Lecture 2-1: Invented Facts
목차 31
- Lecture 1-1: What the Word "AI" Points To
- Lecture 1-2: How a Language Model Makes Words
- Lecture 1-3: Sounding Right Is Not Knowing
- Lecture 1-4: Reading a Conversation Closely
- Lecture 1-5: Practice Lab, Read an AI Answer
- Lecture 2-1: Invented Facts
- Lecture 2-2: Old Knowledge, Missing Sources, Uneven Coverage
- Lecture 2-3: Scripture and Theology, Where to Be Most Careful
- Lecture 2-4: Building a Checking Plan
- Lecture 2-5: Practice Lab, Find the Errors
- Lecture 3-1: Asking Well
- Lecture 3-2: Checking What You Are Given
- Lecture 3-3: What Must Never Be Typed In
- Lecture 3-4: Keeping Notes and Giving Credit
- Lecture 3-5: Practice Lab, Mend the Prompts, Mark the Secrets
- Lecture 4-1: Hand It Over, Help With It, Keep It
- Lecture 4-2: Office Work and Communication
- Lecture 4-3: Teaching and Study Materials
- Lecture 4-4: Where Teachers and Leaders Draw the Line
- Lecture 4-5: Practice Lab: Hand Over, Help, or Keep
- Lecture 5-1: Preaching and AI: A Theological Question
- Lecture 5-2: Care and Counselling
- Lecture 5-3: When Members and Young People Use AI
- Lecture 5-4: The Minister's Own Life
- Lecture 5-5: Practice Lab: Judge the Scene, Write a Position Paragraph
Chapter 1 gave you a plain picture of an AI chat assistant: a tool that continues words plausibly. This chapter asks what follows when you rely on it. The first and most common answer is that it will sometimes state, in a calm and fluent voice, something that is not so.
People call this "hallucination", a word borrowed from medicine that suggests the tool sees things. That picture is misleading, and this lecture uses a plainer phrase: an invented fact. An invented fact is a statement that sounds like knowledge, is built from the same material as true statements, and has nothing behind it.
The central claim of this course is that an AI system is a tool that continues language plausibly and not a knower. Invented facts are what that claim looks like in practice. If you understand why they happen, you will not be surprised by them, and you will know where to look.
이 17분으로 얻는 것
- Explain, using the picture from lecture 1-2, why a language model produces invented facts without any intention to deceive.
- Name the kinds of statement most often invented: quotations, books and articles, references, verses, names, numbers, dates and events.
- Say which requests and which conditions make an invented fact more likely.
- Explain why a confident tone is not evidence that a statement is true.
- Treat any specific claim from an AI tool as a draft claim until it has been checked.
A simple man believes everything, but the prudent man carefully considers his ways." — Proverbs 14:15
World English Bible
You shall not give false testimony against your neighbor." — Exodus 20:16
WEB
Therefore putting away falsehood, speak truth each one with his neighbor. For we are members of one another." — Ephesians 4:25
WEB
These verses are not about machines. They remind us why a statement carries weight once a person passes it on. A minister who repeats a quotation, a date or a story speaks it in their own voice, to people who trust them. Whatever the tool did or did not intend, the one who says it to the congregation is the one who answers for it.
1. What an invented fact is
Lecture 1-2 gave the picture: a language model has learned, from a vast body of text, which words tend to follow which. When you ask a question, it does not look up the answer in a store of facts. It continues your words with the words that most plausibly come next. Most of the time, plausible and true go together, because the text it learned from was mostly sensible people saying true things. But the tool has no step in which it asks, "Is this so?" It has only the step in which it asks, "What would come next here?"
So when the true answer is rare, or unknown to the tool, or absent from what it learned, the tool does not stop. It continues anyway, with something of the right shape. A request for a quotation from an early teacher gets a sentence shaped like a quotation from an early teacher. A request for a book on a subject gets a title shaped like a title, with an author's name shaped like an author's name.
Invented fact
A statement produced by an AI tool that has the form of knowledge (a name, a date, a quotation, a reference) but was made by continuing a pattern, not by finding something that exists. It may be partly right, wholly wrong, or accidentally correct.
Notice two things this does not say. First, it does not say the tool is lying. A lie needs someone who knows the truth and says otherwise. The tool has no such position. Second, it does not say the answer is always wrong. Often it is right. That is exactly what makes invented facts dangerous: they arrive mixed in with true ones, in the same voice.
2. What gets invented
Anything specific can be invented. Some kinds appear again and again.
- Quotations. A saying attributed to a named teacher, a pope, a reformer, a monastic writer, a saint. The words sound right for the person, and the person never said them.
- Books, articles and sources. A title with a plausible author, publisher and year, for a book that does not exist, or a real author given a book they did not write.
- References and page numbers. A real book given a page or chapter that says something else.
- Verses. A verse number that does not match the words, words attributed to the Bible that are not in it, or a book of the Bible said to contain something it does not. Lecture 2-3 returns to this.
- Names, numbers and dates. The year of an event, the number of people at a gathering, the name of a person in a story.
- Events and biographies. A summary of a person's life that blends facts with details made up to smooth the story.
- Web addresses and contact details. A plausible address that leads nowhere or to something else.
Here is a short example. (An invented example; the conversation and the book are both made up.)
Person: Can you give me a short quotation from an early Christian teacher about patience in suffering, with the book it comes from?
Assistant: Certainly. In his Letters on Endurance, the teacher wrote: "The patient soul is a house built on rock, for the storm cannot find a door." (Book 2, Letter 7.)
