The words spoken in the Unit 9 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 9 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 9 — Grounding and citation checking
Narrator: On the routes with the mathematics, Units 6, 7 and 8 build a reference by hand, compare value by value, and stop at the first divergence. That works there because the case has a definite answer. If your route skipped them, you meet the method here.
Most things are not like that
An oracle is a source of correct answers you can trust independently of the thing being checked. Most of what you will actually check has none: there is no computable answer. So this unit runs the same method on a case that has none — and the point is that the method comes through intact.
The same shape
In the Bernoulli case you asked for a routine, a short program. Here you ask for a summary with references. There, you derived a table by hand. Here, the reference is what the sources actually say. There, you compared value against value. Here, you compare claim against source. And instead of the first index that disagrees, you look for the first claim the source does not support. Build the reference. Compare item by item. Stop at the first divergence. Name the fault. Say what you can defend. This is not a lighter version of the lab. Building the reference is still the work. What changes is that the judgements at the end are yours, because no program in this course makes them.
What a citation establishes
A citation produced alongside a claim is more generated text. Same output, same process, no more evidence attached than the sentence it follows.
Questioner: Then why ask for a citation at all?
Narrator: It is still the most useful sentence in the output. Not because it is reliable, but because it is the one part that is cheap to prove wrong, if it is wrong. What it gives you depends on what comes with it. There are three cases, from least to most: an unbounded claim, a claim with a locator, and a quotation with its locator. Each leaves you a different amount of checking.
[On screen: A card at the top reads 'A claim', attached: no citation. Below it a field of thirty-five schematic documents is enclosed by a dashed search-area boundary, and the label reads: it leaves you a research task, the search area is everything. A locator is attached: the boundary shrinks onto one document, which opens as a page beside the field, and the boundary settles on one place in it: a bounded task, which source to open and where to look. The quotation with its locator is attached: the boundary shrinks to one line, which turns green: a certificate, one sentence to read. Two results appear: the source says this, checked; the source is right, not shown. Then: relied on? a separate check, and the source is a poor place to make it. Finally arrows leave the chosen document for three other documents in the field, and the label reads: whether to rely on it, lateral reading: read what other sources say about who stands behind it. A note says the sources are schematic and no real source is shown.]
An unbounded claim leaves you a research task. A claim with a locator leaves you a bounded one: which source to open, and where to look. Ask for the quotation with its locator, and you have a certificate: the exact sentence the claim rests on, and where it is. Checking it means reading one sentence. And a checked quotation certifies that the source says this, not that the source is right. Whether the source deserves to be relied on is a separate check, and the source is a poor place to make it. Professional fact checkers leave it, and read what other sources say about who stands behind it. That is called lateral reading. None of the six failures that follow is new. People misattribute quotations, repeat superseded figures and turn associations into causes. What generation changes is the volume: eight referenced claims in seconds, and still an afternoon to check them.
Pictures and recordings
Everything so far applies when the evidence is a picture, a recording or a screenshot, with one difference. A picture looks like a record of what happened, and it may not be one. In a 2026 study, reviewers with PhD-level training were shown research images altered by a generative model to change the result, and spotted them no better than chance. The best automated detector tested did better, but was far from reliable. So treat an image as a claim about where it came from, and check it the way you check a citation. Find where it was first published, and what else records the same thing. Looking right proves nothing, and a missing label proves nothing either. And if an image, a recording or a voice message asks you to act on something that matters, confirm it through a channel you found yourself, not one the message supplies. Tracing one is lateral reading applied to a picture. Trace it in four steps. First, search by the image itself, with more than one engine. The earliest copy with a named source is a lead, not yet the origin. Second, read its caption, credit, date and any file details as claims, since all of them can be edited. Third, find an independent record of the same thing: another photograph, a report from someone who was there, the original paper and its data. Fourth, write down what you could and could not trace. Untraced is a finding. It is not the same as generated.
Three questions
Three questions, and they get very unequal attention. Does the work exist? Does the locator exist? And does it support this claim, as stated, at this strength? The first is the cheapest. A citation can pass the first two and still fail the third, and the third is the one that costs actual reading. A summary can cite a real paper, at a real page, that genuinely discusses the topic -- and still misstate what it found. Every surface signal points the wrong way, including the first two answers being right. And the first question, cheap as it is, cannot be assumed to have been asked. References to works that do not exist have been found in papers that passed peer review and were published. At one large machine-learning conference, a scan found one hundred fabricated citations in fifty-three accepted papers. In one audit of biomedical papers, about one paper in two hundred and seventy-seven from early twenty twenty-six had a fabricated reference. Those figures measure what reached print, not how often a system invents a reference. What they show is that publication does not guarantee a cited work exists. Ask the first question every time.
