Understand the technology
Make responsible decisions about image data.
For May and November 2027. This lesson is for SL and HL students. Learn through practical studio examples, then practise separating evidence, permission, quality and ethical judgement.
01 | A DATASET IS A COLLECTION OF DECISIONS
Visionary Studios wants images that are useful for creative work. That starts partly with the examples from which a model learns, not just the prompt entered at generation time.
Dataset curation means deliberately selecting, organising, checking and documenting data for a purpose. It includes visual suitability, accurate descriptions, useful representation and the basis for using the material.
The case study highlights avoiding unauthorised copyrighted material and respecting intellectual property. This shared SL and HL lesson explains the questions involved; it does not assume one legal rule applies everywhere.
02 | TRAINING DATA AND REFERENCE INPUTS ARE DIFFERENT
A training dataset contributes to learning model parameters. A reference image supplied during ordinary generation can guide an output without updating those parameters.
Those technical roles differ, but both can raise permission, confidentiality and contractual questions. Uploading a client image to an external service is a separate action from storing it in a local project.
The studio should identify the intended use precisely: collection, training, fine-tuning, reference conditioning, sharing and publication may require different checks.
03 | WHAT IS INTELLECTUAL PROPERTY?
Intellectual property , or IP, concerns legal rights associated with certain creations, inventions and commercial identifiers. Copyright, trade marks, patents and registered design rights are distinct examples.
For this image case study, copyright in artworks and photographs is particularly relevant. Brand logos can also involve trade mark and other rights.
Do not treat all IP as one permission. A licence addressing one right or activity may not resolve every issue affecting a final advertisement.
04 | COPYRIGHT FROM FIRST PRINCIPLES
Copyright can protect qualifying creative expression, such as photographs and illustrations. It concerns protected expression rather than granting ownership of every general idea or theme.
A broad idea such as an explorer beside a lake is different from copying a particular protected illustration. However, deciding whether specific material is protected or an output infringes depends on facts and applicable law.
Protection, ownership and exceptions vary across jurisdictions. Avoid turning an introductory definition into an automatic decision about a disputed AI training use.
05 | ACCESS IS NOT THE SAME AS PERMISSION
An image appearing on a public website tells us that it is accessible. It does not, by itself, establish permission to copy it, train a model with it or use it commercially.
A download button or absent watermark is also not sufficient evidence of broad rights. The relevant licence, ownership, contract or applicable exception needs investigation.
For the studio, an unknown source should become an unresolved record rather than silently being marked “free to use”.
06 | OWNERSHIP AND LICENCES
A rights holder owns relevant rights. A licence grants specified permission under stated terms; it does not necessarily transfer ownership.
Read the scope: which material, activities, users, duration and conditions are covered? Permission to display a stock image in one campaign does not automatically establish permission for model training.
Client-supplied material also needs a clear basis for the studio’s intended use. Receiving a file is not proof that the client controls every necessary right.
07 | READ CONDITIONS RATHER THAN ASSUME
Some licences require attribution, restrict commercial uses or impose other conditions. Open access and openly licensed are not identical descriptions.
Attribution acknowledges a source; it does not replace permission where permission is required. Paying for access also does not automatically purchase unlimited rights.
Keep the actual licence and version with the source record. A vague note saying “found online” is poor evidence for a later review.
08 | PUBLIC DOMAIN NEEDS VERIFICATION
Public-domain material is not restricted by copyright in the relevant way, but its status should be checked for the applicable jurisdiction and the specific item.
A website’s label can be mistaken, and other considerations may remain, such as privacy, trade marks, confidentiality or contractual restrictions.
For a dataset, record why the public-domain conclusion was reached. Do not infer it merely because an image is old or has been reposted many times.
09 | EXCEPTIONS AND AI TRAINING ARE CONTEXT-DEPENDENT
Copyright systems can provide exceptions or limitations for particular uses. Their scope and conditions differ, including how they apply to text and data mining or AI training.
There is no safe universal claim that “all training needs a licence” or “all training is permitted because the model learns patterns”. The applicable rules and facts matter.
Students should identify the uncertainty and the checks needed. For a real deployment, the studio should obtain appropriate advice about unresolved legal questions rather than rely on a classroom rule.
10 | PROVENANCE MAKES THE DATASET REVIEWABLE
Provenance records where material came from and how it has been processed. Useful fields include source, creator or rights holder where known, acquisition date, licence terms, intended use and review status.
