Glossary

Data minimization

Written with AI from the sources below and checked against our facts sheet. Last checked 7 Nov 2026. How this site is made · Spot an error?

Part of: AIGP certification and exam

Data minimization is the privacy principle that personal data must be adequate, relevant and limited to what a stated purpose needs (GDPR Art. 5(1)(c)). For AI, it applies to training data, prompts, logs and outputs. It is a balance, not a blind cut: data that genuinely makes a model safer or more accurate can be necessary.

In plain English

Packing for a trip, you take what the trip needs, not everything you own just in case. Data minimization is the same discipline for personal data. Collect, use and keep only what the purpose needs.

The GDPR states it in Art. 5(1)(c): personal data must be "adequate, relevant and limited to what is necessary" for the purposes of processing. Its sibling, storage limitation (Art. 5(1)(e)), says identifiable data should not be kept longer than needed.

AI makes this harder because models are data-hungry. The test is purpose-relative. Ask, field by field: does this input genuinely improve the model for this purpose? Then ask whether less identifying data, such as aggregated, pseudonymized or synthetic data, would do the job.

Example

Tidewell Energy, a fictional utility, builds a model to forecast neighborhood electricity demand. The data science team asks for smart-meter readings every second, customer names and full payment histories, all kept indefinitely.

The privacy team applies minimization:

Request	Decision	Reason
Readings every second	Readings every 30 minutes	Forecasts need demand patterns; second-by-second data can reveal what people do at home
Names and account numbers	Removed; meter IDs pseudonymized	The model forecasts by neighborhood, not by person
Payment histories	Not collected	Not relevant to demand forecasting
Keep indefinitely	Raw readings deleted after 24 months; aggregates kept	Storage limitation

Testing shows the 30-minute data forecasts as well as the per-second data. The team documents that result as evidence the cut was safe.

Why it matters

GDPR fines. Breaching the Art. 5 principles can bring fines of up to €20 million or 4% of worldwide annual turnover, whichever is higher (Art. 83(5)(a)). Art. 25(2) adds data protection by default: only the data necessary for each purpose is processed.
Models remember. The EDPB says an AI model trained on personal data is not automatically anonymous (Opinion 28/2024). Keeping someone out of training data is far cheaper than removing them from a trained model.
EU AI Act. A rule added by the 2026 Digital Omnibus sets a strict necessity test for bias detection. High-risk providers may process special categories of personal data only if synthetic or anonymized data would not work (Art. 4a).
US. California's CCPA requires collection and use to be reasonably necessary and proportionate to the purpose (Civil Code section 1798.100(c)).
NIST. The AI RMF names data minimizing methods, such as de-identification and aggregation, as support for privacy-enhanced AI.

How the AIGP exam tests it

Lesson II.A.2 teaches minimization alongside privacy by design. Three traps recur.

Trap 1: minimize regardless of effectiveness. Any option that cuts data "regardless of the model's accuracy" is wrong. The best answer balances privacy and utility for the stated purpose.
Trap 2: encryption as minimization. Encryption is a security control. Encrypted data you do not need is still data you should not have.
Trap 3: "just in case." Keeping raw data indefinitely for possible future models breaches minimization and storage limitation. Future projects are assessed when they are proposed.

The tell: the right option ties each data field and each retention period to the approved purpose.

Frequently asked questions

What is data minimization under the GDPR?

Art. 5(1)(c) of the GDPR requires personal data to be adequate, relevant and limited to what is necessary for the purposes of processing. In practice, collect only the fields a defined purpose needs and process them no further than needed. Pair it with storage limitation, so identifiable data is deleted when the purpose is done.

How does data minimization apply to AI training?

Justify each training feature against the model's purpose, and drop fields that add risk without improving performance for that purpose. Remove direct identifiers that are not needed. Prefer aggregated or synthetic data where it works equally well, and set retention periods for training sets, prompts and logs. Then test whether the model memorized personal data.

Does data minimization conflict with AI accuracy?

Sometimes, but the principle is purpose-relative. Data that is genuinely necessary for a safe and accurate model can be justified. What minimization rules out is data collected just in case, or kept with no defined purpose. Testing whether a smaller dataset performs as well is the usual way to settle the question.

What is the difference between data minimization and storage limitation?

Data minimization limits what personal data you collect and use: only what is adequate, relevant and necessary (Art. 5(1)(c)). Storage limitation limits how long you keep it in identifiable form: no longer than the purpose requires (Art. 5(1)(e)). An AI project needs both, a lean dataset and a deletion schedule.

Sources

Spot an error? Tell us