The words spoken in the Unit 4 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 4 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 4 — Choosing the tool
Narrator: Unit 3 said what a language model does. Unit 1 sorted tasks by where the reference lives. This unit puts them together and asks a practical question.
The claim
Deployed systems differ in which layers they have around the base model. Every layer you add changes what you have to check. Not by removing a check. By moving it somewhere else, and adding one of its own.
[On screen: One question, what would I have to look at to find out whether this answer is right, stays fixed at the top while four kinds of system are tried below it, one row each, against five places a check can land: the claim itself, the cited source, the inputs, the result and the media itself. A magnifier marks each place you have to look. For plain generation, with no source offered, the check lands on the claim itself: find a record and check it. For a retrieval-backed system, which fetches sources, it moves to the cited source: does it say this, and is it the right source? For a tool-calling system, which runs code, it splits in two, onto the inputs (was it asked the right thing?) and the result (did it come back right?). For a multimodal system, which takes an image, audio or video, it moves onto that media itself, which has no search field: look or listen yourself, by eye or by ear. Every row keeps its check; the closing line says the check moved each time and was never removed.]
Watch where the check goes as one question meets four kinds of system. Plain generation: the check goes onto the claim. Retrieval-backed: it goes onto the cited source. Tool-calling: it splits between the inputs and the result. Multimodal: it goes onto the image, the audio or the video itself, checked by eye or by ear.
Four kinds
Four kinds, named by what they add. Not four products — a product can be several of these at once and can change which it is between releases. Ask what the system is doing on this occasion, not what it is called. The right-hand column is the new failure each one brings, and the scenes that follow take them in turn.
Plain generation
A base model with post-training and nothing else. Everything in the answer is generated, so a recall task stays a recall task — the reference is somewhere else and you must go and get it. There is nothing wrong with this. It is the right choice when you supplied the content, and it is usually fastest and cheapest. What it is not is a way of finding things out.
Retrieval-backed
The system searches, or is handed documents, and answers from what came back. This is a real difference, and it is the most useful single upgrade for a recall task. It splits the check into two questions. Does the source say what the answer says it says. And is it the right source to settle this. The second one is the one people skip, because a citation looks like the end of an argument rather than the start of one. Whether it searched is not always your choice. As of November 2026, one provider's documentation says its assistant may search the web without being asked. The same product can be plain generation on one occasion and retrieval-backed on the next. So record whether it searched: the checks you need differ.
Research modes
[On screen: Headed 'Research modes [D5]'. A row of boxes builds from left to right: your question, a plan, many searches, a long report. Below them is a grid of small cells labelled 'the field of possible sources (illustrative)'. Ten searches run one after another, each lighting a cluster of cells, mostly on the left of the field; the right side is never searched. Most of what was found then turns grey as dropped, and the long report appears with thin lines linking it to only the kept cells, forty of them. A caption says one provider described its version this way in December 2024 [D5]. An arrow loops from the searches back to the plan, labelled 'retrieval, run in a long loop', and the report is marked: 40 citations, 40 checks, each asking whether the source says it and whether it is the right source. A caption asks what was searched, and what was left out, and the legend for dropped and never searched lights up. Captions say that which searches ran, which sources were kept and when it stopped are not shown by its citations. Two cited cells are then checked and turn green: they hold. The dropped and never-searched cells stay grey, and the caption says those two checks tell you nothing about them. Last, a dashed box reads 'Compare with a source you chose yourself: a review article, a library database, your reading list', and rings mark what that source points to across the field, some on cited cells, some on dropped ones and some in the never-searched part.]
Some assistants offer a research mode. It turns your question into a plan, runs many searches one after another, and returns a long report with links to its sources. One provider described its version this way in December 2024. It is retrieval run in a long loop, and the volume is the cost. Forty citations owe forty checks of the two kinds you have just seen. And it raises a question no single citation can answer. What was searched, and what was left out? You may be able to revise its plan, but which searches it ran, which sources it kept and when it stopped are not shown by its citations. Checking two of its citations tells you those two are sound. It tells you nothing about whether the report covers the field, or only the part that was easiest to find. So before you rely on one for an essay or a literature review, compare with a source you chose yourself: a review article, a library database, your reading list. And say what remains uncertain. In research of your own, there may be nothing to look up at all. Then the check is the research itself: a derivation, a computation, an experiment. What any system gives you there is a candidate, a hypothesis to test, and testing it is part of the task, not an extra.
