Signal Signal 2: synthetic data is projected to reach 80% of AI training input by 2028

Summary

Deloitte's second watch signal concerns the changing value of data for AI. Because foundation models learn from much the same public data, owning data no longer confers an edge by itself, and older data grows less useful as conditions shift. Synthetic data, meaning AI-generated content used to train other AI, is filling gaps, with predictions that it will make up 80% of the data AI tools use by 2028, up from 20% in 2024. It has a performance ceiling, however, typically reaching 90% to 95% of the quality of real data. Worse, models trained mainly on AI-generated content risk model collapse, losing information about rare patterns, confusing concepts and producing bland, repetitive output. The report concludes that those with fresh data, such as real-time interactions and proprietary business information, hold the advantage, favouring companies that control the interaction layer.