Chapter 3 · Creativity, Work and Safety
Copyright, Licences and Attribution: Using AI Output and Source Material
- Page 11 of 22
- 21 min read
The last two pages were about people's data. This page is about people's creative work: the photographs, drawings, articles, songs, code and datasets that other people made. AI tools are built from such material and can produce material that looks like it. That creates three practical questions for you: may I use this source, may I use what the tool gave me, and what do I owe the people whose work is involved?
Some of these questions are contested, and the law is moving. This page does not give a verdict. It shows you how to ask the questions in order, how to read the strongest arguments on several sides, and how to decide carefully when you use AI in work that others will see or pay for. It builds on the decision pattern from Everyday AI Ethics: Who Benefits and Who Could Be Harmed and on the habit of asking about sources from Privacy, Consent and Data Use: What You Share with an AI Tool.
What you will learn
- The difference between copyright and plagiarism, and between owning a copy and having permission to use it.
- Why the legality of training AI on creative work is contested, and the strongest arguments for broad access and for creator control, side by side.
- The difference between input rights (the material that goes in) and rights in generated output (what comes out).
- How to check the licence of an image, a text or a dataset, and why attribution is different from permission and does not solve every rights issue.
- How to review an AI-generated illustration for a client and the licence conditions of an AI-generated asset before you reuse it.
Legal note (as of 2026-10-12; not legal advice): Copyright rules are set by each country's law and by courts, and countries differ on what is protected and what exceptions exist. As of this date, positions on AI and copyright differ by country and are changing; this page therefore describes no universal rule. The few legal statements below come from sources we fetched (a court opinion, government pages and a creators' association announcement) and are about one country or one case each. Check the current official text for your country and sector.
Copyright and plagiarism are different questions
In everyday terms, copyright is about legal control over copying, sharing and changing creative work. It concerns the rights of the creator or owner, and the details depend on where you are. Plagiarism is about honesty: presenting someone else's words, ideas or work as your own. It is a rule of schools, publishers, workplaces and professions, not a court rule. The two overlap but they are not the same, and you can break one without the other.
| Situation | Plagiarism? | Rights question? |
|---|---|---|
| You copy a stranger's blog post and publish it under your name. | Yes: you claim it as yours. | Yes: you copied and shared it without permission. |
| A classmate gives you permission to hand in her essay, and you do. | Yes: it is dishonest toward the teacher. | Probably not: the owner agreed. The school rule still applies. |
| You share a whole paid course video with friends and say clearly where it came from. | No: you named the source. | Possibly yes: the seller may not allow sharing. |
| You summarise an article in your own words and cite it. | No. | Unlikely to be a problem, but check the source's terms for how it may be used. |
This matters for AI because a tool's output can raise either question. If a chat assistant writes your report and you submit it as your own, the problem may be an integrity rule (Learning, Integrity and Disclosure: Using AI in Study and Assessment returns to that). If an image tool produces something that copies a known picture closely, the problem may be a rights question. Keep the two questions apart when you describe a concern.
Owning a copy is not the same as having permission
People often confuse owning a thing with owning the rights in it. You can own a book, a downloaded file or a screenshot without having the right to republish it, adapt it or use it in a commercial product. Permission usually comes from one of three places: the owner of the work, a licence that the owner offered to the public, or a rule in the law. A fourth thing, "it was easy to download", is not permission.
| You have... | That tells you... | It does not tell you... |
|---|---|---|
| A purchased copy of an image | You may have paid for a particular use. | Whether you may print it on products. |
| A file you can open on the web | The file is available to you. | Whether the owner allows reuse. |
| A photo you took of a mural | You made a photo. | Whether the mural's creator may have a say in how you use it. |
| A client's instruction: "use this picture" | The client wants it used. | Whether the client has the right to give permission. |
The practical question is always: who holds the rights, and what did they allow? If you cannot answer, treat the use as unresolved and either find the answer or choose another source.
