Data minimization is one of the core principles of the EU General Data Protection Regulation, which requires personal data to be “adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed” (GDPR Article 5(1)(c)). Similar principles appear in other privacy laws and frameworks.
For AI work, the principle turns into practical questions:
- which fields does this task actually need, and which can be left out
- can identifiers be replaced while the relationships the task depends on are kept
- does the AI need the original values, or only their meaning
- how long does the data need to be kept for this purpose
How is data minimization different from purpose limitation?
- Purpose limitation says personal data should be collected for specified, explicit and legitimate purposes and not further processed in ways incompatible with them.
- Data minimization says that, for a given purpose, only the necessary data should be processed.
- The two work together: the purpose defines what is necessary, and minimization limits the data to it.
Where does data minimization break in AI work?
A team sends full customer records to an external language model to summarize support tickets, because it is easier than selecting fields. The summaries need the issue, the product and the history, not names, account numbers or health details. Minimizing the input would have kept the summaries useful and the exposure smaller.
Minimization does not mean stripping data until the task fails. Removing the relationships a task depends on makes the data less useful without making the task work. When the issue is sending sensitive text to an external model, LLM Capsule replaces sensitive or controlled values with context-preserving substitutes and restores the original values in the response. Data usability covers the wider question of using data within the limits that apply.