The words spoken in the Unit 13 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 13 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 13 — Privacy and data handling
Narrator: The usual advice is one paragraph: do not paste confidential or personal material into an AI system. That is sound, and it is not enough to act on.
Why one sentence is not enough
It does not say what happens to what you paste. Without that, you cannot judge any of the cases the sentence does not cover. And most real cases are cases it does not cover. What this unit will not do is state your institution's rules or the law. Those exist, they differ by place, and they are not this course's to invent. What it gives you is the shape, so that when you read your own rules you know what they are talking about.
What leaves
A hosted service is one you reach through a browser or an app. With a hosted service, what you supply leaves your device. That covers your prompt, the conversation so far, any file you attached, and anything a tool fetched on your behalf. The attachment and the conversation are the two that surprise people.
The attachment goes whole
[On screen: Headed 'You sent a question. This is what crossed.', a dashed line divides your device from the provider, and whoever it works with. You typed one question and attached one file, and the question crosses. More pieces follow that you did not think you were sending: the attached file, whole, not only the clause you asked about; inside it, what you did not ask about; and the earlier turns of the conversation. A conversation carries its history: re-sent by your device with every turn, or kept by the service and read back in. Either way the earlier turns are processed again with each new turn, and their counter climbs from used x1 to used x5. Then everything that crossed is dimmed where it is and marked 'already sent'. A new question, a rewritten clause with the identifying detail gone, crosses on its own; on your side, 'not sent this time: the file and the earlier turns'. Closing line: change what you send, and less of it crosses.]
You asked about one clause. The document went. And if the file has tracked changes, comments, hidden columns or a second worksheet, those are text too.
[On screen: Headed 'You sent a question. This is what crossed.', a dashed line divides your device from the provider, and whoever it works with. You typed one question and attached one file, and the question crosses. More pieces follow that you did not think you were sending: the attached file, whole, not only the clause you asked about; inside it, what you did not ask about; and the earlier turns of the conversation. A conversation carries its history: re-sent by your device with every turn, or kept by the service and read back in. Either way the earlier turns are processed again with each new turn, and their counter climbs from used x1 to used x5. Then everything that crossed is dimmed where it is and marked 'already sent'. A new question, a rewritten clause with the identifying detail gone, crosses on its own; on your side, 'not sent this time: the file and the earlier turns'. Closing line: change what you send, and less of it crosses.]
Checker: A text extraction includes them, and the service receives them.
Earlier turns keep going
[On screen: Headed 'You sent a question. This is what crossed.', a dashed line divides your device from the provider, and whoever it works with. You typed one question and attached one file, and the question crosses. More pieces follow that you did not think you were sending: the attached file, whole, not only the clause you asked about; inside it, what you did not ask about; and the earlier turns of the conversation. A conversation carries its history: re-sent by your device with every turn, or kept by the service and read back in. Either way the earlier turns are processed again with each new turn, and their counter climbs from used x1 to used x5. Then everything that crossed is dimmed where it is and marked 'already sent'. A new question, a rewritten clause with the identifying detail gone, crosses on its own; on your side, 'not sent this time: the file and the earlier turns'. Closing line: change what you send, and less of it crosses.]
Narrator: A conversation carries its history. Depending on the product, your device re-sends it with every turn, or the service keeps it and reads it back in, sometimes as a summary. Either way, something you pasted twenty messages ago is still in play: it is processed again with each new turn, not once. And deleting the conversation removes it from your view. What that does to stored copies, backups and logs is a question about the service, not about the button.
Three questions to go and answer
Retention depends on the service, the plan, and settings that change. So rather than facts that will be wrong next year, three questions you can go and answer. Is my input used to train future models? How long is it kept, and who can read it? And where is it processed, and under what terms? If you cannot answer all three for the service in front of you, that is itself the finding. Then be strict about what you send. The four bands later in this video show how. A product can also keep material by design. One provider's documentation, as of November twenty twenty-six, describes two such features. Saved memory and instructions carry into later conversations, and memory can be on by default. A shared link can be read by anyone who has it, and copies made from it survive after you withdraw the link. A connector to your mail or files sends material without you pasting anything. Find out which of these are on for you.
