The words spoken in the Unit 1 video, scene by scene, as its captions show them, with what the animations show marked [On screen: ...] where they play. Each speaker is named where the speaker changes. The Unit 1 reading covers the same material.
The narration is spoken by synthetic voices made for this course, each shown on screen as a drawn robot. No person is speaking.
Unit 1 — What generative AI is for
Narrator: Before you use one of these systems for something that matters, there is a question worth asking. Not "is it any good at this". A different one.
The question that dates
"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. So this unit asks something else, and the answer does not move.
The question that does not
For this task, where would you find evidence that the output is right? That question has only a few possible answers. They are the useful way to sort the work, because they shape what checking costs. And checking is the cost most people forget to count before they start.
Two pairs of tasks
Translating a paragraph and reformatting a table look nothing alike. They are the same class of work, because in both cases you already hold the thing you would check against.
Questioner: And the second pair? Are those two the same kind of task?
Narrator: These two look almost identical, and they are not the same class at all. In one case the document is in front of you. In the other you would have to go and find it. That difference can cost you an afternoon.
Five classes
Five classes, and they differ in one thing only. Not difficulty, not importance, not subject. Where the reference lives. Read that column downwards and you are reading, typically, how much work you have left to do after the output arrives.
[On screen: Headed 'The class does not change the output. It changes the distance.', one box labelled the output stands alone, captioned: same output, the question is what you would check it against. Then the reference it would be checked against appears and moves, one task class at a time, each with its caption. Transformation: the reference is your own input, right beside it; the reference is already in your hand. Grounded answering: the source you supplied; a step away, the source you gave it. Generation to a specification: the specification you wrote; a little further, check it against what you asked for. Recall: a source you must find, furthest away; find a source, open it, read the right part. Judgement: no single reference to travel to; instead three separate things fan out from the output, its factual premises, the stated criteria and the trade-offs named, because there is no one answer to compare against but there are still things to check.]
On screen, the output stays put and the reference moves. Beside it for a transformation; a step away for a supplied source; a little further for a specification you wrote; furthest for recall. For judgement no single reference appears, but its premises, its criteria and its trade-offs can still be checked.
Transformation
You hand over the content and ask for it in another form. The reference is what you handed over, so the check is something you can do at your desk with nothing else open. The failure to look for is not invention. It is omission — something that went in and did not come out. That is easy to miss, because what comes back reads as complete.
Checker: The table in the last scene was made from a four-column table on this unit's page. Compared with it row by row, all five classes came through. Two columns did not: what you supply, and what checking costs. Nothing in the shorter table shows that they are missing.
Grounded answering
Narrator: Attach a document and ask a question about it, and the check is bounded: it has a clear limit. Read the source at the point the answer relies on.
Questioner: Does attaching a source remove the check?
Narrator: Attaching a source changes where you aim the check. It does not remove it. Two questions stay live. Does the source say what the answer says it says. And is it the right source to settle this. 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. Then an answer can rest on pages it fetched, not on ones you chose. So record whether it searched. Then open each link it attaches. A link claims two things: that the source exists, and that it supports the answer. Check both.
Generation to a specification
State what would count as correct, and ask for something that meets it. This is the class the middle of this course lives in. The failure to look for is particular: the output meets the specification you wrote, rather than the one you meant.
Recall
Ask a question with nothing attached and the answer arrives with no evidence attached either, and no marker separating the claims that happen to be right from the ones that do not. This is where the gap is widest. Eight referenced claims can be produced in seconds and take you an afternoon to check. It shows in print. References to works that do not exist have been found in published, peer-reviewed papers, and one audit of early 2026 found fabricated references in about one biomedical paper in 277. Publication does not guarantee that a cited work exists, and checking that it does is the cheapest check there is.
Judgement
Which design is better. Whether an argument convinces. Whether the writing is any good. There is no single answer to be right about, so what comes back is a proposal. There are still things to check. Its factual premises can be wrong, and it can be assessed against criteria you state. What you cannot do is settle it by comparison. A proposal can be genuinely useful: a critique can name something you had missed. Treating a proposal as a finding is the error. And a confident tone is a property of the text, not evidence about the question.
