Lecture 2-2: Old Knowledge, Missing Sources, Uneven Coverage
Contents 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
Lecture 2-1 looked at statements the tool makes up. This lecture looks at a quieter problem: statements that were once true or are only partly true, because of what the tool learned and what it did not.
Three limits come from the same root. A language model is built from a body of text collected up to a certain point. That body is finite, it is not a complete library, and it is not an even sample of what people have thought. So the tool's knowledge is old, gappy and lopsided. None of this is a fault of the person using it. But a minister who does not know it will be caught by it, usually in the place where the community is smallest and its customs most particular.
What these 17 minutes give you
- Explain what a training date is and why it makes some answers out of date.
- Explain why a tool cannot name where a sentence "came from", and why a missing source is not the same as an absent fact.
- Describe how what is common in the learning material becomes common in the output.
- Recognise when a smaller tradition, language or local practice is likely to have been flattened into a general picture.
- Say what to check first when a question concerns a recent event, a local custom or a minority position.
He who gives answer before he hears, that is folly and shame to him." — Proverbs 18:13
World English Bible
The heart of the discerning gets knowledge. The ear of the wise seeks knowledge." — Proverbs 18:15
WEB
A quick answer is not the same as a heard one. The wise person seeks knowledge, and one way of seeking is to ask what an answer rests on before taking it. These verses do not mention machines, but they describe the stance this lecture asks for: listen to the answer, and also ask what it was built from.
1. Old knowledge
A language model is made by exposing it to a large body of text. The text was gathered up to some point in time. After that point, the model learns nothing more. The point is often called the training date or the cutoff, and in practice it means this: the tool speaks as if the world stopped when its learning stopped.
Time does not stop. Months and years go by between the end of the tool's learning and the day you use it. In that gap, people are elected and replaced, rules change, books are published, buildings close, a bishop moves, a church changes its schedule, a city renames a road. The tool does not know. And it does not always say so. Asked about the present, it often answers about the time it learned, in the present tense.
An example. (An invented example.) A volunteer asks a chat assistant, "Who leads the national association of churches in my country, and how do I contact them?" The assistant names a person and gives a general description of the office. The person named may have left the post. The contact details, if given, may be years old. The answer was a fair summary of what the tool had learned. It was not a statement about now.
Some services add a step that searches the web or reads documents you supply before answering. This changes the picture in two ways. It can bring in recent information. But it brings its own risks: the search may pick a weak page, the tool may misread a good one, and the answer may blend what it found with what it already "knew" without telling you which is which. If your service does this, find out how, and ask to see the page it read. If you cannot tell whether it searched, assume it did not.
Do not trust the tool's own account of its date
Asking the tool "How recent is your knowledge?" gives an answer in the same fluent voice as any other, and may be wrong. The service's own description, where it gives one, is better than the tool's guess. For anything that depends on the present, check the present: a current notice, an official page, a phone call.
2. Missing sources
When you read a book, you can see where a claim comes from: a footnote, a name, a page. A language model has no such trail. It did not store documents to consult. During learning, the text left a mark on the model's settings, much as years of reading leave a mark on a person's way of speaking without leaving a list of books. When the model writes a sentence, it cannot go back and point to where the idea came from.
This is why asking a tool "What is your source?" is an uncertain move. It will produce something source-shaped, as lecture 2-1 explained. A reference it gives may be real, partly real or made up. The correct route is the other way round: start from a source you can open, and then see whether the tool's sentence matches it.
There is a second kind of missing source. The learning material was never a complete library. Much of what people know is not on the open internet in text form:
- books still under copyright and not freely available in digital form;
- the records of a parish, a diocese, a mission or a small denomination, kept on paper or in a drawer;
- what elders tell, what is sung but not written, what is done at a feast or a funeral in a particular village;
- writing in languages with few texts online.
When something is absent from what the tool learned, the tool does not say, "I never saw this." It continues as it does with everything. The result is a general answer where a particular one was needed. An absent fact becomes a smoothed-over statement, and the reader has no sign that anything is missing.
3. Lopsided coverage
The third limit follows. The learning material is not an even sample of human thought. What is written about most often is written about most often. A large, well-documented body of opinion appears in the material many times, in many forms. A small one appears a few times, or only as described by someone else.
The tool's output follows the same shape. What is common in the learning material tends to be common in the answers. This gives several patterns a minister should know.