Six codes
A closed list, exactly like the failure codes in the lab. Use these names and no synonyms. Several can apply to one claim. So work down the list in order. The first one that applies is the diagnosis. From the top, the six are: fabricated locator. Stale version. Misattributed. Over-specified. Unsupported inference. And right source, wrong claim. The order runs from the cheapest check to the most expensive. First, whether the page exists. Then which edition. Then which work. Only then, what the passage means. Two pairs are close, and the closeness is the work. Unsupported inference and over-specified are both "the source says less" — but one strengthens the claim and the other sharpens a number. Misattributed and right source, wrong claim can look alike. Misattributed: the cited work does not contain the finding, but another work does. Right source, wrong claim: the cited work is the right place to look, but it says something different. That one is harder to catch, because everything looks fine until you read the passage.
No program does this
Questioner: Could a program not check the citations for us?
Narrator: The lab's checker names a failure by re-running the code against a reference it can compute. Nothing equivalent exists here, and this lab does not automate these classifications. Whether a source supports a claim is a judgement about meaning. Automated tools may assist with parts of it, and their scope and errors need checking. A checker that claimed to settle it outright would be making exactly the kind of overreach this course trains you to catch. So the discipline is not automation. It is the closed list. Having to choose between two codes forces you to say what kind of wrong you found. "The citation was bad" says nothing, and points to no repair.
Build the reference first
The step people skip. On the routes with the mathematics it is the hand derivation in Unit 6; here it is reading the sources before the summary. For each source: what you could retrieve, what the passage at the cited place actually says, and what it supports — at the strength it supports it. And note any limit on it: a scope, a version, a passage that replaces another. That is where two of the six codes are caught. And record what you could not retrieve. A paywall, a dead link, an edition you cannot get. Those are findings. They change what you can claim, and they do not invalidate the rest.
The first divergence
[On screen: Headed 'Compare in order, and stop at the first one that fails', under a tag marking the specimen and its sources as fictional, constructed for this exercise. Three column headings appear one at a time: generated claim on the left, what the cited source says on the right, and supported? at the end. Then the rows fill in, one cell at a time. Claim 1, that the kerbside scheme began in 2011, citing Council 2019 section 1.1, agrees with Source A section 1.1: supported, yes. Claim 2, that participation increased by 19.6 percentage points, 2012 to 2018, citing Hallam 2021 page 94, does not: Source B page 94, quoted in part, says '...increased by approximately 20 percentage points between 2012 ...'. That row turns red and is marked no, over-specified, and a line below reads: first divergence at claim 2, over-specified. Closing line: stop at the first one that fails, and name its code.]
Take the claims in order. Stop at the first one the passage does not support, and name its code. The left column holds each claim from the summary, with the source it cites. The middle column holds what that source says at that place. The last column records whether the source supports the claim. The first row is claim one: the kerbside scheme began in 2011. On the right, Source A, section one point one, says the scheme began in 2011. It is the same fact at the same strength, so the answer is yes, and you go on to the next claim. The second row is claim two: participation increased by nineteen point six percentage points between 2012 and 2018. On the right, Source B, page 94, says approximately twenty. The two no longer say the same thing, so the answer is no. It is the first divergence, so you stop here and name its code: over-specified. Why that code, and not another, needs a closer look.
[On screen: Headed 'One claim, read against its source', with a subtitle that ends on 'Name the failure, and say what it is not.' The claim states that participation increased by 19.6 percentage points between 2012 and 2018, citing Hallam 2021 page 94. Below it, the passage it cites, Source B page 94, tagged fictional and constructed for this exercise, reads: in Brackenford, participation increased by approximately 20 percentage points between 2012 and 2018. An amber box marks 19.6 percentage points in the claim and approximately 20 percentage points in the passage, joined by a dashed line. A green panel headed 'These agree' says: same population, same period, same direction, same scope; the work exists, the page is right, the passage is there. An amber panel headed 'Over-specified' says: only the precision is invented; everything else about it is right. Closing line: a citation that resolves, at the right page, to a passage that exists, and the claim still fails.]
Here it is claim two. Same population, same period, same direction and scope; only the precision is invented: nineteen point six against approximately twenty. Over-specified. Later claims may look fine. Their looking fine tells you very little, once you know the summary can produce an unsupported claim in this register. But stopping at the first divergence stops the diagnosis, not the work. You still check the rest before you rely on any of it.
What you can say
Not "the summary is accurate", which nothing you did establishes. For another summary, you might write: of eight claims, I retrieved sources for six. Five are supported as stated. One cites a section that does not exist, though the claim itself is supported elsewhere in the same document. Two I could not check. Every part of that can be interrogated. "I verified the references" cannot.
Next
The activity gives you a complete specimen — a summary and the sources it cites — so you can do a full comparison without leaving the page. The sources are invented for the exercise and should never be quoted anywhere else. Six of the eight claims fail. Finding them is the easy part. Saying which code is the work. Then take the three checkpoints.
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.