Why training-data legality is contested
AI models are built by processing very large amounts of text, images, audio and code, much of it made by people. The contested question is whether copying that material to train a model is something that needs the creators' permission, or something that can be done under a legal exception or another justification. The answer may depend on the country, the kind of work, how the material was obtained, what the model does with it, and what the courts decide on the facts of each case.
You will often read confident headlines. Be careful with them. Two corrections are worth learning early. First, a settlement is not a ruling: parties who settle agree to end a dispute, and the court does not decide who was right on the legal question. Second, a court decision applies to its own facts and its own country; a different case with different facts can end differently.
Legal note (as of 2026-10-12; not legal advice): According to the Authors Guild, a United States federal court granted final approval on 2026-07-20 to a class settlement in the case Bartz v. Anthropic, which the Guild reports as USD 1.5 billion. The Guild says the release covers past acquisition and copying of the listed works through 2025-08-25, and that claims about outputs and future conduct are not released. This is a settlement, so it does not decide whether training on copyrighted works is lawful. Separately, in Thomson Reuters v. ROSS (opinion No. 25-2153, filed 2026-09-29), the United States Court of Appeals for the Third Circuit held that the headnotes at issue were copyrightable and that ROSS's use was not fair use, affirming a partial summary judgment on an interlocutory appeal. The opinion itself says ROSS's system was not a generative AI and distinguishes the generative-AI cases, so the decision is specific to its facts, and further proceedings are possible. The European Commission's pages say that providers of general-purpose AI models under the EU AI Act must have a copyright-compliance policy and publish a summary of their training content. These are examples from two regions and say nothing about your country. Check the current official text for your country and sector.
We do not predict how these questions will be settled. Treat the field as open. What you can do now is understand the arguments, and make careful choices in your own use.
Arguments for broad training access and for creator control
Here are the strongest arguments people make. They are shown in the order of a typical debate, without a verdict. Views inside any one region differ, so the labels below describe arguments and not what a whole country thinks.
| Theme | Arguments for broad training access | Arguments for creator control |
|---|---|---|
| Learning from works | People have always learned from existing works. A model that learns patterns, and does not store copies for readers, resembles that process. (A researcher at a university library in Helsinki might say this.) | A model is not a person. It copies at massive scale, and the creators whose work made it possible were not asked. (A photographer in Lima might say this.) |
| Fair payment | Requiring permission for every item could make building models impossible except for the biggest firms, which would reduce competition. | If a product earns money from creators' work, creators should be asked and paid. Payment keeps cultural work sustainable. (A poet in Nairobi might say this.) |
| Access and benefit | Broad access can produce tools for education, translation and accessibility, including for languages with little written material. (A translation start-up in Bangalore might say this.) | The same tools can compete with the creators whose work trained them, and can reduce the value of their work in the markets they depend on. |
| Practicality | The data are enormous and often cannot be traced to owners, so case-by-case permission is not workable. | Difficulty is not a justification. Practical mechanisms, such as licensing schemes or opt-out signals, can be designed. |
| Control and dignity | Models can still serve the public interest even if some creators would have declined. | Creators should decide whether their work is used to make systems they may oppose, and should be credited or compensated where they agree. |
The names above are invented personas who illustrate arguments. Between the two columns, people propose middle positions, such as opt-out signals, licensing arrangements, payments, or limits on what outputs may copy. Each has its own critics. As a learner, your task is to state each side fairly enough that its supporters would say "yes, that is what we mean", and then to explain which factors matter most for your decision and why.
Where reasonable people disagree
The disagreement is partly about facts that are still being tested (how models use training data, what courts decide) and partly about values (how to weigh creators' control against broad public benefit). New evidence can settle the first kind; the second kind remains a matter of judgement, and thoughtful people end up in different places.
