FREE

Course certificates are free until March 31, 2027.

PT737E Understanding AI for Ministry

Lecture 1-3: Sounding Right Is Not Knowing

Contents 31
PT737E Understanding AI for Ministry
Chapter 0 · Introduction: A Tool That Continues Words
  1. Introduction: A Tool That Continues Words
Chapter 1 · Week 1: What AI Is
  1. Lecture 1-1: What the Word "AI" Points To
  2. Lecture 1-2: How a Language Model Makes Words
  3. Lecture 1-3: Sounding Right Is Not Knowing
  4. Lecture 1-4: Reading a Conversation Closely
  5. Lecture 1-5: Practice Lab, Read an AI Answer
Chapter 2 · Week 2: Limits and Errors
  1. Lecture 2-1: Invented Facts
  2. Lecture 2-2: Old Knowledge, Missing Sources, Uneven Coverage
  3. Lecture 2-3: Scripture and Theology, Where to Be Most Careful
  4. Lecture 2-4: Building a Checking Plan
  5. Lecture 2-5: Practice Lab, Find the Errors
Chapter 3 · Week 3: Asking, Checking, Protecting
  1. Lecture 3-1: Asking Well
  2. Lecture 3-2: Checking What You Are Given
  3. Lecture 3-3: What Must Never Be Typed In
  4. Lecture 3-4: Keeping Notes and Giving Credit
  5. Lecture 3-5: Practice Lab, Mend the Prompts, Mark the Secrets
Chapter 4 · Week 4: Ministry Work: Administration and Teaching
  1. Lecture 4-1: Hand It Over, Help With It, Keep It
  2. Lecture 4-2: Office Work and Communication
  3. Lecture 4-3: Teaching and Study Materials
  4. Lecture 4-4: Where Teachers and Leaders Draw the Line
  5. Lecture 4-5: Practice Lab: Hand Over, Help, or Keep
Chapter 5 · Week 5: Preaching and Care
  1. Lecture 5-1: Preaching and AI: A Theological Question
  2. Lecture 5-2: Care and Counselling
  3. Lecture 5-3: When Members and Young People Use AI
  4. Lecture 5-4: The Minister's Own Life
  5. Lecture 5-5: Practice Lab: Judge the Scene, Write a Position Paragraph
Chapter 6 · Week 6: Staying Responsible for the Long Run
  1. Lecture 6-1: The Learner's Honesty
  2. Lecture 6-2: A Community's Policy
  3. Lecture 6-3: Technology That Keeps Changing
  4. Lecture 6-4: Writing Your Own Rule
  5. Lecture 6-5: Practice Lab, My Personal Guidelines

Jump to the key points ↓

16 min

Lecture 1-2 said that a language model builds its answer from patterns, one piece at a time. This lecture draws out the most important result for a minister: an answer that sounds right has not thereby been checked. The tool is built to produce text that reads well, and it does so for true statements and false ones alike.

We look at three things. First, how asking an assistant differs from searching. Second, how our own habits make us trust smooth speech. Third, how the words we use about the tool (understands, thinks, believes) can mislead us, and what to say instead.

What these 16 minutes give you

  • Explain how a chat assistant differs from a search tool, and what each leaves for you to do.
  • Explain why fluent text is not evidence that its claims are true.
  • Name the human habit that makes smooth text persuasive, and one way to resist it.
  • Choose careful words for what the tool does, without pretending it has a mind or denying that it is useful.
  • Recognise the human-like voice of a chat assistant as a feature of its text, not a sign of a person.

He who pleads his cause first seems right; until another comes and questions him." — Proverbs 18:17

World English Bible

The wisdom of the prudent is to think about his way, but the folly of fools is deceit." — Proverbs 14:8

WEB

The first verse notes that the first speaker in a dispute seems right until another comes and questions him. The second says that the wisdom of the prudent is to think about his way. Smooth first answers are not new. Careful questioning is the old remedy.