What can happen to a copy
Once your material reaches a service, there is a copy you do not control. A policy says what the provider intends to do with it, and what it commits to. It does not decide everything that can happen to it. It can be kept and read. A provider may keep copies, in logs, backups or for abuse monitoring, and its policy states for how long and who may read them. It can be learned from. Where input is used for training, a model can occasionally repeat what it was trained on. Researchers pulled hundreds of passages word for word out of one model, names, phone numbers and email addresses among them. It can be ordered kept. In twenty twenty-five a court required one provider to keep chat logs it would otherwise have deleted, for a lawsuit about copyright. It can be exposed by mistake. In twenty twenty-three a bug showed some users of one chat service the titles of other people's conversations, and some subscribers' billing details. And it can be taken on purpose. A person with access inside the provider, or at a contractor it works with, can misuse it, and an attacker who breaks in can copy it. That can happen wherever data is held. Some of that nobody at the provider intended: a bug, a log kept longer than the policy says, a court order. Some of it is somebody's deliberate misuse. None of it needs the provider to mean to break a promise: its terms are a promise about intent, not a guarantee of the outcome. So before you send something, ask whether it would matter if this copy got out. If it would, the classification decides, not the promise.
Classify by whose it is
Four bands, sorted by who is harmed if this becomes readable by someone else. Public is not a permission slip: published material still carries rights and terms, and a published page can still hold someone's sensitive details. Yours alone, with no other person's details in it, settles that nobody else's privacy is at stake. Not whether you are permitted. Someone else's is broader than it sounds. A message a friend sent. A colleague's draft. A photograph with a person in it. The test is not whether it is marked confidential. It is whether somebody gave you this expecting it to stay where you put it. What this band settles is that you must not paste it as it stands; the next slide is what to do instead. Not yours to decide is the band people miss entirely, because nothing about the material feels sensitive. A dull internal spreadsheet is often covered by an agreement somebody else signed. The decision is not yours: ask before, not after.
Change what you send
The move is almost never "do not use the tool". It is to change what you send. Replace names, dates, identifiers, distinctive details. The task usually survives untouched, because you were asking about structure, not about a person. Redaction reduces the risk; it does not make the use authorised. An organisation may forbid sending even de-identified material to a service it has not approved.
[On screen: Headed 'You sent a question. This is what crossed.', a dashed line divides your device from the provider, and whoever it works with. You typed one question and attached one file, and the question crosses. More pieces follow that you did not think you were sending: the attached file, whole, not only the clause you asked about; inside it, what you did not ask about; and the earlier turns of the conversation. A conversation carries its history: re-sent by your device with every turn, or kept by the service and read back in. Either way the earlier turns are processed again with each new turn, and their counter climbs from used x1 to used x5. Then everything that crossed is dimmed where it is and marked 'already sent'. A new question, a rewritten clause with the identifying detail gone, crosses on its own; on your side, 'not sent this time: the file and the earlier turns'. Closing line: change what you send, and less of it crosses.]
And you rarely need to send the document to get what you want. A rewritten clause with the identifying detail gone is a different question from pasting the contract, and it usually gets the same answer. Or bring the tool to the data: a local model or an approved service changes what leaves, though not always to nothing.
Removing a name is not removing a person
[On screen: On the left, a redacted note: the name is struck out and removed, and four details are kept, a role, a place, a date and an unusual circumstance. On the right, the people of a small institution, shown as identical figures. As each kept detail is applied, the people who no longer fit fade: the role leaves a handful, the place fewer, the date two, and the unusual circumstance one person, who is circled. No name was used. The last frame asks whether a colleague who knows the setting could work out who this is.]
A role, plus a place, plus a date, plus an unusual circumstance. In a small institution that identifies somebody as surely as a name. Ask whether a colleague who knows the setting could work out who this is. And notice that you are the worst possible judge of that, because you already know the answer.
Two questions, not one
Your unsubmitted coursework is yours alone as far as privacy goes. Whether you may use an AI system on it is a completely different question, with a different authority: your institution's. Keep them apart. "It is my own work, so there is no privacy problem" can be true at the same time as "and I am not permitted to do this". A rule like this protects people only if someone can check that it is followed. If nobody can check it, an institution may believe it is protected when it is not. It is fair to ask what such a rule can achieve on its own. That is a question for the people who write the rule. It does not make the rule optional for you. Those who write it have to ask what people can realistically do, and what can be checked. Then they design the rule, and the support around it, on that basis.
Next
The activity asks you to find out, for the service you actually use, what happens to what you send, and to record where you found each answer. "I could not find out" is a legitimate answer. Note how long you looked, and let it make your classification stricter. Then sort eight items, redact one, and take the two 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.