Input rights and rights in generated output
Keep two separate questions, which are often blurred together.
| Question | What it asks | Example |
|---|---|---|
| Input rights | Do I have permission to use this material as input: to upload it, to train or tune with it, or to paste it into a tool? | You upload a photographer's picture to an image tool as a reference. Did the photographer allow that? |
| Output similarity | Does the output copy or too closely follow someone's existing work? | The tool returns a poster that looks almost the same as a known illustration. |
| Rights in the output | Can anyone own or control the output, and who? | A client asks to be the sole owner of an AI-generated logo. |
Legal note (as of 2026-10-12; not legal advice): According to the United States Copyright Office's report on copyrightability, published 2025-01-29, copyright protects original expression in a work created by a human author, even if the work includes AI-generated material. The report concludes that purely AI-generated material is not protected, and that where a human's control over the expressive elements is insufficient, protection is not available; it says that, based on the functioning of currently available technology, prompts alone do not give sufficient control. It adds that a human author can have copyright in their own perceptible contributions, in the creative selection, coordination or arrangement of material, and in creative modifications, and that the amount of human contribution is assessed case by case. The Office says it will keep monitoring. This is one country's official view; other countries differ, and the next page returns to that. Check the current official text for your country and sector.
Why this matters in practice: if an output may not be protected, a client who thinks they will "own" it exclusively may find that others can copy it. That is a reason to be honest with clients about what you can and cannot promise.
Checking the licence of a source image, text or dataset
A licence is the owner's written permission, with conditions. Before you use a source, especially in work for others, check it in this order.
Worksheet:
LICENCE CHECK
1. Where did it come from? (original owner, a site that re-uploaded it, a dataset page?)
2. Is a licence stated? (copy the text or the link; note the date you saw it)
3. May I use it commercially? (yes / no / not stated)
4. May I change it? (crop, edit, translate, mix with other things)
5. What must I do? (credit, link, share changes under the same terms)
6. Who else has rights in it? (a person pictured, a model, another creator, a brand)
7. What is NOT covered? (the licence of the dataset is not the licence of each item in it)
8. Result and who decided: (use / ask first / do not use; name, date)Three traps to watch for. First, "no licence stated" does not mean "free to use"; it usually means you must ask. Second, a re-upload on another site may not carry the original owner's permission. Third, a dataset's licence is not always the licence of the items inside it, so a dataset labelled as open can still contain items with conditions. If you cannot settle the question, choose another source or ask the owner in writing and keep the reply.
Attribution is not permission, and does not solve everything
Attribution means saying who made something or where it came from. Permission means the right to use it. Giving credit is good practice and honest, and some licences make credit a condition of the permission. But credit alone does not create a right to use.
| You... | Credit given? | Rights issue solved? |
|---|---|---|
| Use a photo and write "Photo: A. Rahman" without asking. | Yes | No. You still need the right to use the photo. |
| Use a photo whose licence says "reuse with credit", and you credit it. | Yes | Yes for that condition, but check the other conditions too. |
| Use a licensed photo of a person for a product advertisement. | Yes | Not necessarily: the person pictured may have a say, and the licence may not cover advertising. |
| Credit an AI tool for the image you made. | Yes (for the tool) | No: this says nothing about the rights of the sources or the terms of the tool. |
| Use text copied from a site that wrongly claimed it. | Yes | No: the person who gave permission did not have the right. |
A good habit is to treat attribution as a separate line in your checklist: "Do I have permission?" first, then "How do I credit, honestly and as the licence asks?"
A hypothetical AI-generated illustration for a client
Invented scenario:
Ananya is a freelance designer in Pune. Her client Rafael owns a bakery in Bogotá and wants a
poster illustration of a rooster for a spring campaign. Rafael plans to print 5,000 posters and
put the picture on the shop's packaging. Ananya generates several options with an image tool.
