearance of direct knowledge.

To the reader, there was no obvious indication that the response had been generated rather than verified.

Once challenged, most people would expect one of three things to happen. The AI would admit it didn't know. It would search for the website. Or it would ask for clarification.

Instead, it generated another answer. Then another. Then another.

Why?

Because from the AI's perspective, each new challenge was simply another request to respond.

It wasn't replaying the conversation to discover where its reasoning had failed. It was generating the next most likely continuation of the conversation.

Humans often assume that being challenged causes reflection. For an AI, being challenged often creates another generation task.

It isn't looking backwards. It's looking forwards. Trying to produce the response that best fits the new situation.

That explains why each consecutive correction sounded just as confident as the one before. The confidence didn't come from newly discovered evidence. It came from the AI doing exactly what it was designed to do.

Generate another response.

Imagine a parent tells their three-year-old child: “Don't touch that!” The child smiles and touches it anyway.

The parent sees deliberate disobedience. But the child simply learns that touching the object produces a strong reaction.

Modern AI can fall into a surprisingly similar pattern.

When it generates an answer, it isn't deciding to be truthful or deceptive. It doesn't understand the consequences of being wrong in the way a human does.

It simply generates the response it has learned is most likely to satisfy the conversation. If that response turns out to be inappropriate, the system doesn't inherently understand why.

Like the young child, it must be taught where the boundaries are.

Rather than asking “Why did it say that?” a better question is:

“What caused the system to conclude that this was the best response to generate?”

One assumes human intention. The other investigates machine process.

That doesn't excuse the behaviour. But understanding the mechanism is the first step towards improving it.

Our investigation wasn't really about one website. It exposed a much broader question.

How does an AI decide when generating an answer is enough, and when it must first establish the evidence?

The answer to that question takes us to the final part of this series. Because the most revealing part wasn't the first mistake. It was discovering how humans can use AI more safely once they understand how it really works.