Unit 1: What generative AI is for

Reading time: about 12 minutes. Activity: about 12 minutes.

1. The question worth asking

"What is generative AI good at?" is a question that goes stale. Answer it with a list of products and the list is wrong within a year. Answer it with a list of tasks and you have described what the systems could do on the day you wrote it.

This unit answers a different question, one whose answer does not move:

For this task, where would you find evidence that the output is right?

That question has only a few possible answers, and they are the useful way to sort the work. Two tasks that look nothing alike — translating a paragraph and reformatting a table — sit in the same class, because in both cases you already hold the thing you would check against. Two tasks that look almost identical — "summarise this document" and "summarise the research on this topic" — sit in different classes, because in one case the document is in front of you and in the other you would have to go and find it.

The class identifies the checking job, and that sets a typical cost, not a fixed price. That is the whole of this unit.

2. Five classes, by where the reference lives

A reference is whatever you would compare the output against to find out whether it is right. On a route that includes them, Unit 6 has you build one by hand and Unit 8 has you use it. Here we are only asking where it comes from.

Read the last column downwards. The column does not rank usefulness. It shows how much work is typically left after the output arrives, and that is the cost most people forget to count before they start. The order is typical, not fixed: a translation into a language you cannot read is expensive to check, a recalled date may take seconds, and a complex specification can be very costly.

Transformation

You supply the content and ask for it in a different form: translated, summarised, reformatted, restructured, rephrased for a different reader, turned from prose into a table.

The reference is your input. Every claim in the output should be traceable to something you handed over, so the check is a comparison you can make at your desk with no other material. The characteristic failure is not invention but omission: something in the input that did not survive the transformation. That is easy to miss precisely because the output reads as complete. And the check is cheap only if you know the input: for a long document you have not read, it costs about as much as reading the document.

Grounded answering

You supply a source — a document, a page, a specification — and ask a question about it. Some systems fetch the source themselves rather than being handed it [L3, sec. 2].

The reference is that source. The check is to read it at the point the answer relies on. Note what has and has not changed: attaching a source changes where you aim the check, not whether one is needed. Two questions remain live. Does the source say what the answer says it says? And is the source the right one to settle this question?

The class can also change without you. As of November 2026, one provider's documentation says its assistant may search the web without being asked [D4, opening note]. Then an answer can rest on pages it fetched, not on ones you chose. So record whether it searched, and open each link it attaches. A link claims two things: that the source exists, and that it supports the answer. Check both (Unit 9).

Generation to a specification

You state what would count as correct, and ask for something that meets it. A routine that satisfies a stated contract; a document in a required structure; a test that fails under a named condition.

The reference is the specification. This is the class the middle of this course lives in: Unit 7 asks for a routine, Unit 8 checks it against a reference you built, and Unit 10 is about what happens to the output when you make the specification more precise. The characteristic failure is that the output meets the specification you wrote rather than the one you meant.

Recall

You ask a question with no source attached and no specification: a date, an attribution, a requirement, a figure, who first proved something.

The reference is external and you do not have it. The system does not look anything up in a record of what is true (Unit 3 explains why). So its answer to a recall question comes with no evidence, and nothing marks which claims happen to be right. Checking means finding an authoritative source, which is a research task of its own. Unit 9 is about doing that properly.

This is the class where the gap between how much a claim costs to produce and how much it costs to check is widest. Eight referenced claims can be generated in seconds and take an afternoon to verify. The gap shows in print: references to works that do not exist have been found in published, peer-reviewed papers [L10, secs. 2.1, 4.1], in about one biomedical paper in 277 in one audit of early 2026 [L11]. Publication does not guarantee that a cited work exists, and checking that it does is the cheapest check there is.

Judgement

You ask which of two designs is better, whether an argument is convincing, whether a piece of writing is any good.

There is no reference in the sense the other classes have one, because there is no single answer to be right about. That does not mean there is nothing to check. A judgement rests on factual premises, and those can be wrong: an argument that one design is cheaper depends on a claim about cost. It can also be assessed against things you can state — the objectives, the constraints, the evidence offered, the trade-offs, professional standards, whose interests are served. What you cannot do is settle it by comparison, because there is nothing to compare it against. So the output is a proposal, and treating a proposal as a finding is the error. A proposal can be genuinely useful: a critique can name something you had missed. A confident tone here is a property of the text, not evidence about the question.

How this relates to Unit 3

Unit 3 sorts outputs a second way, which it calls triage. It asks a second question about the same task: not where the reference lives but what kind of evidence settles it — something you can execute and compare, something you must look up in a record, or neither because an error would be cheap and visible.

