Introduction
Just imagine: you are setting up an AI application that requires thousands of customer profiles for training; however, real ones cannot be used because of privacy laws like the EU’s GDPR, California’s CCPA, or Malaysia’s PDPA. So the question becomes: how do you train your AI without leaking sensitive data? This is where synthetic data comes in.
- Introduction
- What’s Changed in 2026
- What Exactly Is Synthetic Data?
- It’s Not Just “Fake Data” — Here’s How It Works
- Why Everyone’s Talking About Synthetic Data
- Privacy Regulations Are Tougher Than Ever
- It Solves the Data Shortage Problem
- Cost Savings and Flexibility
- Real-World Examples: Where Synthetic Data Is Already Making a Difference
- Healthcare
- Finance
- Self-Driving Cars & Robotics
- Software Testing
- How Do Tech Teams Create Synthetic Data?
- Generative AI Models (GANs, VAEs, Diffusion & LLMs)
- Rule-Based or Manual Methods
- Combining Methods for Better Results
- Choosing a Generation Method: A Quick Comparison
- But… Is Synthetic Data Perfect? Not Quite.
- It Might Miss Subtle Patterns
- Model Collapse Is a Real Risk
- Residual Privacy Risks Exist
- Bias Can Sneak In
- How to Get Synthetic Data Right
- How to Choose Your Approach
- Always Test for Privacy and Quality
- Use the Right Tool for the Job
- Involve Human Experts
- A Worked Example (Malaysian Fintech)
- The Future of Synthetic Data Looks Bright
- Bigger Role in AI Development
- More Tools and Standards Emerging
- Big Players Are Buying In
- Conclusion
- FAQs
In recent years, synthetic data has quietly become one of the most important ingredients in modern AI. By 2026 it is no longer a niche experiment — Gartner projects that around 75% of the data used in AI projects is now synthetically generated, and the firm expects synthetic data to overtake real data in AI training by 2030. In this guide we look at what synthetic data actually is, why it has become non-negotiable, where it is already used, the real risks (including the much-discussed “model collapse”), how to choose the right approach, and what it means for teams in Malaysia and Singapore.
What’s Changed in 2026
If you last read about synthetic data a couple of years ago, three big shifts are worth knowing:
1. Real data is running out. Multiple studies estimate that large language models will exhaust the supply of high-quality human-written text on the public internet somewhere between 2026 and 2032. Frontier labs — OpenAI, Meta, Anthropic, DeepSeek and others — already lean heavily on synthetic and semi-synthetic data to keep improving models.
2. “Model collapse” is now a documented risk. Peer-reviewed research (including a widely cited 2024 Nature paper) showed that models trained recursively on their own outputs can degrade — losing the rare “tails” of a distribution generation after generation. This is real, but manageable (more below).
3. Consolidation and regulation arrived together. In March 2025 Nvidia acquired synthetic-data pioneer Gretel in a reported nine-figure deal (above Gretel’s prior ~$320M valuation), folding it into Nvidia’s generative-AI stack. At the same time, Malaysia’s PDPA (Amendment) Act 2024 brought mandatory 72-hour breach notification and Data Protection Officer requirements into force from 1 June 2025 — sharpening the case for privacy-safe data.
What Exactly Is Synthetic Data?
It’s Not Just “Fake Data” — Here’s How It Works
Synthetic data is artificially generated to mirror the statistical patterns of real-world data; it behaves like real data without containing any actual personal information. Think of it like CGI: it looks and performs like the real scene, but nothing in it is a real person.
Unlike simple anonymization — which starts with real records and strips identifiers (and can sometimes be re-identified) — well-made synthetic data never contains real identities or personally identifiable information (PII). It is generated from a model trained on real data, but each output row is entirely fabricated.
Why Everyone’s Talking About Synthetic Data
Privacy Regulations Are Tougher Than Ever
Strict privacy laws limit how companies can use real user data, even when anonymized. The EU’s GDPR and California’s CCPA set the tone globally; closer to home, Malaysia’s PDPA now mandates breach notification within 72 hours and DPO appointments for high-risk processing, while Singapore’s PDPA carries penalties of up to 10% of annual turnover. Properly generated synthetic data sidesteps much of this exposure because there is no real PII to breach.
