Skip to content
ExamSpine

Training data ethics: consent, copyright and attribution

Image generators learn from huge numbers of images made by real people. Who agreed to that, who owns the result, and who should be told it is AI?

Challenge 2SL and HL

Where the training images come from

Large image models are trained on hundreds of millions to billions of image and caption pairs. Most were collected automatically from the public web, along with the alt text or nearby words as captions.

=="Publicly visible" is not the same as "permission to train on".== A photographer who posts a portfolio online has not necessarily agreed for it to be used to build a commercial product that may compete with them.

The main ethical questions

  • Consent. Did the creators agree? Options range from none, to an opt-out (creators can ask to be removed), to opt-in (only images whose owners said yes).
  • Copyright and ownership. Copying images to train on may infringe copyright, depending on the country and on whether it counts as fair use or a permitted exception. Courts in several countries are still deciding. Who owns a generated image is also unsettled: in some places, work with no human author cannot be copyrighted at all.
  • Attribution and compensation. A generated image draws on many training images at once, so crediting or paying individual creators is technically hard. Some companies license data from stock libraries and share revenue.
  • Style imitation. Prompting "in the style of" a named living artist can produce work that competes with them using what they made.
  • Memorisation. Models occasionally reproduce near-copies of training images, especially ones duplicated many times in the data. That can leak copyrighted work or private photos.
  • Transparency. Should people be told an image is AI-generated? Labels, invisible watermarks and signed provenance records (content credentials) are approaches.

What a responsible company can do

  • Use licensed or public-domain datasets, or its own images, and keep records of where every image came from (data provenance).
  • Honour opt-out requests and do-not-train signals, and remove duplicates to reduce memorisation.
  • Block prompts that name living artists, real people or trademarks.
  • Label generated images clearly and attach provenance metadata.
  • Publish how the model was trained, so customers and creators can judge it.

Key idea: In an exam answer, an ethical point earns most when it names who is affected and what the company could do about it. "It is unethical" earns little; "photographers whose interior shots were scraped get no credit or payment, so the company could license a stock library instead" earns well.

Key terms

Training data
The examples a model learns from: for image generators, very large sets of images with captions.
Consent
Permission from the creator or subject of an image for it to be used, here for training a model.
Opt-out
A way for creators to ask that their work be removed from training data. Opt-in, by contrast, uses only work whose owners said yes.
Copyright
The legal right of a creator to control copies of their work. Whether training a model on it is allowed is still being decided in several countries.
Fair use / copyright exceptions
Legal rules that allow some uses of copyrighted work without permission. Whether they cover AI training differs by country.
Attribution
Crediting the people whose work was used. Hard for generators, because each image draws on many training images at once.
Memorisation
When a model reproduces a training image almost exactly instead of making something new.
Watermark
A mark, visible or hidden in the pixels, that identifies an image as AI-generated or shows who made it.
Provenance
A record of where data or an image came from. For generated images, signed metadata (content credentials) can record that AI made them.
Licensed dataset
Training data used with the owners' permission, usually paid for, with a record of where each item came from.

See all 65 terms in the glossary

Exam-style questions

In the style of Paper 1 Section B. Write your answer first, then open the marking points and mark yourself honestly. The questions use our own company, Hearth & Oak: an online furniture shop that wants to generate room scenes around its products.

Worked example · Discuss · AO3

[6 marks]

Hearth & Oak is choosing between a model trained on images scraped from the web and a model trained only on licensed stock photos and its own studio photos. Discuss which it should choose.

How to approach it

"Discuss" means both sides and a conclusion. Plan two or three points for each option, always linked to this company (a furniture shop, a brand built on trust). Leave two lines for a conclusion that says which and why. The conclusion mark needs a reason that weighs one side against the other.

0 words

Marking points and model answer

A model answer

The web-scraped model has learned from far more images, so it is likely to produce more realistic and varied rooms, from tiny city flats to farmhouse kitchens. However, the photographers and designers whose work it learned from did not agree and are not credited or paid. That is an ethical problem, and also a business one: Hearth & Oak could face copyright claims, especially if the model reproduces a near-copy of a well-known interior photo, and a story about "stolen" images would damage a brand that depends on trust.

The licensed model avoids most of this. Hearth & Oak can show where every training image came from, and the generated images are safer to use in adverts. The cost is a smaller, narrower dataset, so rooms may look less varied, and licences cost money.

On balance Hearth & Oak should choose the licensed model. Its main need is realistic scenes around its own products, which its studio photos already cover well, and the legal and reputational risks of the scraped model are much harder to fix later than a narrower range of room styles.

For each point, does your answer make it? Be strict: a marker would be.

