Making it up carefully

In August 2024, the United States National Science Foundation gave Dr. Wei Zhai USD 75,000 to meticulously photograph, document and track the interiors of people’s houses.

He and his team of researchers, mostly his colleagues at the University of Texas San Antonio, planned to gather personal data about the interiors and exteriors of private residences in San Antonio’s Westside neighborhood and then build digital twins, or hyper-realistic simulations, of these homes.

Zhai is researching extreme heat in Westside, the hottest neighborhood in one of the hottest cities in America. With those digital twins, he and his colleagues hope to gather enough information about how houses trap heat to develop newer and more cost-effective methods of cooling. This, Zhai hopes, will help save lives in Texas, where at least 334 people and possibly hundreds more died from heat in 2023.

When Zhai pitched the project to Westside residents in 2025, they were understandably wary of the significant intrusion on their privacy this would entail. Because there simply isn’t much data on how homes are constructed and used in under-resourced communities like Westside, Zhai’s team needed to collect data on the homes’ interior layouts and properties. That required constant camera monitoring; data on temperature requiring their own finely-tuned sensors; and information about the construction of the homes themselves, which in turn required highly sophisticated LIDAR sensors to produce an accurate 3D model of the buildings from the foundation up.

Gathering the volume of real-world data necessary to build a realistic heat model of an entire neighborhood would be a challenge in any setting, but it would be especially daunting given the extensive detail they needed to collect about the most private areas of people’s homes, armed only with a USD 75,000 grant and a homespun PR campaign. Their problem, then, was twofold: convincing an understandably reluctant population to fork over private information, and then figuring out how to build a usable, highly detailed model based on what data they could glean.

To solve both problems, Zhai and his colleagues turned to an AI-powered solution: synthetic data, or artificial sets of training data that statistically replicate real-world sets that are too sensitive, or too meager, to use in live AI tools.

“We already have at least 20 homes where we’ve installed sensors, but the community is still kind of a data desert,” Zhai told IBM Think. “That’s why we’re using synthetic data to simulate data for other homes where they might not have the resources to install the sensors.”

Zhai has become an evangelist of sorts for the use of synthetic data for research in low-income communities, writing in a November 2025 essay about how it can ensure not just accuracy, but privacy for neighborhood residents. He also co-wrote an October 2025 article for the Journal of Planning Education and Research about how synthetic data can address “key challenges of privacy, reproducibility and technical feasibility.”

Those challenges apply widely to synthetic data projects that are becoming more common across various fields. Stripping data of its personal markers is appealing in finance and healthcare, where data use is increasingly regulated, but accuracy and reliability remain crucial. Similarly, synthetic data played a role in combating the COVID-19 pandemic, with researchers creating anonymized datasets that allowed them to ethically track how real-world symptoms matched associated infections.

Companies across the world are increasingly turning to synthetic data—artificially generated datasets that retain the statistical properties of real-world information—to gain a real-world business advantage while keeping private information private. Research firm IDC Futurescape predicts that in the EU privacy regulations will push 35% of organizations to rely on synthetic data by 2027, while research firm Omdia recently recommended that all cloud service providers and vendors offer synthetic data services.

“Ethics may require AI companies to use synthetic data,” University of Michigan statistics professor Ambuj Tewari wrote in an essay for The Conversation about how the technology can offer privacy benefits to companies training AI models.

But synthetic data carries its own risks, such as the potential for deanonymization, to reproduce biases that already exist in data and failing to represent outliers in real datasets. And organizations that adopt it must remain vigilant, flexible and responsible in how they put it into practice. While “reduced costs and increased safety (by simulating unsafe scenarios) have obvious economic benefits,” Tewari wrote to IBM Think in an email, synthetic data “is meant to get us faster to systems that perform well in the real, not virtual, world. So final performance and safety evaluation needs to happen in the real world.” Some researchers have questioned whether useful data can truly be anonymized, or anonymized data can truly be useful.

Synthetic data was born out of an effort to protect citizens’ privacy, a project that long predates modern AI models.

In 1993, Harvard University professor Donald Rubin wanted to analyze data from the 1990 US Census, which included details like addresses, phone and Social Security numbers. In a 1993 paper published in the Journal of Official Statistics, he proposed that agencies replace sensitive data in published datasets with statistically indistinguishable substitutes derived from a then-yet-to-be-determined method, something he argued “should have many beneficial side-effects for handling missing data in complex surveys in general.”