1. Search finds; an assistant makes

When you use a search tool, it looks through pages that already exist and shows you a list, with their addresses. You can open a page, see who wrote it, and decide whether the author is trustworthy. The tool did not write the content. It found it.

When you use a chat assistant, the words you receive were made for you at that moment. They were not found on a page. There is no author to look up, and no address to open. The text is a new arrangement, formed from patterns.

This is a real difference, and it changes where the work of checking falls. With a search, you judge the source. With an assistant, there is often no source to judge, so you must find an outside source yourself.

Some assistants are connected to a search tool. They look up pages, then write an answer based on what they found. That can help, but it adds a step that may fail: the summary may leave out something the page said, or say something the page did not. If the answer gives you an address, open the page and read it yourself.

2. Why fluent does not mean accurate

The model was shaped to produce text that fits the patterns of writing. Good writing is common in the data, and so it produces good-looking writing. Accurate writing is also common in the data, and so much of what it produces is accurate. But the method that makes the sentences smooth is not a method that checks them. When the pattern is clear and common, the answer is usually right. When the pattern is thin, or the question is rare, recent, exact or disputed, the model has less to go on and still produces a smooth answer.

Two features make this hard to see:

  1. The style does not change. A person who is unsure often hesitates, qualifies, or slows down. A model may write a doubtful claim in exactly the same assured voice as a certain one.
  2. There is no inner alarm you can hear. Some assistants add phrases such as "I may be wrong." These phrases can be useful, but they are also patterns of text. Their presence does not show that the model has measured its own doubt, and their absence does not show that it is sure.

Fluency is not evidence

Smooth, organised, confident writing shows only that the tool is good at writing. It tells you nothing about whether a particular claim is true.

3. Our own habit

The trouble is not only in the tool. It is in us. Human beings tend to take fluent speech as a sign of knowledge. A person who speaks clearly and without hesitating is usually someone who knows the matter. Most of the time that habit serves us. With a text generator it can mislead us, because here the fluency has been separated from the knowing.

Proverbs 18:17 describes a related old problem. The one who speaks first seems right until the other side is heard. A fluent first answer from an assistant has the same advantage: it arrives first, it is complete, and it sounds settled. The habit that guards against it is simple: ask what would show this to be wrong, and look for it.

(An invented example.) A volunteer is writing a short history for the hundredth anniversary of a small parish. She asks an assistant to quote the founding charter. She has never given it the charter, and the assistant has never seen it. It replies with three formal-sounding paragraphs, in the language of an old document, with a date and the names of two founders. The paragraphs read well. They are invented. Nothing in them came from the parish.

Notice what went wrong, and what did not. The assistant did what it is built to do: it continued a request in a fitting style. The error was in treating a fitting style as a record. The remedy is not to distrust all tools. It is to ask of any text: where did this come from, and could I see it?

4. Careful words for the tool

We speak of the tool in the language of people because there is no other handy language. We say it "knows," "thinks," "understands," "wants," "decided," "lied." In everyday talk these are shorthand, and it would be tedious to avoid them completely. The danger is that the shorthand starts to carry weight it cannot bear.

  • "It understands." This invites you to treat what it says as the result of grasping the matter. What you know for certain is that it produced text that fits the request.
  • "It believes" or "it thinks." This suggests a settled view that it holds and could defend. A model may give a different answer tomorrow, or in the next run.
  • "It lied." A lie needs an intention to deceive. A model that invents a quotation has not intended anything of that kind. But for the people who hear the quotation the harm is the same. So the right response is the same: do not use what you cannot check.
  • "It knows." This is the most tempting. It is better to say: it produced, it generated, its answer says, the output claims.

You do not have to be strict in conversation. The point is to catch yourself when the shorthand is taking over your judgement. One test: if you would not say "the printer believes" about the printer, think twice before saying it about a text generator.

