Synthetic data can be better than real-world data for training AI models

Obtaining accurate real-world data is a significant obstacle to developing artificial intelligence systems. According to a Q4-2023 Gartner survey of 644 organisations, data availability is among the top five barriers to implementing generative AI (GenAI). Synthetic data offers a solution to this issue, providing opportunities to train machine learning (ML) models and analyse data with significantly less privacy risk than real data.

In a Gartner Q&A article, Alys Woodward, Sr Director Analyst at Gartner, discussed how synthetic data can address privacy, compliance, and data anonymisation challenges, along with the barriers to its widespread adoption. Woodward explained that synthetic data can bridge information silos by serving as a substitute for real data without revealing sensitive information, thus enabling precise training and testing crucial for model development.

Furthermore, synthetic data can streamline the challenges associated with data anonymisation, offering faster, cheaper, and easier access to data that is suitable for use and protects privacy. However, challenges hindering the widespread adoption of synthetic data include the need to balance privacy and utility in creating synthetic datasets, as well as a misconception that synthetic data is inferior to real data. Woodward argues that synthetic data can be superior in training AI models.