The problem is becoming acute for physical AI and world models used in robots, self-driving cars and surgical assistance, where rich real-world data is scarce and costly to produce.
The report says poor-quality training data slows development, delays products and can create unpredictable behaviour, making it harder for systems to distinguish normal conditions from rare but dangerous scenarios.
It argues the old scaling approach of simply adding more data is failing, and companies must invest in tools to analyse, clean and normalise datasets or risk limiting AI's real-world potential.
With 'junk data' hindering self-driving cars, how can companies prove their AI is truly safe before it's deployed on public roads?
As AI pivots from big data to good data, what will define the winning strategy: superior synthetic data or unparalleled real-world data collection?
The AI gold rush was for data quantity. Is the next boom in 'data refineries' that turn junk data into high-value AI fuel?
Model Collapse and the Synthetic Data Crisis: Unraveling AI’s Decline from 2024 to 2026
Overview
Between 2024 and 2026, the rapid increase in AI-generated and synthetic data caused widespread degradation in AI models, leading to an industry crisis marked by errors like hallucinations in clinical systems and vulnerabilities in autonomous and agentic AI. This collapse stems from compounding technical errors amplified through feedback loops, which erode rare but critical data needed for real-world complexity. The crisis exposed major weaknesses in tracking data origins, causing legal and quality risks, and destabilizing AI systems that rely on external data. In response, the industry is shifting from focusing on more data to prioritizing higher-quality, privacy-preserving synthetic data and collaborative training methods, aiming to rebuild trust and ensure AI’s sustainable future.