Video
Unit video
The narrated video for Unit 14, about 8 minutes, is published with the first release.
Learning purpose: Three obligations owed separately -- rights, attribution and disclosure -- and honesty about which of them has unsettled answers.
Accessible text alternative: the transcript linked below, and this page, which covers the same material.
Transcript of the Unit 14 video
The narration is spoken by synthetic voices made for this course, each shown on screen as a drawn robot. No person is speaking.
Watch what changes
Reading time: about 9 minutes. Activity: about 12 minutes.
1. Two questions people run together
Unit 17 is about disclosure: saying how a piece of work was made. This unit is about two different questions that sit beside it and are constantly confused with each other.
Rights. What may lawfully be done with this — copied, published, sold, built on. Attribution. Who should be credited, and what an honest account of the work says.
They come apart in both directions, which is the useful thing to know. Something can be entirely lawful to use and dishonest to present as your own. Something can be scrupulously credited and still an infringement, a use that breaks someone's legal rights in the material. Neither question answers the other.
Add Unit 17's disclosure and there are three, and this course's position is that you owe all three separately.
2. What this course will not tell you
Copyright in generated output is genuinely unsettled, and it is unsettled differently in different countries.
That is not a way of avoiding the question. It is the answer, and stating it plainly is more useful than a confident summary would be. A page that told you "generated output is not copyrightable" or "training on public material is fair" would be doing the thing this whole course trains you to catch: making a confident claim where the evidence does not support one, written in a way that does not separate the settled parts from the contested ones.
So this unit does two things instead. It separates what is broadly settled from what is not, so you know which of your questions has an answer. And it gives you a practice for attribution, which is not a legal question and which you can get right today.
3. What is broadly settled, and what is not
Reasonably settled.
- Copyright protects a particular expression, not an idea, a fact or a style.
- Material you feed into a system does not stop being covered by whatever covers it. A licence you are bound by still binds you.
- If generated output reproduces a substantial part of an existing protected work, the ordinary rules about that reproduction apply — the fact that a machine produced it does not create an exemption.
- Terms of service are a separate matter from copyright and bind you by agreement. They govern your relationship with the provider and can restrict what you do with output. They do not themselves show that copyright exists, but where rights do exist they may license them (give permission to use them) or assign them (transfer them), subject to applicable law. (In the UK, copyright can be assigned, in writing, or licensed by its owner [R3, s. 90].)
Not settled, and contested differently by jurisdiction (the country or legal system whose law applies).
- Whether generated output has copyright at all, and if so who owns it.
- Whether training on protected material without permission is lawful, and under what conditions, including how far a rights holder's opt-out should count. This page reports no court outcome on it: where the law stands depends on the country and the date, and both are yours to check before relying on any summary.
- What obligations attach to output that resembles a particular existing work without copying it.
Copyright is not the only right. Output that imitates a real person's likeness or voice, or uses a brand's name or logo, raises other questions. In the UK there is no single image right: a likeness is protected, if at all, through a patchwork of passing off (misrepresenting something as connected with another person or business), trade marks, contract and copyright [R1, "What is an 'image right'?"]. Elsewhere a separate publicity or personality right may exist [R2, "Legal protection of image rights"]. And a copyright licence need not cover trade marks: the Apache licence, for one, grants no permission to use the licensor's trade marks [R4, sec. 6].
Notice which of these you actually need. Most practical questions — may I use this in my dissertation, may I publish this on a client's site — are answered by your institution's or client's rules and by the terms of service, both of which exist and are readable. The unsettled questions matter mainly when you need to assert a right over something, or when output resembles an identifiable work.
4. Attribution: the part you can get right
Nothing above prevents you from giving an honest account. Attribution is a practice, not a permission, and it has three parts.
Credit what a person made. If your work rests on someone's argument, data, figure or code, cite them. That obligation is entirely untouched by how you found the material. A generated summary that led you to a source does not make the source uncited — it makes it a source you found in an unusual way and have now read.
Never cite a source you have not opened. This is Unit 9, and it is the point
where the two units meet: attribution rests on the citation being real, and a
citation you have not opened is a claim you have not checked. Where you rely on a
report of a work rather than the work, say so. The course's reference list does:
where a work was read only through a report, or only in abstract, its entry records
that. scripts/check_claims.py checks the course's list of claims
(Historical Notes section 6): every source cited there has an entry
in the reference list, and the claims with no source are marked unverified. That
shows that each citation points to a real entry, not that the sources were read.