The move
Here is the part worth taking away. The class is not a fixed property of the task. It is usually your choice. "What does the regulation require" is recall, and expensive. "Here is the regulation — what does clause four require" is grounded answering, and the check is reading clause four. Same question. Same system. Minutes instead of an afternoon, because you moved yourself into a cheaper class before you started. The same move works forwards. A vague request becomes generation to a specification the moment you write down what would count as correct. That is Unit 10. And inside any class, ask for evidence you can check: the sentence of clause four it relies on, or tests you can run. Unit 0 called that a certificate. It is checked against the source, not taken on trust.
One task, four classes
[On screen: Headed 'One task, four classes', a fixed question sits at the top: 'Does our coursework policy allow me to use an AI system to draft an essay plan?' Below it a box labelled the answer stays in place while a reference box moves. Asked with nothing attached, it is recall: the reference is the policy, which you must go and find, and it moves furthest away; to check it you must find the policy, the whole task you were trying to avoid; what checking costs: high, the check is a separate research task. Asked with the policy pasted in, it is grounded answering: the reference moves in to the clause the answer relies on, in the policy you pasted, a step away; find the clause and read it, a minute, not an afternoon; cost: low to moderate, read the source at the cited point. Asked for a table of every clause that mentions AI, it is transformation: the reference is the document you supplied, right beside the answer; every row is checkable against the document, look for a clause left out; cost: low, compare output against input. Asked whether it is a good policy, it is judgement: no reference appears, because no clause settles it; what comes back may be worth reading and is not a finding; cost: check the factual premises and assess the proposal against stated criteria. A row of four labels along the bottom lights up as each way of asking is reached, and a closing line says the difference between an afternoon of checking and a minute of it is the class you chose. Then the question changes to 'What is B8?'. Asked with nothing attached it is recall, and the reference, a source for B8, moves furthest away. Asked with the recurrence handed over, it is generation to a specification: the reference moves in to the recurrence, the specification you gave, while a dashed outline marks where the recall reference had been; cost: moderate, test the output against the specification.]
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. 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. Find the clause the answer relies on and read it. You may still find a clause about brainstorming stretched into one about drafting, but in a minute rather than an afternoon. Ask instead for a table of every clause that mentions AI, with its number and its exact wording, and you are in transformation. Every row is checkable against the document, and the failure to look for is a clause that quietly did not make it into the table. Ask whether it is a good policy and you are in judgement. No clause settles it. What comes back may be worth reading, and it 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 holds for this course's case. It uses a sequence called the Bernoulli numbers, and B-eight is one of them. Ask for B-eight with nothing attached, and it is recall. Hand over the recurrence, the rule that builds each number from the ones before, and ask for the computation. Then it is generation to a specification.
What they are not for
Not a list of forbidden tasks. One question: if this is wrong, how would I find out, and what would that cost? Three answers that should stop you. There is no reference obtainable at a price you can pay. You would not recognise a wrong answer — which is about you rather than the system, and is the sharpest of the three. Or the doing was the point, and an output that arrives without the work has skipped the thing the work was for: a correct answer and having learned to produce one are different results.
What the classes cost
Questioner: So should I avoid recall and judgement?
Narrator: Recall and judgement usually cost the most to check, which makes them the classes you owe the most checking — not the ones to avoid. And the cheap end is not the safe end. A transformation can drop the one qualification that mattered, and it will do so fluently. These are typical costs, not fixed prices. A translation into a language you cannot read is costly to check, and so is a long document you have not read. A recalled date can take seconds.
Next
The activity asks you for five tasks of your own, sorted, with the reference named for each. Then it asks you to take the most expensive one and move it. The row where you write "I do not know where I would check this" is the most useful thing you will produce.
Closing, 20 seconds, no speech. A candidate record on the left, a check against an independent reference in the middle, and a bracketed claim on the right fade in side by side. Below them: Continue with the course material, ai.qyber.dev/edu-ai-literacy-ada-verification/, Course playlist: youtube.com/playlist?list=PLPsBWTfgvakM, All links are also in the description, and Keep your record. Check the next claim.