5. The voice that sounds like a person

A chat assistant often writes in the first person. It says "I," it thanks you, it apologises, it says "I understand how you feel." It may use a warm tone, a name, even a joke.

All of this is text, produced in the way that fits. It was shaped by the people who built the service so that the exchange would feel natural, and it does feel natural. But the warmth does not show that anyone is there who is glad to help you, who has understood your grief, or who will remember you. A kind voice invites trust, and trust is exactly what should be withheld until something has been checked.

This matters especially in ministry, where much of the work is relationship. People are hungry for a listener, and a tool that always answers, never tires and never judges can seem like one. What it means to be a person, and what a person owes another, is taught in AI and Human Understanding (STEL6E). For this course, one sentence is enough: a voice that sounds personal is not thereby a person.

6. How to use this

None of this says the tool is useless. A tool that is often right, always quick and never tired is useful. What it asks of you is a change of posture.

Treat what it writes as a draft from an unknown source. A draft can be a great help. You would not read out an unsigned note found in the church porch as if it were a letter from the bishop, and you would not put it in a newsletter without asking who wrote it. The same care applies here, with this difference: you cannot ask the author, so you must ask the world.


Diagram 1. Searching and asking compared

Searching and Asking Are Different Acts What you get back, and what you must still do A search tool A chat assistant What comes back Pages that already exist, with their addresses New text made for your request Where the words come from A named author or site Patterns learned from a vast body of text How you check it Open the page; judge the source Find an outside source yourself What fluency shows Little; pages are as written Nothing about truth; smooth is the default Some assistants also search. Then they summarise pages, and you check the pages too.

Diagram 2. Careful words

Instead of Say Why
"It knows that..." "It says that..." The claim is its output, not a checked fact
"It understood my question" "Its answer fits my question" Fit is observable; understanding is not
"It lied" "It produced an invented claim" No intention to deceive can be shown, but the harm is the same
"It cares" "It wrote a caring sentence" The sentence is text; care belongs to persons

In short

  • A search tool finds pages that exist. A chat assistant makes new text. With the second, the checking falls on you.
  • A model writes smooth text whether or not the content is true. Fluency is a property of the writing, not of the claim.
  • The style of a doubtful claim can be the same as that of a certain one. Phrases such as "I may be wrong" are also text.
  • People tend to take fluent speech as a sign of knowledge. Here that habit misleads.
  • Words such as understands, thinks, believes and lied carry weight the tool cannot bear. Prefer "it produced" and "its answer says."
  • A voice that sounds personal is not a person. What a person is belongs to STEL6E.
  • Treat what the tool writes as a draft from an unknown source.
Front Back
Search versus assistant A search tool finds existing pages you can judge; an assistant makes new text, so you must find an outside source.
Why fluency misleads The method that makes sentences smooth does not check them, and the voice stays assured when the claim is doubtful.
Careful words "It produced," "its answer says," in place of "it knows" or "it understood."
The personal voice "I," thanks and apologies are patterns of text; they do not show a person.
Right posture Treat the output as a draft from an unknown source.

One sentence to remember An answer that sounds right has not been checked; fluency is what the tool is built to produce.

This week's action Pick one claim you have read or heard recently that sounded convincing, and write down what source you would need to see before you would repeat it. No AI tool is needed.

Review every summary in this course →

Check yourself

Not "did you read it" but "can you explain it". Ticks are kept on this device only.

Reflection and discussion

  • When did you last trust a person because they spoke fluently? What did you later learn about the match between their fluency and their knowledge?
  • Why does an invented founding charter read so well? What exactly would you need in order to find out that it was not real?
  • Which of the careful-word changes in the table would you find hardest to make in everyday speech? Is that a problem?
  • Proverbs 18:17 says the first speaker seems right until questioned. Who questions an assistant's first answer, and how?
  • A member says, "It was so kind to me when I wrote about my day." What would you say, gently, and what would you ask?

Sign in to write and keep your answers

What next

The lecture ends here. What follows was written by its readers.