TwiddleLab / Artificial Intelligence / Generative AI: Making New Things
Copyright and Consent
What you will learn
- "Is it allowed?" splits into separate questions: may it be trained on, may the output be shared, who is the author. Different people decide each one.
- Consent means the owner agreed. Your own work, CC0 and public domain are open; living artists and other people's photos need a yes.
- Memorization: a model can reproduce a training picture almost exactly, mostly ones repeated many times. Removing repeats helps.
- Laws differ between countries and are still changing. This is not legal advice.
How the lesson goes
An AI makes a picture in the style of your favourite artist. Did the artist agree? Do they get anything?
What do you think?
Build a training set: what may go in?
Drag each card into Fine to use, Ask first or Leave out.
Your own work, CC0 and public domain can be used by anyone. A living artist's portfolio and your friends' faces belong to them: ask. Private documents and sites that forbid scraping stay out.
This model copies one picture almost exactly. Remove the repeats until it stops.
Tap the copies of Sunset Critter in the training set. Keep one.
Consent, memorization and authorship.
Three different questions behind "is it allowed?".
Who decides each question?
Drag each question to its decider.
Lawmakers and courts set the general rules, a licence sets the terms for one work, each platform sets its own rules, and an artist decides what to allow with their own work.
You post your art online. Switch on protections until every risk is reduced.
Tap protections. The table shows which risks each one helps with.
Code: your AI helper suggests code from an open-source project. Can you use it?
Switch between the two licences and compare their rules. Then pick.
MIT lets you use the code anywhere as long as you keep the copyright and licence notice. The GPL lets you use it too, but if you give out a program that includes it, you must share that program's source under the GPL.
This generated picture is almost a copy of one training picture. Which?
Tap it, then Check.
A near copy keeps the layout, shapes and colours of one training picture. That is memorization, and it usually happens with pictures that appeared many times in the data.
Earn the Fair Data badge: pick 6 pictures for a new dataset.
Every licence must allow training, and you need critters, landscapes and buildings. Fewer submits, more stars.
Previous lesson: Music and Voice Generation · Next lesson: Can We Detect Generated Content? · Part of the course: Generative AI: Making New Things
Lessons are interactive and need JavaScript. Turn it on to play them.
