Synthetic data

Synthetic data is data generated by an AI model instead of collecting real data. They are used when there is little real data (a startup without a history), it is expensive (manual marking), confidential (medicine, finance), or specific rare scenarios are simply needed.

In a marketing context, I encountered synthetic data in two cases. The first was additional training of the request classifier: the client wanted to automatically sort incoming requests by type, but there were few real examples of each type. We generated synthetic calls via GPT-4 → additionally trained the classifier → the accuracy turned out to be higher than expected. The second was testing recommendation systems: it was necessary to test the algorithm on the behavior of thousands of users before launching the product; synthetic profiles made it possible to do this.

Risks: Synthetic data may reinforce biases in the original model (if GPT-4 generates “typical cases”, it relies on its data, not real customers). Models trained on synthetics must be validated on real data before being launched into production.