Lesson 3 — The Confident Guesser: How the Machine Actually Works

“The one who states his case first seems right, until the other comes and examines him.” — Proverbs 18:17 (ESV)

This lesson takes about 25 minutes.

Try this at a dinner table: say “Give us this day our daily…” and stop. Everyone’s mind supplies bread before you finish — not because they looked it up, but because they’ve heard the pattern a thousand times.

That reflex — completing a pattern from long exposure — is, at heart, what a large language model does. This lesson opens the hood, with no math and no code, because one honest look at the machinery will protect you better than a hundred warnings. Once you see that it’s a guesser, you’ll never again mistake its confidence for knowledge.

What you’ll learn today

You’ll learn the three facts about how AI works that every user needs: it predicts rather than knows, it learned from a snapshot of human writing, and its confident tone is a style, not a signal of truth. Together they explain nearly every strange thing AI does — the brilliance and the blunders alike.

Fact one: it predicts the next word

When you send a message, the model does not go look anything up. It computes, from all the patterns it absorbed, what word most plausibly comes next — writes it — then asks the question again, and again, until the answer is done. A prayer’s cadence, a recipe’s shape, a lawyer’s phrasing: it has seen so many examples of each that it can continue almost any pattern you start.

This is why it’s so good at form: letters, outlines, summaries, explanations. Form is exactly what patterns capture. And it is why the machine has no built-in preference for truth. When the true next words and the plausible-sounding next words are the same — which, on well-trodden subjects, is most of the time — it shines. When they differ, it will choose plausible every time, because plausible is all it can see.

Fact two: it learned from a snapshot

The model’s “knowledge” is the writing it was trained on — an enormous but finished library, with a cutoff date. Three consequences follow:

• It can be out of date. Ask about last month’s news or this year’s prices, and a plain model answers from a frozen world (some tools can search the live web when asked — Lesson 4 uses that — but the base machine does not).
• It inherits our errors. The library was written by fallen humanity: mistakes, biases, and popular falsehoods went into the training data, and patterns of them come out.
• It knows the sound of citations. It read millions of footnotes, so it can produce perfectly formatted references… to studies, books, and court cases that do not exist. It isn’t lying — lying requires knowing the truth and choosing against it. It’s completing a pattern, and the pattern of a citation is easy to complete.

Fact three: confidence is a writing style

Here is the fact to underline twice. Human experts wrote most of the confident prose in the training library, so the model learned to write confidently — about everything, including its mistakes. A true answer and a false one arrive in the same assured, fluent, well-organized voice. There is no stammer to listen for, no shifty eyes to watch.

Proverbs saw this coming three thousand years ago: the one who states his case first seems right, until the other comes and examines him. The machine always states its case first, and it always seems right. Your job — the Christian’s old job — is to be the one who comes and examines.

Coach’s Note — People get burned by AI in exactly one way: they let fluency stand in for verification. Not because they’re foolish — because fluent, structured, confident text has always been our shortcut signal for competence. The machine breaks that signal. From today, let confidence trigger curiosity instead of trust: the smoother the answer, the more deliberately you check the load-bearing fact. Lesson 4 gives you the method.

When the world asks: “If it’s just autocomplete, why is it so smart?”

Fair question — the machine really does draft contracts and pass exams. The honest answer: an almost unimaginable amount of pattern, compressed. Human writing encodes human thought, so a machine that models our writing borrows the shape of our thinking — genuinely useful shape, at industrial scale.

But notice what that explanation concedes: all of it is downstream of minds. The library it learned from is the fossil record of actual understanding — people who knew things, writing them down. The machine is impressive the way an echo in a cathedral is impressive; the voice came from somewhere else. Christians need not shrink this achievement to feel secure — we can admire the engineering honestly and still see clearly that pattern is not perception, and prediction is not knowledge. “For who knows a person’s thoughts except the spirit of that person, which is in him?” (1 Corinthians 2:11, ESV). There is no spirit in the machine — only the residue of ours.

Practice right here

Below this lesson is a practice panel called “Plausible or True?” It deals you a series of confident, fluent, AI-style statements — the kind of sentence that sounds like an encyclopedia. Some are true. Some are inventions wearing truth’s clothing. You decide, one at a time, and each reveal shows you the tell.

Expect to get some wrong — that’s the point, and it’s the safest possible place to feel the trap. By the last card you’ll notice the change in yourself: fluency has stopped feeling like proof.

Now do it for real

You’re going to catch the guesser in the act, using questions where you hold the answer key:

1. Open your assistant (the free account from Appendix A).
2. Ask about something you know cold — your town’s history, your trade, a hobby you’ve spent years on, your church body’s history. Example shape: What can you tell me about [your topic]? Include specific dates and names.
3. Grade it like a teacher. Mark each claim right, wrong, or “suspiciously specific.” Asking for dates and names matters — that’s where a guesser drifts.
4. Now ask it to keep going: Tell me three more little-known facts about [same topic]. Little-known facts are where invention blooms. Grade again.
5. Notice the tone never changed. The wrong claims sounded exactly like the right ones. You’ve now seen, with your own eyes, the single most important fact in this book.

If everything it said was true — good! Common topics are its home field. Try once more with something more local or more niche, and watch the edges fray.

Faith foundations

Scripture never says “trust every confident voice”; it says nearly the opposite, in both Testaments. Israel was commanded to test even impressive claims — a prophet whose sign comes true but who leads toward other gods must be refused (Deuteronomy 13:1–3), and “if the word does not come to pass… the prophet has spoken it presumptuously” (Deuteronomy 18:22, ESV). Confidence was never the criterion; faithfulness to truth was.

The New Testament hands the same discipline to the church: “test everything; hold fast what is good” (1 Thessalonians 5:21, ESV); “do not believe every spirit, but test the spirits” (1 John 4:1, ESV). The Christian mind was trained for the AI age centuries in advance — we were never permitted to outsource belief to eloquence.

And here the machine, honestly understood, becomes almost a parable: a voice that speaks with total assurance and no knowledge of truth. We have met that pattern before, and Scripture told us what to do about it — examine, test, hold fast what is good.

For your group

1. Where else in life do we treat confidence as evidence? What does that habit cost us?
2. In the “grade the machine” exercise, what did it get wrong — and how did the wrong parts sound?
3. Read Deuteronomy 18:21–22 and Proverbs 18:17. How did God’s people test claims before checking was easy? What’s our equivalent?
4. “It isn’t lying — it’s completing a pattern.” Does that distinction change how you think about the tool? Does it change your responsibility when you pass its words along?
5. Is there a claim you’ve believed mainly because it was said smoothly? How would you check it now?

You did it

You’ve looked under the hood and found no oracle — a guesser, trained on our library, fluent in our confidence. That knowledge doesn’t make the tool useless; it makes you safe to use it. Prediction plus your verification is a genuinely powerful combination. Prediction alone is a hazard.

Up next: Lesson 4 turns “verify it” from advice into a five-minute habit — the Berean method, updated for machines.