Courses · FAC 105 The ethical use of AI in teaching and research · Lesson 1 of 4
What the tools are
Lecture 1 of 4 · about 6 minutes
Before you begin, ask yourself. The lecture answers these questions, and you will meet them again at the end.
- How does a language model produce an answer?
- On which subjects are the tools weakest?
- May an AI tool be cited as a source?
What you will be able to do
- Say what a language model does.
- Name what it does well and how it fails.
- Keep the rule: never a source.
Unit 1 of 20 · the plan
- Before the session1½ h
- The lesson1½ h
- The case½ h
- The seminar1½ h
- The check¼ h
- Hand in¾ h
About6 hours
The artificial-intelligence tools that write, summarise and answer questions are built on large language models. To use one honestly you need to know, roughly, what it does.
A language model is trained on an enormous quantity of text. From it the model learns which words tend to follow which. When you type a question, it produces an answer one piece at a time, each piece being a likely continuation of what came before. It does not look anything up, unless it has been connected to a search tool. It does not hold a store of facts in the way a database does. It produces text that resembles the text it learned from. This is why its writing is fluent, and why fluency tells you nothing about truth.
Such a model does several things well. It drafts quickly. It summarises a long document you give it. It explains a familiar idea at whatever level you ask for. It rewrites clumsy prose, translates between major languages, and suggests an outline or a set of questions. Used as an assistant by someone who knows the subject, it saves time.
It fails in ways that are just as regular. It states falsehoods with complete confidence, because a confident sentence is as easy for it to produce as a true one. The usual name for this is hallucination. It invents sources: authors, titles, page numbers, court cases and quotations that look exactly right and do not exist. It is weakest on precisely what this University teaches: small nations, local history, unpublished records, and anything about which little was written. Asked about the Mohawk of the Grand River, a tool will often blend several communities into one, repeat a common error, or fill a gap with something plausible.
Pause. Look away from the page and say in your own words what “Large language model” means. Then read on.
It reflects its sources. A model trained mostly on what settlers wrote about Indigenous peoples will reproduce those writers’ assumptions as though they were neutral. It has no means of knowing what a people knows about itself and never wrote down.
It also handles languages unequally. In English it is strong. In Kanien’kéha, where little text exists, it will produce words that look right and are not words. That is worse than useless, since a learner cannot tell.
Two more facts complete the picture. What you type into a tool may be kept by the company that runs it and used to train later models. And the tools change every few months, so that anything said about their abilities will date. Their nature will not change.
From all this comes the rule the University teaches, printed below. A tool can help you find and phrase. It can never be what you rely on. The thing you cite, quote or teach from is always a source that a person has read.
The source
AI tools are never a source; they may point to sources.The University’s rule on AI as a source · The Master Content Card of this course, Mohawk University.
This lesson as a unit
Before the session1½ hours
Read. UNESCO, Guidance for Generative AI in Education and Research (2023).
Bring 150 words. Why does fluency tell you nothing about truth?
The case: The term file½ hour
The term file. A made-up instructor’s first term teaching one made-up online class, from its policy on tools to its safety page. Open the file. Name the class and its subject; enter where a tool would help the instructor and where it would mislead the class.
The seminar1½ hours, live
The room asks a tool three questions of local history and checks each answer.
Hand in: casebook entry 1 · What the tools are45 minutes
The class, and the tools. Marked to a published guide, with written feedback in three parts: what the evidence supports, what is still assumed, and what to do next.
Key terms
- Large language model
- A program trained on text to produce likely continuations of what it is given.
- Hallucination
- A tool’s confident statement of something false.
- Training data
- The text a model learned from, whose gaps and biases it reflects.
The questions of this lesson
- Choose oneHow does a language model produce an answer?
- By generating likely continuations of text
- By looking facts up in a database
- By asking an expert
- Choose oneOn which subjects are the tools weakest?
- Common English grammar
- Summarising a document you supply
- Small nations and local, little-written history
- Choose oneMay an AI tool be cited as a source?
- Yes, if it sounds sure
- No; it may only point to sources
- Yes, with a date
They are answered and marked in My Register, where this sample lesson is open to every member: there it asks 6 questions in all, in two ways, with how sure you are and what is still unclear. No answers are published on this site.
Summary
Go back to the questions at the top. You should now be able to answer each in a sentence, and to explain these terms: Large language model, Hallucination, Training data.
Sample questions of the test →
This was the sample lesson
The ethical use of AI in teaching and research has three more lessons, each with its reading, its terms and its self-check, and a test drawn from the whole course. They are for enrolled learners, in My Register.
The first lesson is open to everyone. The full course is $60, and free for members of the Nation. What it costs