In an interview with IBM Think, Rubin said this innovation arose from a frustration with how other sociologists and statisticians seemed to play whack-a-mole with privacy issues as they arose, developing bespoke algorithms for each type of data.

“I have a strange brain, and I can’t stand working on lots of special problems,” Rubin said. “I wanted a big solution, and a big solution is that you never release anybody’s private data, and you don’t have any special cases. You make up the whole dataset.”

In the years to come, statisticians and computer scientists would invent numerous methods for generating useful artificial data. These include both via traditional correlation-based statistics methods and, more recently, the transformer models and generative adversarial networks (GANs) that developers now use to generate synthetic data for machine learning (ML) model training.

To generate completely new data that retains the useful properties of real-world data, researchers and engineers have traditionally used statistical models. There are two primary approaches to this. One is distribution-based, where statisticians define a function and then samples are taken from the distribution and turned into new data points. The other is correlation-based, where researchers interpolate or extrapolate new data points based on the correlation between existing ones.

With the increase in power and sophistication of modern machine learning models, researchers have new tools on hand for generating this data. GANs run a training loop where two models, one that has the real data and another that only has fake data, play a game in which the latter tests the former’s ability to distinguish between the two. The game continues until the synthetic data model’s knowledge becomes complete enough to accurately reproduce a likeness of the real-world distribution.

This approach makes the resulting dataset more complex and multi-dimensional, but less transparent. That’s because researchers often do not have access to the inner workings of the GAN model, both because of their extreme, recursive complexity and the fact that the model often belongs to an external organization that doesn’t wish to expose its inner workings.

Transformer models generate synthetic data similarly to how a large language model (LLM) generates text: by tokenizing data and analyzing it at scale until the model can reliably predict which number should come next in a sequence, capturing the dependencies between each data point. These models allow researchers to train on a much larger amount of data than with GANs or traditional statistical distributions. But transformer models are expensive to train on account of the billions of parameters they include, and as with LLMs, it is possible for errors in tokenization to distort relationships between the individual data points.

In a real-world setting, for example, researchers might want to use census data to study the relationship between migration and household income. Because census data is full of personally identifiable information (PII), the government would need to offer researchers an anonymized, synthetic dataset. The researchers would likely use a transformer model for its ability to preserve dependencies between individuals in the household, which are crucial in census data for what they reveal about demographic trends.

They would fine-tune an existing LLM to the census data by training it to predict the correct data contained there. Then, they would use that fine-tuned LLM to create a new census dataset that maintained the same relationships that are useful for researchers’ analysis while replacing real PII with made-up names, incomes and locations. In the private sector, a similar strategy might generate synthetic patient records for hospitals that train diagnostic models, profiles of potential borrowers to check lending algorithms for bias or customer profiles to train product recommendation algorithms.

Despite its versatility, synthetic data still faces a significant barrier to adoption in government agencies. October 2024 research from software developer SAS found that 32% of government decision-makers wouldn’t consider using synthetic data, significantly more than the 23% of all other respondents who said the same. One expert told Government Technology that in his experience, “government agencies abhor synthetic data” as “not real.” When dealing with these agencies, he said, he euphemizes the practice as “data shuffling.”

In 2021, however, the United States Census Bureau decided to buck that trend, announcing that it would work to integrate synthetic data into the annual American Community Survey. (The Bureau had already begun experimenting in 2018 with a synthetic version of the Survey of Income and Program Participation.) Their stated goals, as in Zhai’s project in San Antonio, were to protect citizens’ privacy in the dataset and produce higher-resolution data in more sparsely populated areas. The announcement kicked off a firestorm among researchers and pundits, with critics questioning how useful data can really be to understanding the economy once it’s stripped of its real-world characteristics.

One academic told the Associated Press that such an effort would simply “not be suitable for research.” and accused the Census Bureau of “sharply reduc[ing] public access to data.” Another, however, described how synthetic data can allow the government to “actually get far more detail into the data than with traditional methods,” particularly when trying to understand areas that aren’t particularly data-rich, like a sparsely populated rural county or a tiny neighborhood block.

The debate revealed some of the key trade-offs about the use of synthetic data: there’s no guarantee that it can fully replicate the exact level of detail of its real-world counterpart, which raises the question of whether one can make accurate decisions based on it.

Furthermore, there are numerous ways that the methods used to create synthetic datasets can unintentionally distort the real-world information: they can impose a distributional framework that’s inaccurate, twisting the data to fit an erroneous pattern; they can overfit the pattern to outlier data or miss it entirely; and, if not trained correctly, they can miss dependencies, such as the within-household correlations in census data.