The common view stands in for all views. Ask how a Christian community understands some practice, and the answer may describe the views that are most written about in the languages the tool learned from. A smaller tradition, a minority position inside a larger one, or an older practice may be missing or described only through someone else's summary.
Traditions are blended. The tool may run the language of two traditions together, using a term in a sense that belongs to one while describing the other. A reader who does not already know the difference will not see it. Lecture 2-3 gives examples in Scripture and theology.
Language matters. A tool is usually stronger in the languages with the most text. A question put or answered in a language with fewer texts, or about a place where that language is spoken, is more likely to carry errors, odd phrasing or borrowed assumptions. If the answer will be read by speakers of that language, have one of them read it.
The local is flattened. A local feast day, a community's own way of keeping a season, a regional saint, a town's history: these are the details the tool is least likely to have, and most likely to replace with a general description.
An example. (An invented example.) A minister in a small mountain parish asks a chat assistant to write a short note for the feast of the parish's patron. The assistant writes a warm paragraph about the patron's life and about how the feast is "celebrated with a procession and a communal meal in many communities". The minister's parish keeps a vigil, with no procession. The patron's life, as given, merges two people with similar names. Each sentence is reasonable. The note as a whole is about someone else's feast.
This is not a complaint about fairness
These patterns are described here as limits that show up in the output, so that you know where to look. Whether the way such tools are built is just, who is left out, and what should be done about it are questions of ethics, and they belong to Christian Ethics of AI (STEL5E). This lecture only asks you to notice what is missing from an answer.
4. What to do with these limits
Three questions, asked before you rely on an answer, cover most of the ground.
- Does it depend on the present? Names of office-holders, schedules, rules, prices, recent events, current teaching documents. If so, check a present-day source.
- Does it depend on something local or small? A parish's own practice, a family of churches, a regional saint, a language with few texts. If so, check with someone who belongs to it, or with the community's own records.
- Does it say "usually" or "many"? A general statement may be correct for the many and wrong for your people. Ask whether it fits your case.
These questions are the start of the checking plan built in lecture 2-4. For now it is enough to hold on to one idea: the tool's answers come from what was written, and what was written is old, incomplete and uneven.
A limit of this lecture. It does not say which topics or traditions are best or worst covered by any particular tool. That differs from tool to tool and changes over time, and it is not for a course to rank. It asks only that you ask the question for each use.
Diagram 1. Old, gappy, lopsided
Diagram 2. Three questions before relying on an answer
| Question | If yes | First move |
|---|---|---|
| Does it depend on the present? | The tool may speak from its training date | Check a present-day source |
| Does it depend on something local or small? | The tool may have flattened it | Ask a member of the community, or its own records |
| Does it say "usually" or "many"? | It may not fit your case | Ask whether it fits your people |
In short
- A tool's knowledge stops at its training date; the present is outside what it learned.
- A language model keeps no list of sources, so it cannot show where a sentence came from.
- Much of what people know was never written in a form the tool could learn: records, oral custom, small languages.
- What is common in the learning material is common in the output; small traditions and local practice are likely to be flattened.
- The ethical argument about who is left out belongs to Christian Ethics of AI (STEL5E); here it is only a limit to notice.
- Before relying on an answer, ask whether it depends on the present, on something local or small, or on a general "usually".
| Front | Back |
|---|---|
| Training date | The point at which the tool's learning stopped; it speaks as if the world stopped there too. |
| Why a tool cannot cite its source | It keeps no list of documents; what it learned left a mark, not a reference. |
| Missing sources | Books, records, oral custom and small languages that were never in the learning material. |
| Lopsided coverage | What is written about most is most likely to appear in answers. |
| Three questions | The present? Local or small? "Usually" or "many"? |
One sentence to remember The tool's answers come from what was written, and what was written is old, incomplete and uneven.
This week's action Pick one question about your own community's practice, ask yourself what an AI tool would probably say, and write down how it would differ from what your community does.
Check yourself
Not "did you read it" but "can you explain it". Ticks are kept on this device only.
Reflection and discussion
- What is a training date, and what kinds of question in your ministry depend on the present?
- Why is "What is your source?" an uncertain thing to ask a language model?
- Name one thing your community does that is unlikely to appear in writing. What might an AI tool say about it?
- What is the difference between a gap in the learning material and a false statement? Why is the first harder to notice?
- If a tool describes your tradition in a way you do not recognise, what could explain it, and what would you do?
- Which of the three questions at the end of the lecture would you ask most often in your own work, and why?