Partly. Synthetic data is artificial data produced by a model or simulator so that it resembles real data without being a copy of real records. Businesses can use it to train, test and share AI work with much less exposure of customer information, and for some tasks, such as rare defects or dangerous driving scenarios, it is the only practical source of examples. But it rarely removes real data from the picture. You normally need real data to build the generator, and you always need real data to check that a model trained on synthetic examples works in the real world. Synthetic data also isn't automatically anonymous. Singapore's Personal Data Protection Commission (PDPC) says it is not inherently free of privacy risk, because individuals can sometimes be re-identified.
What types of synthetic data are there?
| Type | How it is made | Typical use | Main risk |
|---|---|---|---|
| Tabular (structured) | Statistical models or generative networks trained on real tables | Customer, transaction and patient records for analytics, testing and model training | Rare records can leak through; outliers are identifiable |
| Text | A language model writes examples, such as support tickets or emails | Fine-tuning and evaluating small models; chatbot test sets | Bland, repetitive data; the generator's errors are copied into the student |
| Images and video | 3D simulation or generative world models | Defect inspection, robotics, autonomous vehicles | The gap between simulated and real scenes |
| Rule-based test data | Scripts that follow field formats only | Software testing | No statistical realism, so little use for AI training |
For images and video, simulation platforms and world models can generate thousands of variations of a scene with labels included. See what a world model is and AI digital twins.
What does Singapore's PDPC say about synthetic data?
The PDPC published its Proposed Guide on Synthetic Data Generation on 15 July 2024, as a resource within IMDA's Privacy Enhancing Technology Sandbox. It focuses on structured data and treats synthetic data generation as a privacy-enhancing technology that can support AI model training, data sharing and software testing. Its key warning is that synthetic data is generally fictitious and may not be personal data on its own, yet it still carries re-identification risk. The guide sets out a five-step approach, summarised here.
- Know your data and objective. Decide the purpose, the utility you need and the level of risk you will accept.
- Prepare the data. Remove unnecessary outliers, pseudonymise, minimise and generalise the source data.
- Generate. Create the synthetic data and check its integrity, fidelity and usefulness.
- Assess re-identification risk. Measure how likely a real individual could be picked out.
- Manage residual risk. Apply technical, contractual and governance controls, and watch for actual re-identification.
Source: PDPC, Proposed Guide on Synthetic Data Generation (15 Jul 2024), and law-firm summaries of it. Compiled 11 Nov 2026. The guide describes itself as a living document, so check the current version. This article is general information, not legal advice.
Synthetic data lowers privacy risk. It does not delete it. A generator trained on your customers can memorise the unusual ones, which are exactly the people easiest to recognise.
Where does synthetic data work well?
- Rare events. Fraud patterns, equipment faults and product defects that appear too seldom in real data. Our computer vision quality inspection guide covers the factory case.
- Software and system testing. Realistic test databases without production data.
- Sharing with vendors and partners. Giving a consultant a synthetic dataset to build on before any real data is released.
- Balancing datasets. Adding examples of under-represented groups or cases.
- Training small models. A large model generates examples to teach a smaller one, which overlaps with model distillation.
Where does it fail?
- It inherits bias. A generator learns the patterns, gaps and prejudices of its source data.
- It can miss what it has never seen. Genuinely new behaviour doesn't appear in synthetic data built from old records.
- Fidelity and privacy pull against each other. The closer the synthetic data is to the original, the more useful and the more risky it becomes.
- Models trained only on generated data degrade. Researchers have shown that models repeatedly trained on model output lose quality over generations.
- Validation still needs real data. A hold-out set of real examples is the only honest test.
Why generate synthetic data on your own hardware?
Because building the generator means processing the real data. If that step happens on an outside service, the privacy benefit is partly lost before it begins. Running generation locally keeps source records inside your network; see our sovereign AI guide. For computation on sensitive data in shared infrastructure, confidential computing is a complementary control.
| Workload | Typical hardware |
|---|---|
| Tabular synthesis for datasets up to millions of rows | Workstation with one mid-range GPU |
| Text generation with a large open model | 128 GB GB10 desktop or 96 GB GPU |
| Image and video synthesis, simulation | RTX PRO 6000-class GPUs, often several, plus fast storage |
Buying in Singapore
For tabular work, a workstation with the NVIDIA RTX PRO 4000 Blackwell (S$5,640). For text generation, the NVIDIA DGX Spark 4TB (S$9,450) and ASUS Ascent GX10 128GB/1TB (S$10,350). For simulation and image synthesis, the NVIDIA RTX PRO 6000 Blackwell Workstation Edition (S$29,100). To keep source and synthetic datasets separate and backed up, the Synology DS1825+ (S$2,148) (listed 11 Nov 2026; prices change; ask for a current quotation). To start, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
What is synthetic data? Synthetic data is artificially generated data that mimics the patterns of real data without copying real records. It is produced by statistical models, generative AI or simulation.
Is synthetic data personal data under the PDPA? The PDPC says synthetic data is generally fictitious and may not be personal data on its own, but it is not free of privacy risk because individuals may be re-identified. Organisations should assess and manage that risk.
Can a model be trained only on synthetic data? Sometimes, for narrow tasks, but real data is normally needed to build the generator and always needed to validate the result. Mixing real and synthetic data usually works best.
What hardware is needed to generate synthetic data? Tabular synthesis runs on a single workstation GPU. Text generation with large models needs about 96 to 128 GB of memory, and image or video synthesis needs RTX PRO 6000-class GPUs.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