For the best one she gave the tool a photograph of a rooster that she found on a travel blog,
as a reference. She hasn't checked the photo's licence. Rafael says: "I want to be the sole owner
of this image so no competitor can copy it."This is a small job, but it contains all three questions from this page. Run it through the pattern, with the steps that matter here.
Worksheet, filled in (steps 2 to 8):
2. WHO IS AFFECTED Rafael (will print and sell), Ananya (professional reputation), the travel
blogger or photographer who made the reference photo, other illustrators
whose work is similar, bakery customers.
3. BENEFITS/HARMS Benefit: a custom picture fast and cheaply. Harm: the reference photo may be
used without permission; the print may resemble an existing image; Rafael
may be told he owns something nobody can exclusively own.
4. OPTIONS A. Use the image as it is and tell no one.
B. Regenerate without the reference photo; compare the result with known
images; give Rafael an honest statement about what is protected.
C. Ananya draws the rooster by hand, or edits the output heavily and keeps
her drafts as a record of her contribution.
D. Buy a licensed illustration whose terms cover printing and packaging.
5. EVIDENCE Known: the photo's licence is unknown. Unknown: the tool's commercial terms;
whether the output resembles an existing image. Check: read the terms;
search for the photo's owner; run an image search for look-alikes.
6. DECISION B or C. Drop the reference photo. Read the tool's terms and save a dated copy.
Tell Rafael plainly what Ananya can and cannot promise about exclusive
ownership, in writing. Stop if the licence cannot be checked.
7. ACCOUNTABLE Ananya answers to Rafael for her process; Rafael decides how to use the result.
8. REVIEW DATE Before printing, and again if the tool's terms change.Notice that "tell Rafael honestly" appears as a condition. Whether to disclose AI use to a client is covered elsewhere, but promising something you cannot deliver is a plain honesty problem. Reasonable designers may differ: one may prefer D, because a licensed illustration gives the clearest paper trail; another may prefer C, because their own drawn work is the strongest in both ethics and rights. Both are defensible if the reasons are stated.
Licence conditions before reusing AI-generated assets
The terms of the tool you used also matter. Before you reuse an AI-generated image, text, audio or code in client or public work, read the tool's current terms and record the answers.
Template:
ASSET RECORD
Asset: (name or file; what it is for)
Made with: (tool and plan; date of generation)
Terms read on (date): (keep a copy or a screenshot of the terms)
Commercial use allowed? yes / no / not stated
Who owns or may control it? (as the terms say; and note: terms do not settle every question)
Restrictions: (credit, no use for some purposes, plan limits, anything else)
Inputs used: (any uploaded or reference material and its licence)
Human contribution: (what I wrote, chose, edited or drew; keep drafts)
Similarity checked? (how, and the result)
Client told? (what I said, and when)
Decision and person: (use / ask first / do not use; name, date)Two limits to remember. First, a tool's terms describe what the provider promises or allows, and they cannot settle the rights of third parties, or whether the law will protect the output. Second, terms change, which is why you record the date. A short record like this also protects you if a client asks, months later, where an image came from.
Where this shows up at work
A marketing assistant who writes the source and licence next to every image in a campaign folder can answer "where did this come from?" in seconds. A teacher who tells pupils the difference between citing a source and having permission to reuse it prevents a common confusion. A freelancer who keeps drafts and a short asset record can show their own contribution, and can say honestly what a client does and does not receive. A small publisher who asks authors where an illustration came from avoids problems after printing.
Common mistakes
- Mixing up plagiarism and copyright. Fix: ask two questions: "Am I claiming this as mine?" and "Do I have the right to use it?"
- Assuming that a file you can download is free to use. Fix: find the owner and the licence, and write them down.
- Treating a settlement or one court decision as the answer to everything. Fix: remember that a settlement is not a ruling, and that a decision applies to its own facts and country.
- Believing that credit fixes the rights problem. Fix: get permission first; credit second, as the licence asks.
- Promising a client exclusive ownership of an AI output. Fix: read the tool's terms and your country's rules, and say honestly what you can promise.