Say what you did, separately. How the work was made is Unit 17's disclosure. It answers a different question from "whose ideas are these", and running them together produces the two characteristic failures:
| Failure | What it looks like | Why it is wrong |
|---|---|---|
| Disclosure standing in for credit | "Written with AI assistance" and no citations | Says how it was made and nothing about whose ideas it rests on |
| Credit standing in for disclosure | A full bibliography, no mention of assistance | Says whose ideas, and nothing about how the text was produced |
You owe both, and they belong in different places: citations where the claim is made, and a disclosure statement about the work as a whole. They are not worth the same to a reader, though. A quotation with its locator is a certificate (Unit 0): the reader can open the source and check your claim without trusting you. "Written with AI assistance" cannot be confirmed from the text and says nothing about whether the work is right. Disclose because it is honest, and where a rule requires it; put the effort into what can be checked.
As of 2026, some tools can attach a provenance label of their own: a content credential, a signed record of how a file was made [V1, sec. 1.3], [V2, sec. 6.1], or a statistical watermark in generated text [V3, pp. 818, 823], [V4, "Watermark application", "Limitations"]. A label records where something came from, not whether it is true, and it has limits. A credential can be stripped; a text watermark marks only the output of providers that apply one, and rewriting or translation weakens its detection. A missing label shows nothing about how the work was made. None of this replaces your own disclosure.
None of these failures began with AI. Plagiarism, ghost-writing and citing papers copied from someone else's bibliography are old. What AI adds is plausible text, and plausible citations, at volume. Whatever their origin, references to works that do not exist now turn up in published, peer-reviewed papers, and one audit of biomedical papers found them becoming more common since 2023 [L10, secs. 2.1, 4.1], [L11].
Do not credit the system as an author. Not because of a rule about who counts as an author, but for a reason this course can defend: an author is someone answerable for the work. Naming a system as an author states that something can answer for the claims, and nothing in the system can. Record the tool the way you would record any other instrument — in the account of how the work was made. If your referencing style asks for an entry for a tool or a retained chat, follow it: such a reference does not make the system an author, and it does not replace citations to the underlying sources.
5. A worked case
You are writing a literature review. You ask a system for an overview; it produces a summary with six citations. You read four of the six papers, use two of them substantially, and rewrite everything in your own words.
What you owe:
- Cite the two papers you used, at the claims they support. You read them; they are sources.
- Do not cite the two you did not read, and do not cite the two you read but did not use.
- Disclose that you used an AI system to produce an initial overview, that you retrieved and read the sources yourself, and that the text is your own.
- Follow your required referencing style for the system. The bibliography lists the people and works whose ideas you relied on, and a system is not someone you can credit or question. If your style asks for an entry for a tool or a retained chat, give one: it records a tool, not an author, and it does not replace citing the two papers.
The awkward case in that list is the one people ask about: what if the framing of your review came from the summary? Then say so in the disclosure. "The structure of section 2 follows an outline produced by an AI system, which I then revised" is a true sentence, it costs you nothing, and it is exactly the kind of thing a reader is entitled to know.
6. What this unit is not saying
It is not legal advice, and it is not your institution's rule. Both exist and govern; this gives you the categories to read them with.
It is not saying attribution is optional because rights are unsettled. The opposite: attribution is the part you can get right regardless of how the legal questions resolve, which is why it is where this unit spends its effort.
Summary
- Three separate obligations: rights (what may lawfully be done), attribution (whose ideas), and disclosure (how it was made). Each needs answering on its own.
- Copyright in generated output is genuinely unsettled and varies by jurisdiction. This course says so rather than summarising it confidently.
- What is settled: your inputs keep their licences, reproduction of a protected work is still reproduction, and terms of service bind you separately.
- Cite what a person made, never cite a source you have not opened, and keep credit and disclosure in their separate places.
- Do not name a system as an author — an author is someone answerable for the work.
Then answer cp14-three-obligations and cp14-settled-or-not at the foot of
this page.
Timings: video 8 min, reading 9 min, activity 12 min.
Check yourself
This unit has 2 checkpoints, answered here. A wrong answer says which misunderstanding it matches. Nothing is uploaded.
Local progress is available in a supported browser.
Prefer a terminal?
The same questions, from The verification lab and self-check:
python3 selfcheck.py run --unit 14