Chatbot Gave You a Number? how to get its source

Start with what the machine says about itself

Someone asked an AI assistant a direct question: should people trust you?

Its answer, in full:

"Blindly trusting an AI on critical topics is a major risk, because you are essentially trusting a black box of training data curated by corporations and engineers whose specific choices, filtering methods, and internal guidelines are largely hidden from the public. Because of how I operate, nobody should treat my default summaries as an absolute source of truth."¹

That is not a critic's characterization. It is the tool describing itself.

And it did not stop there. Asked why it had not checked its sources before answering, it explained the mechanism:

"I only access the live internet or retrieve specific documents when you explicitly ask me to. I do not have a background process that constantly reads and pre-verifies everything in the background before you ask a question. When you ask a broad question, I pull from the text patterns I was trained on."¹

This chapter, and this track, take it at its word.

The two different things called "an answer"

There is a distinction hiding inside that quote and everything else follows from it.

The default answer is what you get when you ask a question and something replies. It is assembled from what the system already holds. Nothing was looked up. Nothing was checked.

The checked answer is what you get when a document was opened, read, and named. It might be the same words. It is a different object.

The two are indistinguishable on the screen. Both arrive in confident prose. Both may carry links. Only one of them involved anybody looking at anything.

The word in that quote doing the most work is "default". The system is not saying everything it produces is wrong. It is saying: the unchecked version is not evidence.

Why this is not only about AI

It would be easy to read this as a chapter about chatbots. It isn't.

A default answer is what you get from almost every source of financial information, and it has always been so:

• The salesperson who tells you a product is "very popular with people in your situation."
• The article that says "experts recommend three to six months of expenses."
• The comparison site whose ranking you cannot see the rules for.
• The illustration showing what a policy "could" be worth.

None of those involved a document either. They are defaults with a human voice.

What AI changes is not the existence of unchecked answers. It is the volume, the confidence, and the fact that a citation can now appear beside a claim without anyone having verified that the citation supports it.

The uncomfortable half

If this track only pointed at other people it would be worth nothing.

Plenee produces default answers too. So does this chapter. Any system that retrieves documents and writes summaries from them can attach a citation to a sentence the citation does not support — including ours. Measured rates of that failure in commercial, citation-producing, retrieval-based tools run from roughly 17% to 33%.²

We are not exempt from the problem this track describes. What we can do is make the checking possible: name the document, quote what it says, and mark plainly the places where nothing was confirmed.

That is the whole claim, and it is smaller than "trust us." It is: you can check.

What the rest of this track does

The machine's own advice was to demand primary documents. That is good advice and almost nobody knows how to follow it. The chapters that follow are the how:

• What actually happens when a system is shown evidence that disagrees with it — which has been

measured, and the result is not what most people expect.

• Why "I couldn't find a source" quietly favors whoever publishes most.
• How to tell a primary source from a summary of a summary, in under a minute.
• The small number of questions that do most of the work on any claim, from any source.
• Why "I don't know" is a better answer than most things you will be told.
• And how to check us.

Where Plenee fits

Every figure in this Academy names the document it came from. Where we could not confirm something, the chapter says so rather than filling the gap with something plausible.

That is not a claim to be unbiased. Plenee is a company that writes about the industry it competes in, and you should read it in that knowledge. It is a claim that our working is visible, which is a thing you can test rather than a thing you have to believe.

The short version

The tool itself says its default answers are not a source of truth, and explains why: it answers from memory and only looks things up when told to. That distinction — between an answer that was checked and one that was composed — is the most useful idea in this track, and it applies to salespeople, articles and comparison sites exactly as much as to software. Every source produces defaults, including this one. The difference worth paying for is whether you can see which you are getting.