Last checked: 7 November 2026
In short: Synthetic data is artificial data that a computer generates to copy the patterns of real data without containing any real records. Firms use it to test systems, train AI models and share data without exposing customers’ details.
Table of Contents
How it works
- Learn the patterns. A model studies a real data set, such as bank transactions, and learns how the values relate to each other.
- Generate new records. It produces fresh rows that follow the same patterns, but none belongs to a real person.
- Test the copy. The team checks that it behaves like the original and identifies no one.
Where you’ll meet it
Banks and insurers use synthetic data to test software. Tools such as SAS Data Maker generate it, and ChatGPT can produce small sample data sets on request.
What it means for your business
- Faster sharing. Suppliers can work with realistic data without months of privacy approvals.
- Fill the gaps. You can create extra examples of rare events, such as fraud, so a model has enough to learn from.
- Keep real data for the final test. Check any model built on synthetic data against real results before you rely on it.
Watch out for
- Copies of the flaws. Synthetic data repeats any bias or errors in the original.
- Privacy leaks. A poor generator can produce records close enough to real ones to identify people. Ask suppliers how they test for this.
Example
Hazy. In November 2024 the US analytics firm SAS bought the main software assets of Hazy, a London synthetic data company, to add to its SAS Data Maker product. Nationwide Building Society had used Hazy to create synthetic copies of customer transaction data so outside suppliers could test their products. In its own case study, Hazy said this cut the time to prepare and approve data for suppliers from up to six months to three days.
Synthetic data vs training data
| Synthetic data | Training data | |
|---|---|---|
| Where it comes from | Generated by a computer | Real or generated examples |
| Contains real records | No | Often yes |
| How they relate | One possible source | The wider set a model learns from |
Frequently asked questions
Is synthetic data covered by data protection law?
Usually not, if no real person can be identified from it. That depends on how it was made, so get the generation method and privacy tests reviewed.
Can a model trained only on synthetic data be trusted?
Only after testing on real data. Synthetic data can miss unusual cases or exaggerate patterns, so measure the model’s results against real outcomes first.
When does synthetic data make the most sense?
When real data is scarce, sensitive or slow to approve. Software testing, supplier trials and training models on rare events are common uses.