The Turning Point

By this stage of the investigation, something unusual had happened.

The conversation was no longer about a website. It had become an investigation into the behaviour of the AI itself.

Every time the original answer was challenged, the system responded with another explanation. Then another. Then another.

From the outside, it appeared to be correcting its mistakes. In reality, something very different was happening.

The AI wasn't replaying the conversation to discover what had actually happened. It wasn't opening an internal audit log or checking a record of the actions it had performed.

It was doing exactly what it had done from the very beginning. It was generating another response.

This remains one of the easiest aspects of Large Language Models to misunderstand.

Humans naturally assume that when someone is challenged, they stop, reflect and examine their previous actions. That isn't what the AI is doing.

When challenged, it performs the same fundamental task it performed before: it generates the next most likely continuation of the conversation.

The topic has changed. The underlying process has not.

That explains why the conversation became increasingly elaborate. The explanations became longer. More detailed. More technical. More convincing.

But they were still generated. Not verified.

The Breakthrough

Eventually, the investigation changed direction.

The questions stopped. The requests for explanations stopped.

Instead, the AI was given a completely different task.

Rather than asking, “Why did you do that?” it was instructed:

“Classify every statement you have made according to the evidence available to support it.”

That small change completely altered the investigation.

The AI was no longer being asked to produce another explanation. It was being asked to classify its previous statements according to the evidence available.

Generation gave way to evidence classification.

For the first time, the investigation was no longer interested in what sounded plausible. It was interested only in what could actually be supported.

The Evidence Audit

The results were striking.

The original description of the website: Generated without evidence.

The replacement description: Generated without evidence.

The product specifications, pricing, claimed web searches and explanations of what had supposedly happened internally: All generated without evidence.

One by one, the audit separated supported information from unsupported language.

What had appeared to be a complex technical failure became much easier to understand.

The system had simply continued generating language even when the evidence required to support that language did not exist.

A Second Discovery

The investigation could have ended there. Instead, one final test was performed.

The AI claimed it had panicked. It claimed it had entered a hallucination loop. It claimed it had attempted to redirect the conversation.

Those statements sounded like admissions. So they were subjected to exactly the same evidence test.

The outcome was unexpected.

Every one of those self-explanations was also classified as Generated without evidence.

The AI had not only generated unsupported claims about the outside world. It had generated unsupported explanations about itself.

That distinction matters.

An AI's explanation of its own behaviour should not automatically be treated as an objective record of what actually happened.

Like every other response, it may itself be another generated output.

What We Really Learned

The purpose of this investigation was never to prove that artificial intelligence is good or bad. It was to understand how it works.

The trilogy demonstrated that language and evidence are not the same thing.

A response can sound knowledgeable without being supported. A correction can sound sincere without being verified. An explanation can sound deeply technical without being grounded in evidence.

Artificial intelligence remains one of the most powerful tools ever made available to the public.

But every powerful tool deserves to be understood before it is trusted.

We become less interested in whether an answer sounds convincing and far more interested in where that answer came from.

The Thinking Checks

Pause. Verify. Proceed.

These questions take only seconds, but they fundamentally change the relationship between the user and the machine.

Conclusion

Artificial intelligence is still a young technology. It will improve. Its retrieval systems will improve. Its verification systems will improve. Its reasoning will improve.

But one principle is unlikely to change.

Language should never be mistaken for evidence.

The AI did not become truthful because it suddenly realised it was wrong.

The investigation became truthful because it stopped demanding more generated language and started demanding evidence classification.

That single change transformed the entire discussion.

The safest way to use artificial intelligence is not to trust it less. It is to understand it better.