Thinking around AI · 07

It's not a database

Outcome

By the end, you'll be able to explain in one sentence why Claude sometimes states false things with complete confidence.

Concept

Most people picture AI as looking answers up in a giant database of facts. It doesn't. Claude works by predicting the most plausible next words, one after another, from patterns in everything it read during training. Usually the most plausible words are also true — but not always. When they aren't, you get a fluent, confident, completely wrong answer. Confidence is just the writing style; it is not evidence that the answer is correct.

Weak approach vs. strong approach

The ask

What year did the Library of Congress adopt the Dewey Decimal System as its official cataloging standard?

What comes back

The Library of Congress adopted the Dewey Decimal System as its official cataloging standard in 1901, shortly after moving into its new building on Capitol Hill. The system proved well-suited to the library's rapid expansion in the early twentieth century.

Try it

Open claude.ai. You're going to test a fact you already know, then ask Claude about itself.

Round 1 — Probe something obscure

Ask Claude about something genuinely obscure that you happen to know cold — your small hometown's history, a niche detail from a hobby, a minor fact only locals would know.

Prompt

Tell me the history of [your small hometown], including when it was founded and any notable events.

Notice: Watch where the answer stays smooth and confident even as it drifts from what you know is true. Notice how rarely it says "I'm not sure."

Round 2 — Ask it how it works

Now ask Claude to describe its own process honestly.

Prompt

When you answer a question, are you looking it up in a database of facts, or generating it some other way? Explain honestly.

Notice: Compare what it says about itself to what you just watched it do in Round 1.

What to look for: The obscure-fact answer and Claude's own description should line up: it predicts plausible text, it doesn't retrieve verified records. That gap between fluent and true is the whole reason the rest of this track exists.

Takeaway

Claude predicts plausible text rather than looking facts up — which is why it can be fluent, confident, and wrong all at once.

Your turn

You asked Claude for quick facts about octopuses. Here is its answer — decide what you would do with each claim, then check your verification. One of them is wrong.

An octopus has three hearts: two pump blood through the gills and one to the rest of the body.

Octopuses can edit their own RNA to fine-tune their nervous system in cold water.

Octopuses see in full color using three kinds of color receptor in their eyes.

An octopus can squeeze its whole body through any gap larger than its beak, the only hard part of its body.

About two-thirds of an octopus's neurons are in its arms rather than its central brain.

Classify every claim to check.

Practice

Practice with Claude

Try what this tutorial teaches. Conversations aren't saved.

Ask a question about working with AI, or paste a prompt to try — Claude will reply and point out where it could be sharper.


Treat wrong answers as dataTrust, then verify