Statistical errors aren’t the only risk. When the datasets synthetic data models are trained on are biased and inaccurate themselves, synthetic data recreates that bias.

Synthetic datasets can also still carry privacy risks: machine learning models used to generate synthetic data often unintentionally retain and re-publish real-world data, as in several cases where LLMs have published private API keys. Hostile parties can also deanonymize synthetic datasets through “linkage attacks,” which effectively reverse-engineer the synthetic dataset by matching its distribution to real-world data. Moreover, the use of synthetic data can lull organizations into a false sense of security, leading them to bypass or ignore necessary privacy safeguards.

These risks and more have contributed to governments across the globe slow-rolling their use of synthetic data. But the private sector has been more aggressive. And in industries like finance and healthcare, the stakes for privacy and accuracy are at least as high.

Around the same time as the US Census Bureau began to expand its work with synthetic data, JP Morgan Chase directed its AI Research Group to determine how synthetic data might be most valuable to its now USD 800-billion-plus global banking endeavors.

In a 2020 white paper, the group found that synthetic data could enable financial institutions to anonymize datasets, overcoming regulatory silos created by bills like the EU’s GDPR, or the United States’ Family Educational Rights and Privacy Act (FERPA) and Health Insurance Portability and Accountability Act (HIPAA). For example, when information about potential investment customers is walled off from a company investment branch of the bank by law, an anonymized version of the dataset could still provide value to that branch.

This ability to vault regulatory hurdles might explain why enthusiasm for synthetic data within finance is significantly higher than it is within government. According to a 2024 survey from software company SAS, 29% of banks surveyed are already using synthetic data, and 33% are “actively considering it.”

The authors of the Chase paper cite the UCI Adult dataset, which uses anonymized data to predict, based on census information, whether a person has an income of more than USD 50,000 per year—a data point with obvious implications for marketing personal financial products.

The researchers describe some different algorithmic approaches to anonymizing such data, while warning against the potential for that work to be reverse engineered. They cite an infamous case begun in 2006, when Netflix released a trove of insufficiently anonymized data about its userbase as part of a public contest to improve its recommendation engine. A group of University of Texas researchers easily unmasked Netflix customers using statistical methods, matching the data included in the Netflix contest to publicly available information on IMDb.

By the same token, hostile parties might attempt to deanonymize synthetic financial data by comparing it to easily available public information like voter registration or property records. The Chase researchers write that by making such data “privacy preserving” (with “tunable parameters for the trade-off between privacy and realism”), “human readable” (because “an overly technical representation is less likely to be relevant and foster trust”) and “compact” (ideally accessible via open-source software), it can be transferred between parties with a high level of confidence.

Recent years have seen a crop of ambitious projects attempting to replicate this level of confidence in financial data’s shareability. One example is the European Union’s effort to build a data hub within its Digital Finance Platform. The hub’s goal is to host practical tools for sharing otherwise protected financial data between banks and governments using synthetic data.  One test its researchers ran, which compared an analysis of bank losses in a simulated crisis using real data to an analysis that used synthetic data, found a difference in one relevant metric of only 0.012%.

Examples with this level of veracity can be found in almost every field where statistical analysis is useful. A study by Merck researchers found very little deviation between a synthetic dataset featuring assessments of COVID-19 vaccine effectiveness and its real-world counterpart. Census Bureau researchers compared synthetic data based on the Survey of Income and Participation Program, which provides information about household incomes and participation in government benefits, to the real survey findings and found an “imperfect but strong” association.

Joy Deng, IBM Product Manager for AI on Z and LinuxONE, described to IBM Think how the company uses prebuilt datasets to “jumpstart the model building process without violating data privacy” in financial services. She cited the example of a Singaporean financial services firm that used the sets to create a region-specific credit card fraud detection and prevention model that raised fraud detection rates to nearly 90 percent.

Still, as the regulatory landscape shifts, institutions may need to implement additional protective measures to make sure their synthetic data meets legal requirements. In a recently published pre-print, researchers from the University College London and University of California Riverside suggest that combining synthetic data with differential privacy, a framework that involves including carefully placed statistical “noise” in real datasets to mask patterns around private information, can help institutions meet the threshold for data anonymization imposed by the EU’s GDPR.

