As AI blurs the lines between real and synthetic data, strong governance is essential

Once a niche tool used to address data gaps or safeguard privacy, synthetic data (artificially generated information that mimics real-world data) is transforming the use of artificial intelligence (AI) in many industries. It can fill data gaps, protect privacy and enable the testing of new scenarios, providing a scalable and cost-effective alternative when real-world data is limited or sensitive.

But as the proliferation of synthetic data expands, the line between real and artificial blurs, threatening trust, distorting knowledge and embedding systemic risks.

The opportunities for using synthetic data are vast, but success will rely on strong governance, inclusive and high-quality data practices and transparent collaboration among developers, scientists, policy-makers and organizational leaders.

Synthetic data emerged as a solution when high-quality, representative real-world data was unavailable due to incompleteness, bias or privacy restrictions. It became particularly valuable in cases where the required data simply did not exist. Today, synthetic data continues to supplement real-world datasets covering underrepresented languages, health conditions and demographic groups. It enhances equity in settings ranging from clinical trials and criminal justice to financial inclusion.

But the landscape of synthetic data has expanded dramatically since those early days. Synthetic data is no longer just a tool of necessity, it’s a driver of innovation. A recent strategic brief by the World Economic Forum’s Global Future Council on Data Frontiers addressed the range of new methods for generating synthetic data that have spawned myriad novel application areas.

Entire urban environments can now be replicated for autonomous vehicle testing, for example, as self-driving car company Waymo has done. Media companies like ByteDance can generate massive new synthetic training data sets for their recommendation systems. In healthcare, synthetic patient data is being used to test treatment plans at scale without exposing medical records.

Simulated data has become particularly powerful, offering controlled environments for stress-testing financial markets, modelling climate impacts or running “digital twin” scenarios for infrastructure planning.

This promise comes with new risks. Because synthetic data is pervasive, realistic and foundational in shaping AI systems, it can become indistinguishable from authentic data sources. This creates several risks:

If the underlying data used to generate synthetic data is biased or incorrect, the results may reinforce inequities rather than reduce them – especially if the process of generating the synthetic data is biased. This challenge is reinforced by the increasing frequency with which datasets are created for the purpose of training AI systems.

As AI systems are trained on AI-generated outputs, accuracy and reliability degrade, undermining performance across domains like computer vision or natural language processing. Beyond the widely documented context of “model collapse” for generative AI, this risk can present in more subtle ways. For example, computer vision systems may be compromised if trained on rapidly proliferating AI-generated image and video data in which lighting, motion or overlapping objects are not rendered realistically.

From deepfakes to identity theft through voice cloning, unauthorized synthetic media threatens public trust in data authenticity itself. If people no longer believe what they see, hear or read, the consequences ripple far beyond technical systems.

These risks, while familiar, are magnified by the difficulty of distinguishing between AI-generated and real-world data. The benefits of synthetic data become a liability when governance is weak.

Synthetic data can create better outcomes when organizations prioritize robust governance, transparency and multi-stakeholder collaboration. Success means bridging two worlds: the developers and end users who build and apply the technology, and the executives, lawyers and policy advisors who shape its use. Each stakeholder plays a distinct governance role that no other can fulfil.

Developers and end users can drive stronger technical governance, for example, enhancing the quality and transparency of the models that generate their synthetic datasets and championing safeguards like watermarking and dataset nutrition labels.

But perhaps the most important intervention to prioritize is investing in data traceability. Robust provenance systems allow organizations to identify how and when synthetic data was introduced, aiding accountability and reducing risks like bias and AI autophagy. Given the high costs of retroactive tracing, upfront investments in robust data provenance systems should be a priority for businesses.

As is so often the case, however, technical governance is not enough. Executive and policy leadership must treat synthetic data governance as a strategic priority in its own right, rather than integrating it into broader AI governance questions.

They should pursue tailored approaches to synthetic data governance, including:

Realizing the benefits of synthetic data while mitigating known risks is a shared responsibility between engineers, policy advisors, executives and users working collaboratively and proactively. Together, these groups can realize a future that safely unlocks the immense potential of this new generation of synthetic data.