Try it yourself
- Easy. For each case, write P (a plagiarism concern), R (a rights concern), B (both) or N (neither): (a) a student pastes an AI-written essay and hands it in as her own; (b) a shop prints a photographer's picture on mugs, with the photographer's name on each mug but no agreement; (c) a writer summarises a report in her own words and cites it.
- Medium. Omar (from Everyday AI Ethics: Who Benefits and Who Could Be Harmed) wants to use an illustration from a free-wallpaper site in his shop catalogue. The site has no licence text. Fill in the first six lines of the licence-check worksheet for his case and say what he should do next.
- Hard. Lena, a teacher in Berlin, asks an image tool for a cartoon of a robot to put on a school flyer and a school fundraising T-shirt. She also asks whether the school may print the same picture on 300 T-shirts for sale. Decide with the eight-step pattern (use steps 2 to 8), complete a short asset record, and give a fair statement of one point on which two reasonable people could disagree.
Answers
Easy. (a) P: she claims work as her own, so it is mainly an integrity question; whether a rights question also arises depends on the tool's terms and the sources. (b) R: using the picture on products without agreement is a rights concern; the name is credit, which does not replace permission. It is not plagiarism, because the photographer is named. (c) N: own words and a citation, assuming the report's terms allow use. Pass: you separate honesty from permission in each, and you note that credit does not equal permission in (b). Weak: you treat credit as enough, or call every case "copyright".
Medium. A strong worksheet: (1) a free-wallpaper site, which may have re-uploaded the picture from elsewhere; (2) no licence text found; (3) commercial use is not stated, so it cannot be assumed; (4) changing is not stated; (5) no conditions known; (6) unknown. Next step: search for the original creator, ask the site and the creator for written permission, or choose an image with a clear licence that allows commercial use, or ask a designer for an original. Pass: "not stated" is treated as unresolved, the re-upload risk is mentioned, and the next step is concrete. Weak: "it is on the internet, so it is free".
Hard. A strong answer: the people involved are Lena, the school (which will sell the shirts), pupils and parents, other illustrators whose robot drawings might look similar, and the tool's provider. Benefit: a quick, cheap picture; harm: a close resemblance to an existing character, an unclear commercial licence, and a school making a promise it cannot keep. Options: use as is; generate several and compare to look-alikes; draw one with pupils; buy a licensed picture. Decision: generate, check the tool's terms for commercial use and keep a dated copy, run an image search for look-alikes, do not prompt with a specific known character, and ask the head teacher before selling; stop if the terms do not allow sale. Accountable: Lena, with the head teacher approving the sale. Review: before the order is placed. The asset record names the tool, the date, the terms read, commercial use yes/no/not stated, the inputs (none), the human contribution (prompt choices and selection, perhaps edits), and the similarity check. A fair disagreement: one teacher may say the picture is only a flyer decoration and the risk is low; another may say that selling shirts changes the stakes and a licensed picture is safer. Pass: the pattern is used, terms and look-alikes are checked, the sale changes the care needed, and both views are stated fairly. Weak: "AI pictures belong to whoever made them", or no condition for selling.
Summary
- Plagiarism is about honesty; copyright is about rights. A case can raise one, both or neither.
- Owning a copy is not permission. Ask who holds the rights and what they allowed.
- Whether training on creative work needs permission is contested and changing. A settlement is not a ruling, a court decision applies to its own facts, and the arguments for broad access and for creator control both have real force.
- Treat input rights, output similarity and rights in the output as three questions. Official views on who can hold rights in AI-assisted work differ by country.
- Check licences, treat "not stated" as unresolved, record terms with a date, and remember that attribution is not permission.
Next: Creative Rights, Imitation and Freelance Practice: Fair Work with AI looks at how copyrightability varies by country, when to document your own creative contribution, imitating a living creator's style, and fair practice with clients.