Test data are data used to check software or an AI system before release. In conventional software development they serve to run through functions, interfaces and error cases. For AI systems, a test dataset is a held-back portion of the data that allows performance to be evaluated independently of training. The AI Act defines testing data in this sense (Art. 3(32)) and sets quality requirements for them in Art. 10.
A common mistake is copying production data into test environments with weaker protection. The control A.8.33 Test Information therefore calls for deliberate selection and protection of test data. Synthetic or anonymized data are preferred. Where real data cannot be avoided, the same access rules apply as in production, and the data are deleted after testing.