Google launches a way to audit whether AI models actually forget

Google Research just released a new framework for auditing machine unlearning — the process of removing specific data from a trained model without retraining from scratch. The tool checks whether the model actually forgot what it was supposed to, instead of just pretending.

Machine unlearning has been a hot topic, but most tools claim to do it without proof. Google's approach measures the model's behavior before and after unlearning, comparing predictions and gradients to verify real change. It's an open framework, not a black-box product.

The Linux Foundation's OpenSharing project, mentioned alongside, is pushing the same direction: standardizing how AI assets and data move between systems. If models can't reliably forget, they can't reliably adapt — and that's a problem for anyone who needs to remove someone's data, update a dataset, or comply with privacy rules.

Why this matters for us: la migra app, la app de salud, la app del trabajo — all of them are using models that need to forget when we do, and we need to be able to prove it.