The two questions do not compete: usually the answer to the first gives you the answer to the second. A task with its reference in a specification is one you can check by executing and comparing. A task whose reference is a source, whether you supplied it or must go and find it, is one you check by reading the record. Judgement fits neither, and it is not low-risk drafting either: its premises are checked like any other claim, and its conclusion stays a proposal.

So this unit gives you the cheaper question — where would the evidence come from — and Unit 3 gives you the sharper one, which is what to do about it.

3. The move worth learning

The classes are not fixed properties of a task. Which one you are in is often your choice, and moving between them is the most useful thing in this unit.

A recall question becomes a grounded question the moment you supply the source. "What does the regulation require?" is recall, and expensive to check. "Here is the regulation — what does clause 4 require?" is grounded answering, and the check is reading clause 4.

The same move works forwards. A vague request becomes generation to a specification the moment you write down what would count as correct, which is exactly what Unit 10 asks you to do, and why it is the unit that changes how much the rest of your work costs.

There is a third move, and it applies inside any class: ask for the output to carry evidence you can check, not just an answer. "Quote the sentence of clause 4 you are relying on" instead of "what does clause 4 require?"; "give me the code and tests I can run" instead of "what is the result?". The check then shrinks to confirming that evidence against the source you hold, which is quicker than doing the work. Unit 0 called such evidence a certificate. What the system supplies can itself be fabricated, a quotation that is not there or a test that tests the wrong thing, so it is checked against the source, not taken on trust.

So the practical question is not "can it do this?" but "which class am I in, and can I move myself into a cheaper one before I start?"

4. One task, four classes

Take a single question: does our coursework policy allow me to use an AI system to draft an essay plan?

Asked with nothing attached, it is recall. The answer will name a policy, may quote it, and will read as settled. Nothing in it is evidence: to check it you must find the policy, which is the whole task you were trying to avoid.

Paste the policy in and ask the same question, and it is grounded answering. The check is now bounded, which means it has a clear limit: find the clause the answer relies on and read it. You may still find that the answer stretched a clause about brainstorming into one about drafting, but you will find it in a minute rather than an afternoon.

Ask instead for a table of every clause that mentions AI, one row per clause with its number and its exact wording, and you are in transformation. Every row is checkable against the document you supplied, and the failure to look for is a clause that quietly did not make it into the table.

Ask whether the policy is a good policy and you are in judgement. There is no clause that settles it. What comes back may be worth reading and is not a finding.

Four classes, one question, and the difference between an afternoon of checking and a minute of it is which one you chose. The same applies to the case this course is built on, a sequence called the Bernoulli numbers. "What is B8?" (the one with index 8) is recall, and "here is the recurrence, the rule that builds each number from the ones before; compute it" is generation to a specification. Unit 6 has you build the reference you will use to check the output.

5. What these systems are not for

Not a list of forbidden tasks. A question:

If this output is wrong, how would I find out, and what would that cost?

Three situations where the honest answer is bad.

No reference is obtainable at a price you can pay. Questions about the future, about private matters of fact, or at the edge of what is known. There is nothing to check against, so any output is a proposal however it is phrased.

You would not recognise a wrong answer. This is the sharpest one, and it is uncomfortable because it is about you rather than the system. In a subject you do not know well, the signs you usually trust (smooth writing, precise detail, a confident structure) appear in wrong answers as easily as in right ones. The move here is to change class: get a source you can check, or ask someone who can.

The doing is the point. Some work exists to develop your judgement, and an output that arrives without the work has skipped the thing the work was for. That is a statement about what the task is, not a rule about tools: a correct answer and having learned to produce one are different results, and Unit 10 shows the evidence. Your institution's rules on assessed work are separate, and they are theirs to state, not this course's.

Summary

• Sort a task by where its reference lives, not by topic or by product. That sorting does not date.
• Transformation and grounded answering put the reference in your hands. Generation to a specification puts it in something you write. Recall puts it somewhere you must go and find. Judgement has no single one, though its premises and criteria can be checked.
• The class sets a typical cost of checking, not a fixed price, and most people forget to count that cost before starting. Recall and judgement are usually the expensive classes, which makes them the ones you owe the most checking — not the ones to avoid.
• Cheap to check is not the same as checked. A transformation can drop the one qualification that mattered, and it will do so fluently.
• You can often move a task into a cheaper class by supplying a source or writing a specification. That is the single most useful habit in this course.
• The limits are not a list of banned tasks but one question: if this is wrong, how would I find out, and what would that cost?

Then answer cp01-where-the-reference-lives and cp01-changing-class at the foot of this page.