Tool-calling
The system calls a calculator, or writes and runs code, and uses the result.
Questioner: If it used a calculator, is the answer right?
Narrator: The computation is deterministic and either right or wrong in the ordinary way. That half is a genuine gain. The other half is not. Which numbers went in, and what question was actually computed: those were generated. So the failure to look for is a correct computation of the wrong question: the right formula on the wrong column, or a currency conversion at an unstated date. It arrives carrying the authority of arithmetic. And code it wrote can itself be wrong. It runs the same way every time and still gives the wrong number, which is what Units 7 and 8 are about.
Multimodal
When the source is a document you can find the sentence the answer rests on. When it is a photograph, a chart, or a recording of sound or video, search may need a transcript or an image-search tool, and whatever it finds still has to be checked against the original: by eye or by ear. A description of an image or a recording is generated text about it. It can be detailed, written in a confident tone, and about something that is not in the frame or on the recording. An invented citation can be looked up in a catalogue and found missing. For an invented detail there is no catalogue: you have to look or listen yourself. A picture or a recording looks like a record of what happened, and it may have been generated or altered. Treat it as a claim about where it came from: find where it was first published, and what else records the same thing.
The decision aid
Five steps. One: name the class of the task, as in Unit 1. Where would the reference come from? Two: if it is recall, try to move it. Can you supply the source? Three: ask what the answer will contain. Fetched sources, a computed number, a reading of an image or a recording. Each of those is a check you take on. Four: can I do that check? Five: take the cheapest kind that leaves you a check you can do. The step that does the work is step four: can you do the check? Not "is the system good". Can you, with the time and the access you have, carry out the check its output will need. If not, you have picked the wrong kind, or the wrong task. That check is what makes the system usable at all. And doing it is also how you find out whether using the system paid off: whether it saved you time, found what you needed, or did the task better than another way would have. A system that produces claims you cannot check is not more useful than one that produces fewer claims you can. Makers publish scores on standard tests, called benchmarks. A published benchmark score does not answer step four either. It means little if the test questions were in the model's training data. And even a fair score tells you about the test's tasks, not about yours.
What it costs
There are three costs: money, time and energy. They often rise together, because they have the same four causes, shown on the screen. They are a longer prompt, a longer answer, a bigger model, and more steps. Money may be priced by the text going in and coming out, and by the size of the model, and price lists change. Time has the same causes, plus each extra step, such as a search or a run of code. Energy follows the same quantities. Other things equal, more computation uses more resources. Money and time also depend on pricing, on hardware, and on caching, which reuses stored results.
The figure this course will not give
Questioner: How much energy does one query use?
Narrator: You will have seen figures for what one query costs in energy. This course does not print one.
Checker: Every page and narration script in the course was searched for a number with an energy or carbon unit beside it. There is none.
Narrator: A defensible figure would have to name the model, the hardware, the length of the query, whether the cost of training is shared across queries, and the electricity supply. And it would be out of date within months. A number without its conditions is the thing this course trains you to catch. So the course gives you the shape instead, which stays true. If you need a number, that is a recall task, and it needs a source that states its conditions and its date.
When cheap is right
Questioner: Why not always use the most capable system?
Narrator: The instinct to reach for the most capable thing available treats capability as free. It is not. It costs money, time and checking, because a more elaborate system produces more kinds of claim to check. The expensive option earns its cost when the extra layer removes a check you could not otherwise do. Not when it is simply there.
What this unit is not saying
Two things this unit is not saying. It is not saying that retrieval-backed and tool-calling systems are more trustworthy. They are checked in different ways. Each removes one kind of uncertainty and adds a different one. And it is not a recommendation of any product. Named products date, change what they are between releases, and are not the level at which this reasoning works.
Next
The activity asks you to work the five steps on one real task, then ask the same question of two kinds of system and compare what each gives you to work with. The interesting answer is not which is better written. It is what is in the second one that you can check.
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.