For financial institutions and governments alike, meeting those regulatory requirements is critical to avoiding millions of dollars in fines. The GDPR states that it doesn’t apply to “information which does not relate to an identified or identifiable natural person or to personal data rendered anonymous in such a manner that the data subject is not or no longer identifiable.” Businesses that meet that threshold with the use of synthetic data might avoid, for example, the fate of Meta, which was fined a record EUR 1.2 billion for insufficient protection of user data transferred from the EU to the US.

In healthcare, meeting the regulatory burden for privacy with synthetic data can help organizations save not just money, but lives.

The COVID-19 pandemic posed an enormous, unique challenge for researchers: How could they accurately model, track and predict a virus’s spread through the community without infringing on HIPAA-protected private health data? In 2021, synthetic data allowed researchers at Washington University in St. Louis to build a dataset representing 13 million real-life patients tracked during the pandemic. They converted patient information into a synthetic dataset that then allowed researchers to forecast which patients would need a ventilator, or to predict how medications patients were taking for unrelated conditions might help or impede their recovery from COVID-19. None of this would have been possible without the anonymizing power of synthetic data.

Daniel Blumenthal, who was Vice President of Strategy at the healthcare software vendor MDClone during its partnership with Washington University on its COVID-19 research, recently described to a health IT website how while true patient data is siloed within highly regulated institutions, synthetic data “often doesn’t require institutional review board approval and can be shared more freely.”

This, as Blumenthal told Healthcare IT News, “opens the door for a more diverse and inclusive group of people to contribute to healthcare innovation across academia, industry and government,” whereas data featuring real PII tends to be restricted to users within large organizations (which have the legal capability to deal with approval processes and oversight).

But, as in finance, synthetic data alone isn’t a magic bullet. In his essay for The Conversation, The University of Michigan’s Tewari explains that while synthetic data can be game-changing for privacy protection, it can’t solve every problem. For one thing, it cannot fully guarantee anonymization in the face of unmasking attacks. And if researchers don’t actively seek out these biases, creating synthetic datasets based on biased data will merely entrench them.

Tewari told IBM Think that medical researchers using synthetic data will also face issues proving its clinical and biological plausibility, given the privacy concerns associated with evaluating the accuracy of synthetic data against its real-world counterpart.

A group of South African researchers, in an article in the medical journal Clinica Chimica Acta, write that due to the highly complex, overdetermined nature of real-world health outcomes, synthetic data generators are always at risk of creating “patients that look plausible in isolation but collectively diverge from reality in important ways,” leading to models missing, for example, the correlation between an inflammatory marker and age in a certain disease. This risk, they say, requires constant iterative refinement of the model and expert analysis of synthetic data outputs.

In 2024, the Hastings Center for Bioethics published a report on synthetic data in healthcare. Despite the different field, the report placed the same emphasis on community buy-in as Dr. Zhai’s experiment with urban digital twins, writing that “careless adoption of synthetic data practices could undermine trust and lead to doubts about reliability.”

Zhai has come to understand this well. At a community workshop in January 2025, he and his team fielded questions about the heat study from nearly 100 Westside residents, determined to put them at ease with the idea of researchers installing sensors in their homes. Speaking with IBM Think, Zhai said this workshop doubled the number of participants.

Part of building that trust, he said, was ensuring that residents knew that his team were aware of the limits of the technology that they were using, and that they would always be working to refine and police their own use of it.

“We are always going to have noisy data, where it’s not accurate, or not quite clean, or just missing,” Zhai said. “In this case, where we want to make sure the data is still reliable and responsible, that’s where synthetic data can really help us. If we know what the data pattern should look like, we can use what we do get from the sensors to correct it and create an accurate set of synthetic data.”

Some researchers are even starting to use that predictive power for oracular purposes. Synthetic “counterfactual” datasets are produced by training on real-world data and then tweaking certain parameters to simulate “what-if” scenarios. Medical researchers might use this strategy to model what a diabetic population would look like if an experimental medicine were introduced, or a business might use it to see how employee attrition might change with the addition of new incentives. Researchers are also combining statistical “populations” created via synthetic data with agent-based modeling, with the goal of producing realistic models across various fields that replicate the dynamism and mutability of real populations.

These uses for synthetic data are still experimental. But they reflect both its promise in expanding public knowledge while protecting privacy, and the fundamental limitation that keeps it from being a panacea: that synthetic data requires constant testing and vigilance, and that it is only as useful as it can accurately reflect the messy, unpredictable patterns of the real world.