  • Point 1: Scraped model: much larger and more varied data, so likely better quality and range of room styles. (1 mark)
  • Point 2: Scraped model: creators did not consent and are not credited or paid; ethical concern for photographers and designers. (1 mark)
  • Point 3: Scraped model: legal risk (copyright claims, memorised near-copies) and reputational risk for a brand built on trust. (1 mark)
  • Point 4: Licensed model: clear permission and provenance; images safer to use commercially. (1 mark)
  • Point 5: Licensed model: smaller or narrower data, higher cost, possibly weaker or less varied results. (1 mark)
  • Point 6: A justified conclusion that weighs the points for this company (for example licensed, because legal and trust risks outweigh some loss of variety; its own product photos cover what matters most). (1 mark)

Also accepted

  • Mentioning the option of fine-tuning a licensed base model on its own photos, as a compromise, can earn point 6.

Earns nothing

  • A conclusion with no reason.
  • Points about AI in general that are not linked to the two options.

Mark every point to see your total.

Question 1 · Outline · AO1

[2 marks]

Outline what is meant by memorisation in a generative image model.

0 words

Marking points and model answer

A model answer

Memorisation is when a generative model outputs a copy or near-copy of one of its training images instead of a new image. It is more likely for images duplicated many times in the training set, and it can expose copyrighted or private pictures.

For each point, does your answer make it? Be strict: a marker would be.

  • Point 1: The model reproduces a training image exactly or almost exactly, instead of making something new. (1 mark)
  • Point 2: More likely for images that appear many times in the training data / it can leak copyrighted or private images. (1 mark)

Earns nothing

  • "The model memorises how to draw" (that is ordinary learning).

Mark every point to see your total.

Question 2 · Identify · AO2

[2 marks]

Identify two ways Hearth & Oak could make it clear to customers that a room image was generated by AI.

0 words

Marking points and model answer

A model answer

It could show a visible "AI-generated scene" label on every generated image, and attach provenance metadata (content credentials) to the file that records that it was made by an AI model.

For each point, does your answer make it? Be strict: a marker would be.

  • Point 1: A first valid way, for example a visible label or caption on the image ("AI-generated room"). (1 mark)
  • Point 2: A second valid way, for example an invisible watermark, provenance metadata / content credentials, or a note on the product page. (1 mark)

Earns nothing

  • Two versions of the same label.

Mark every point to see your total.

Question 3 · Explain · AO2

[4 marks]

A customer types "a living room in the style of [a named living illustrator]" into Hearth & Oak's "Design my room" box. Explain one ethical issue this raises and one technical measure Hearth & Oak could take.

0 words

Marking points and model answer

A model answer

The prompt asks the model to imitate a specific living illustrator. That artist did not agree to have their work used this way and gets nothing from it, and the result could compete with the commissions they rely on or be passed off as theirs. Hearth & Oak could check prompts against a list of living artists' names and refuse or rewrite those requests. This stops the most direct cases, although a determined user could still describe the style in other words, so it reduces the problem rather than removing it.

For each point, does your answer make it? Be strict: a marker would be.

  • Point 1: Identifies the issue: imitating a living artist's style without consent / payment. (1 mark)
  • Point 2: Explains the harm: it uses their work to make images that may compete with them or be mistaken for theirs. (1 mark)
  • Point 3: A technical measure: filtering prompts for artist names / a blocklist / refusing such prompts. (1 mark)
  • Point 4: Explains how it helps or a limitation (for example users can describe the style without the name). (1 mark)

Also accepted

  • Removing that artist's work from the training data (honouring an opt-out) is a valid measure.

Earns nothing

  • "It is illegal" with no explanation.

Mark every point to see your total.

Glossary

65 key terms, in plain English

Practice questions

Paper 1 Section B style, with full answers

All case study topics

  1. Challenge 1

    SL and HL

    How diffusion models make an image

    Noise, denoising, the network that predicts the noise, and how a text prompt steers it.

  2. Challenge 2

    SL and HL

    Training data: ethics, ownership and bias

    Consent, copyright and attribution for training images, and how curation choices shape what the model makes.

  3. Challenge 3

    HL only

    GANs: a generator against a discriminator

    How two networks train against each other, and why that training is unstable.

  4. Challenge 4

    HL only

    Hybrid models and the trade-offs between them

    Systems that combine diffusion, GANs and autoencoders, judged on quality, speed, diversity and complexity.

Written by Spine in our own words. Spine is not affiliated with or endorsed by the International Baccalaureate. Every question, set of marking points and model answer here is our own, set in our own made-up company. Facts about the case study itself come from public guides and are marked “reported”. Last checked 2026-10-23.