Cut what a model shouldn’t know. Keep the rest.
Pruniax R&D researches and develops machine unlearning: ways to make an AI model forget specific things it has learned without rebuilding it from scratch, and to audit and verify that it really has. We do it to make the world a safer place.
You can’t take the egg out of a baked cake
Think of baking a cake. Flour, milk, eggs and sugar go in, and what comes out of the oven is one thing. You can’t reach in afterwards and take out the milk or the egg.
An AI model is made the same way. Everything it learns from is mixed together during training. So when it turns out to have learned something it shouldn’t, there is no single place to delete it from. It is baked in.
You could bake a whole new cake without that ingredient, but for a large AI model that is slow and very expensive. Machine unlearning takes another route, closer to neurosurgery. The aim is to find exactly where the unwanted knowledge sits, remove it, and leave everything else working as before.
Deleting the data doesn’t delete what the model learned
In technical terms, every training example nudges millions of weights. Afterwards the example isn’t stored in the model as a record, but its influence is spread across all of them. Deleting a row from the training database changes nothing about a model that has already learned from it.
Retraining from scratch without that data works, but for a large model it means paying the full training bill again each time something has to go. Machine unlearning aims for the same result at a fraction of the cost: a model that behaves as if it had never seen the data.
Why unlearning matters
AI models pick up whatever is in their training data, including things nobody wants them to repeat.
-
Hazardous knowledge
A model should not be able to hand out instructions for weapons or dangerous substances. Unlearning aims to remove that knowledge from the model itself.
-
Keeping children safe
An AI that children talk to should not be able to teach them how to commit a crime. Removing the knowledge goes further than filtering the answers.
-
Fake news and manipulation
A model that has absorbed false or manipulative content can repeat it to millions of people and shift public opinion. Unlearning is a way to take it back out.
-
Fairness
Biased data teaches a model to treat people unequally. Removing its influence helps the model decide more fairly.
-
Privacy
People have a right to have their data erased, under laws such as the GDPR. If that data shaped a model, deleting the original record may not be enough.
-
Integrity
Copyrighted, unlicensed or deliberately poisoned data can slip into training. Taking it out lets the people behind a model stand by what it was built on.
Auditing and verification
Removing knowledge from a model is one thing. Showing that it is really gone is another. This is the service we offer alongside our research.
-
Auditing
We examine an AI model for knowledge it should not hold, such as hazardous instructions, personal data or biased patterns, so it is clear what has to be removed.
-
Verification
After unlearning, we test whether that knowledge is really gone and whether the rest of the model still works as before. The result is evidence that others can check.
Three questions every unlearning method has to answer
These are the measures we hold our own work to, and what our verification checks.
- Is it really gone?
- The model should behave like one that never saw the data it was asked to forget, and that has to hold up when someone tests it.
- Is everything else intact?
- Everything the model is meant to keep should work as well as before. A cut that damages the rest of the tree isn’t pruning.
- Is it cheaper than retraining?
- Retraining from scratch is the baseline. A method only earns its place if it gets there with far less computing power.
Our mission
Our mission is to make AI safer, fairer and easier to trust, by removing what a model should never have learned and showing that it is really gone.
The name says how. Pruniax is pruning plus an axe. Pruning takes away what a tree doesn’t need so the rest can grow, and the axe is for when a whole branch has to go.
Rooted in Groningen, guided by European values
Pruniax R&D is established in Groningen, in the north of the Netherlands. This is where our roots are.
Europe has chosen to put people first in how technology is built. Privacy, human dignity, fairness and accountability are protected by laws such as the GDPR and the EU AI Act.
Machine unlearning is a practical way to live up to those values. The right to be forgotten only means something if an AI model can actually forget.
We build our methods to protect people’s data, to keep AI safe and trustworthy, and to let organisations show that they do.
Talk to us
Working with a model that needs to forget something, or researching unlearning yourself? Write to us.
- info@pruniax.com
- Company
- Pruniax R&D
- Based in
- Groningen, the Netherlands
- KvK number
- 42180708