Also record transformations and removals. If a dataset contains copies from several sources, deduplication can reduce repetition while preserving the records needed to understand origin.
Provenance supports investigation. It does not itself prove that a licence is valid or that all rights have been resolved.
11 | BUILD A SIMPLE RIGHTS REGISTER
| Item | Evidence | Decision pending |
|---|---|---|
| Studio-commissioned character sheet | Contract and contributor details | Confirm scope covers intended training and service upload |
| Stock photograph | Licence text and purchase record | Check permitted activities and restrictions |
| Image copied from a social post | URL only | Establish rights and a suitable basis before inclusion |
| Purported public-domain landscape | Archive record | Verify item status and any other relevant conditions |
These are investigation records, not legal verdicts on particular materials.
12 | CURATION ALSO SHAPES WHAT A MODEL LEARNS
Remove corrupt files and check image descriptions. A misleading caption can teach an incorrect association; repeated copies can overemphasise one visual pattern.
Assess relevant variation in people, environments, viewpoints and styles. A dataset can have documented permissions yet still be unrepresentative or unsuitable.
Conversely, visual diversity does not resolve rights. Data quality, fairness and permission are related review dimensions, not substitutes for each other.
13 | A MODEL CAN RETAIN TRAINING INFORMATION
Generative models learn numerical patterns, but can also memorise aspects of examples and sometimes produce close reproductions. Generation is not a guarantee of originality.
Check selected outputs for concerning resemblance, protected artwork, brand identifiers and other project-specific issues. Similarity tools may help investigate, but cannot provide a universal legal clearance score.
The training-data review and output review therefore address different risks. One does not eliminate the need for the other.
14 | INPUT RIGHTS AND OUTPUT RIGHTS ARE DIFFERENT QUESTIONS
Ask separately whether the studio may use input material, whether an output could interfere with someone else’s rights and what rights the studio can claim or grant in its own result.
AI-generated output protection varies with jurisdiction and human contribution. A provider’s commercial-use terms do not automatically establish copyright ownership or exclusive rights everywhere.
Document creative human work, editing and agreements rather than promise clients that every generated image is automatically exclusive.
15 | EVALUATE EXTERNAL PROVIDERS
Ask what the provider discloses about training data and relevant licences. Review terms covering input retention, reuse, output use and responsibility for disputes.
A general statement that a tool is “safe for business” is not detailed evidence. A contractual protection may have conditions and exclusions; it does not erase every concern.
WIPO’s organisational guidance recommends reviewing tool terms, dataset licence coverage, confidential information safeguards and outputs. Use the source links below to explore the questions rather than assume a provider’s claim settles them.
16 | ETHICAL AND LEGAL JUDGEMENTS CAN DIFFER
A use may raise concerns about creator recognition, remuneration or community expectations even when someone argues that it is legally permitted.
Equally, an ethically attractive aim does not automatically make copying lawful. A charitable campaign still needs appropriate checks.
For Visionary Studios, trust with artists and clients can be a business requirement as well as an ethical consideration. Explain the affected stakeholders and the consequences of a decision.
17 | APPLY A CURATION WORKFLOW
Identify sources, purpose and relevant terms.
Check quality, representation and the basis for intended use.
Keep evidence, resolve uncertainty and review outputs separately.
18 | WRITE A CASE-SPECIFIC RESPONSE
In your book, assess a proposed dataset containing client artwork, stock images and unattributed web images. Identify three different checks and explain why crediting all creators would not resolve every issue.
Reveal a possible response
Check client authority and contract scope, review stock licence conditions for the intended activities, and establish a suitable basis for unattributed material before inclusion. Record source evidence, review status and transformations. Attribution can acknowledge creators but does not replace required permission.
Also assess image quality and representation, investigate provider handling and check selected outputs. Distinguish proposed safeguards from a guarantee of legal compliance.
Use the official 2027 booklet alongside these independent teaching notes. This is general educational content; applicable law and contracts need checking for a real use.
Further reading: WIPO: Generative AI and intellectual property ; US Copyright Office: AI reports (US-specific) ; EUIPO: Copyright and GenAI (EU-specific) . Consult current sources; rules differ between jurisdictions.
CHECK YOUR UNDERSTANDING | CLASSIFY THE CURATION CHECK