Synthetic data is being positioned as both a scaling lever and a governance problem. Microsoft Research Asia describes a system for generating synthetic training data at scale, while the World Economic Forum argues that adoption without provenance and accountability controls will backfire.
SynthLLM: Breaking the AI 'data wall' with scalable synthetic data
Microsoft Research Asia introduced SynthLLM, a three-stage system designed to generate scalable synthetic data aimed at easing the AI training “data wall.” The approach uses graph algorithms to recombine high-level concepts drawn from pretraining corpora, producing synthetic data intended to support continued model development without relying exclusively on additional real-world collection.
The core pitch is practical: if frontier training increasingly runs into limits on high-quality, diverse, and legally usable data, a system that can synthesize new training examples from existing conceptual building blocks could reduce dependence on manual labeling and other costly data operations. Microsoft frames this as a way to keep scaling viable when “more data” becomes expensive, slow, or constrained.
- Data teams get a new scaling primitive: If synthetic generation can be operationalized reliably, it shifts some training-set expansion from acquisition and labeling to pipeline engineering and evaluation.
- Quality and traceability become first-class requirements: Recombining concepts from pretraining corpora raises immediate questions about provenance, duplication, and downstream contamination—issues that governance teams will expect to be measurable, not implied.
- Cost structure changes: The promise is a more cost-effective alternative to real-world data and manual labeling, but the true spend may migrate to validation, filtering, and monitoring for distribution drift.
Artificial intelligence and the growth of synthetic data
The World Economic Forum argues that synthetic data’s growth requires governance frameworks tailored to how synthetic datasets are produced and consumed. The piece highlights the need for robust data provenance systems and watermarking so organizations can establish accountability, track lineage, and reduce bias risks as synthetic data moves deeper into model training and analytics workflows.
A key concern is that weak controls can lead to “AI autophagy”—models increasingly trained on their own or other models’ outputs—reducing diversity and potentially amplifying errors or bias. WEF’s message is that governance is not a bolt-on compliance task; it should be treated as a strategic priority embedded in training pipelines to preserve transparency and trust.
- Provenance is becoming non-optional: If you can’t show where synthetic records came from (and how they were generated), you will struggle to defend model training decisions to auditors, regulators, and customers.
- Watermarking shifts from research to operations: Teams should expect pressure to implement watermarking or equivalent labeling to distinguish synthetic from real data across storage, sharing, and training stages.
- Bias management needs synthetic-specific controls: Synthetic data can reduce exposure to sensitive raw records, but it can also reproduce and scale bias unless governance explicitly measures and mitigates it.