Training data

The quality, composition, and governance of training data are the single largest determinant of an AI model's behavior. A model trained on biased data produces biased outputs. A model trained on inaccurate data produces inaccurate outputs. A model trained on improperly authorized data creates legal liability regardless of how good its outputs are.

Training data comes from many sources: publicly available datasets, web scrapes, proprietary enterprise data, licensed third-party data, and increasingly, synthetic data generated by other AI models. The choice of training data involves tradeoffs between coverage, quality, recency, and legal clearance.

For privacy compliance, training data governance involves three primary questions:

• Was this data collected with appropriate consent and under a valid legal basis?
• Is using it for model training compatible with the purpose under which it was collected?
• Does using it comply with applicable regulations, including the GDPR's purpose limitation principle and the EU AI Act's training data documentation requirements?

Organizations that cannot answer these questions face exposure from regulators, data subjects, and copyright holders whose work may have been used without authorization.