Everything in this answer has the shape of a real one: a book title, a numbered letter, a well-turned sentence. A tired preacher on Saturday night could easily paste it into a sermon. But nothing in the answer shows that the book exists, that the sentence is in it, or that the teacher wrote it. The only way to find out is to look outside the conversation.
3. When invented facts are more likely
No one can give you a safe list. But a handful of conditions raise the risk, and you can learn to notice them.
The detail is specific. A general statement ("Early Christians valued patience") has many ways to be roughly right. A specific one ("Book 2, Letter 7") has only one way to be right and many to be wrong.
The subject is rare. The tool learned from what was common. A rarely discussed teacher, a small community, a local custom or a minor text gives it little to go on, so it fills the gap with the pattern of similar things.
The request assumes something exists. If you ask, "Where in Paul's letters does he say that suffering produces patience?", the tool tends to supply an answer, whether or not he says it in those words. The question carried its own answer in. Questions that presume a fact lead the tool toward agreeing with the presumption.
You ask for sources. Asking for references pushes the tool to produce reference-shaped text. Sometimes the references are real. Sometimes they are the most plausible-looking ones. The request itself does not tell you which.
The matter is recent. Lecture 2-2 explains why. The tool learned up to a certain point and has no knowledge of what came after, unless the service adds a way to search.
The answer is long. A long answer has more places for something to go wrong, and a small error early can be carried forward by every sentence that follows it.
You push back. If you say "That doesn't sound right", some tools apologise and produce a different answer, which may be no better. A changed answer is not a checked answer.
4. Why a confident tone proves nothing
When a person speaks with great confidence about a subject, we often take it as a sign that they know it. That is a habit, and it works fairly well among people, because people who do not know usually hesitate.
A language model has no such link. Its tone is part of what it learned: most of the text it learned from was written by people stating things plainly. A fluent, assured sentence is the usual way for text to sound, so the tool produces it whether the content behind the sentence is solid or empty. The tone tells you how the text was shaped. It tells you nothing about whether the content was checked.
A false sign
"It sounded so sure" is not a reason to trust an answer, and "it hedged" is not a reason to distrust one. Some tools add cautions to everything, which makes the cautions worthless as a signal. Judge by what you can check, not by how it was said.
5. A small habit: a draft claim
The simplest protection is a habit of mind. When a specific claim comes from an AI tool, treat it as a draft claim: possibly useful, not yet a fact. A draft claim can be a good starting point. It can tell you what to look for. It cannot be put into a sermon, a lesson, a letter or a notice until someone has checked it against something that stands behind it.
The next lectures show how. Lecture 2-2 explains the limits that come from what the tool learned. Lecture 2-3 looks at Scripture and theology, where the cost of an invented fact is highest for a minister. Lecture 2-4 builds a plan for checking, and chapter 3 carries the plan into everyday work.
Two limits on this lecture. It does not tell you how often a given tool invents facts, because that differs from tool to tool and changes over time, and a number given here would soon be wrong. And it does not say that checking is a reason to avoid the tools altogether. It says that if you use them, the checking is part of the work and not an extra.
Diagram 1. Where the check has to happen
Diagram 2. Kinds of claim and the first thing to do
| Kind of claim | Typical risk | First thing to do |
|---|---|---|
| A quotation with a name | The words may never have been said by that person | Find it in a source you can open, or leave it out |
| A book or article | The title or author may not exist | Look for the book in a library or catalogue |
| A verse | The words and the number may not match | Read the verse in your own Bible |
| A date, number or name | One digit or one name may be wrong | Check one reliable reference |
| A summary of a person's life | True facts blended with filler | Read a reliable short account |
핵심 요약
- A language model continues words plausibly; it has no step that asks whether a statement is true.
- An invented fact has the shape of knowledge (a quotation, a title, a verse, a date) and nothing behind it.
- It is not a lie, because the tool has no position from which to lie, and it is not always wrong, which is why it is hard to spot.
- Risk rises with specific details, rare subjects, leading questions, requests for sources, recent events and long answers.
- A confident tone shows how text is shaped, not whether the content was checked.
- Treat any specific claim as a draft claim until it has been checked against something that stands behind it.
| Front | Back |
|---|---|
| Invented fact | A statement shaped like knowledge, made by continuing a pattern, with nothing behind it. |
| Why it happens | The tool asks what would come next, not whether it is true. |
| Kinds most often invented | Quotations, books, references, verses, names, numbers, dates and events. |
| Does confidence help? | No. Tone comes from how the text was shaped, not from checking. |
| Draft claim | A claim worth following up but not yet fit to say to others. |
One sentence to remember An answer that sounds like knowledge is not yet knowledge, and the one who repeats it is the one who answers for it.
This week's action Think of one specific claim (a quotation, a date or a source) you have met from an AI tool or heard repeated from one, and write down where you could check it.
스스로 확인해 보세요
읽었는지가 아니라 설명할 수 있는지를 묻는 것입니다. 체크는 이 기기에만 저장됩니다.
묵상 및 토론
- Why does a language model produce a quotation when no such quotation exists? Use the picture from lecture 1-2 in your answer.
- Why is "the tool is lying" the wrong description of an invented fact?
- Think of a kind of request you might make in your own work. Does it have any of the conditions that raise the risk?
- A friend says, "It gave me references, so it must have checked." What would you say?
- When a person speaks confidently, we usually take it as a sign of knowledge. Why does that habit not carry over to AI tools?
- What does it mean to treat a claim as a draft claim, and what would you do before using one in front of other people?