A tool that claims to serve workers has to be made in a way workers could accept. These are the rules the project holds itself to, from the texts it learns from to the people who do the work.
Every source is traced
What a model learns from decides what it is. So every text that enters the corpus is documented, and the record is public.
- A text enters only when its license clearly allows the model to be published under the AGPL. When in doubt, it stays out.
- Every decision to admit or refuse a text carries the declared name of the person who made it. A text trains only when someone admitted it and nobody refused it.
- When an author says their work should not be used for training, that wish is honoured.
- Erasure on request: an admitted request removes the passage from the depot, and every later corpus is built without it.
- No distillation. Models are trained from scratch, with no output of another company's model in the data and no existing base model.
- Provenance travels with the model: every model folder names each text it learned from, with its credit and license.
Training takes only what it needs
Compute and energy are shared resources. A common model is not worth wasting them.
- The first model is small. Larger ones come only when the smaller ones have shown what works.
- No run trains beyond what is necessary, and the energy each training run uses is published.
- Recipes and checkpoints are published, so others can continue from them at their own scale instead of training the same thing again.
- Free public compute for open research is preferred to buying hardware.
The people who build it are treated as workers
An AI built on exploited labour would contradict everything the project stands for.
- No disguised unpaid work: volunteering is always free and explicit.
- Labelling, moderation and similar tasks are never handed to badly paid people.
- Every contributor is credited, with their agreement, or stays anonymous if they ask.
- Employees have a say: they are electors from their first day, and after a month or two they stand by default for election as representatives.
- Every salary is voted by the representatives, without the person concerned voting, and its amount is published.
What people write stays theirs
Using a tool should not mean feeding it.
- What people write to the association's services is not used to train the model.
- The only exception is a separate contributor mode: a document the person signs, and a distinct access point that is known to carry that consent.