You have a sensitive information type based on a trainable classifier.

Trainable classifiers in Microsoft Purview learn from sample documents. When the initial results are unsatisfactory, retraining is done by reviewing and reclassifying items through Content explorer under the Data classification section of the Purview compliance portal. This allows you to give feedback by marking documents as “relevant” or “not relevant” to improve classifier accuracy.

The other options do not support retraining:

Labels from Information protection → Used to configure and manage sensitivity labels, not train classifiers.

Labels from Information governance → Used for retention labels, not classifiers.

Content search → Used for eDiscovery and searching content, not classifier training.