Data Sovereignty
The rule that data is governed by the laws of the place it physically sits — which decides whether you are allowed to send it to someone else's model at all.
Reviewed July 17, 2026Stable
When not to use it
- As a synonym for privacy. A system can be sovereign and invasive, or private and unlawfully located.
- As a claim that regional hosting removes all foreign legal exposure. Extraterritorial disclosure powers follow the provider, not the server.
- As a reason to avoid cloud entirely when an approved region genuinely satisfies the requirement.
Reach for something else instead
- Federated learning avoids the transfer by training where the data is.
- Edge AI avoids it by running inference on the device.
- Differential privacy protects individuals in a dataset but does nothing about which law governs it.
- Self-hosted open-weight models trade capability for full control of location.
Sources & further reading
- European Commission, Adequacy decisions and international data transfers — the EU's mechanism for lawful cross-border movement.
- EDPB (2020), Recommendations 01/2020 on supplementary measures — how transfer risk is assessed after Schrems II.
- Carlini et al. (2021), Extracting Training Data from Large Language Models — why weights may carry obligations from the data behind them.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Checking where data is stored and not where it is processed, logged or monitored.
- Assuming a regional endpoint means the whole pipeline is regional.
- Treating prompts as ephemeral. Prompt content is usually retained somewhere and is often more sensitive than the source record.
- Solving it once at training time and forgetting that every inference call is a fresh transfer.
At a glance
FieldSafety & Ethics
Governslocation, not content
Distinct fromprivacy
Main leverwhere the model runs
Costcapability, since the hostable model is rarely the best one
DifficultyPractical
Where this sits
5 concepts come first. Understanding it opens up 2 more.
3Levelsteps in
5Needs firstconcepts
2Opens up1% of 256
2Areascrosses fields
Computed from the prerequisite graph, not assigned. How this works