Privacy law says keep less. The model says keep everything. Nobody wrote down what "it" is.

Two mandates now sit in the same architecture diagram and refuse to make eye contact. One says collect only what you need. The other says feed the model and let posterity sort it out.

Data minimisation is neither new nor vague. GDPR Article 5(1)(c) requires personal data be "adequate, relevant and limited to what is necessary in relation to the purposes for which they are processed." The UK kept the identical wording. The rule rests on one phrase doing the work: the purpose. Minimisation answers a question the organization was supposed to have already asked.

Then the model arrived hungry, and the house rule became keep everything in case. Privacy professionals describe the collision politely: large training sets sit in open tension with a stated purpose, and once data is fused into model weights, deleting it can mean retraining, not a delete button. The principle assumes you can name a purpose; the engineering assumes you would rather not.

Here is the part the steering committee skips. You cannot minimise against a purpose you never defined, nor honor a retention limit on a field nobody owns. "We might need it for AI" is the absence of a purpose, wearing a roadmap. The conflict everyone frames as privacy-versus-innovation is usually definition-versus-nothing. The law did not create the tension; it exposed a blank purpose statement.

Watch which function gets handed the contradiction. It lands on the steward or privacy lead, asked to build a retention schedule for data whose purpose and owner were never written down. That is the reality tax: the cost of skipping the boring sentence. Before the next model eats the warehouse, finish the clause beginning "the purpose of this data is."