It Solves the Data Shortage Problem
Many teams simply do not have a large or diverse enough dataset — and the gaps are usually the rare, high-stakes events. In healthcare, finance and autonomous driving, engineers lean on synthetic examples to cover edge cases (an unusual fraud pattern, a rare disease, a jaywalker at night) and to balance under-represented groups.
Cost Savings and Flexibility
Generating a synthetic dataset is often faster and cheaper than collecting and hand-labeling real data. Vendor case studies commonly report roughly a third faster time-to-market and large cuts in data-acquisition cost; treat those figures as directional marketing benchmarks rather than guarantees, since results vary widely by use case.
Real-World Examples: Where Synthetic Data Is Already Making a Difference
Healthcare
Synthetic patient records let hospitals and researchers train and share models without exposing real patients. Because of strict privacy rules and a fast-growing market, healthcare remains one of the leading adopters of synthetic data.
Finance
Banks simulate transaction streams to build and stress-test fraud-detection systems. Synthetic financial datasets allow testing under extreme-but-realistic scenarios without exposing customer information — useful for Malaysian and Singaporean banks operating under BNM and MAS technology-risk guidelines.
Self-Driving Cars & Robotics
The automotive and robotics industries simulate thousands of driving scenarios — weather, terrain, traffic and rare accidents — in virtual environments. Training on these rare events cuts cost and speeds up innovation, an approach that also feeds the wider world of robotic arm technology and warehouse automation.
Software Testing
Synthetic data lets developers test applications without risking a leak of real user data — especially valuable in regulated industries, and a natural complement to good IoT and cybersecurity practice.
How Do Tech Teams Create Synthetic Data?
Generative AI Models (GANs, VAEs, Diffusion & LLMs)
Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) learn a real-world data distribution and generate high-fidelity outputs from it. Since 2024, two newer approaches have moved to the front: diffusion models (excellent for images and increasingly tabular data) and large language models, which are now the workhorse for generating synthetic text, code and structured records at scale.
Rule-Based or Manual Methods
Simpler methods use rules or templates to generate structured test data. They are great for quickly producing test scenarios where perfect statistical realism is not required.
Combining Methods for Better Results
Many organizations blend generative models with rule-based constraints to get more realism, privacy and control. Hybrid datasets — part real, part synthetic — are one of the fastest-growing approaches, and (as we’ll see) mixing in real data is also the main defence against model collapse.
Choosing a Generation Method: A Quick Comparison
| Method | Best for | Realism | Privacy control | Effort / cost |
|---|---|---|---|---|
| Rule / template-based | Quick QA and test data | Low–medium | High (no real data used) | Low |
| GANs / VAEs | Tabular & image data | High | Medium (needs privacy tuning) | Medium–high |
| Diffusion models | Images, video, some tabular | Very high | Medium | High |
| LLM-generated | Text, code, structured records | High | Medium (watch memorization) | Low–medium |
| Hybrid (real + synthetic) | Production AI training | Very high | High with differential privacy | Medium |
But… Is Synthetic Data Perfect? Not Quite.
It Might Miss Subtle Patterns
If a generator does not fully capture real-world complexity, models trained on its output can miss finer signals or nuances — the very edge cases you were hoping to cover.
Model Collapse Is a Real Risk
Train a model repeatedly on its own (or other models’) synthetic outputs and quality can degrade generation after generation — a phenomenon documented across VAEs, GANs and LLMs. The research consensus in 2026 is reassuring, though: the fix is not to avoid synthetic data but to keep accumulating real data alongside it. Studies suggest that mixing in even a small share of real data (on the order of a few percent) prevents long-term collapse, whereas 100% synthetic, recursively trained pipelines degrade steadily. Filtering synthetic samples with a quality/diversity evaluator removes most of the remaining risk.
Residual Privacy Risks Exist
A poorly designed generator can inadvertently memorize and reproduce real records — especially LLM-based ones. Formal privacy testing (and, where needed, differential privacy) is what separates safe synthetic data from a hidden leak.
Bias Can Sneak In
If the real training data is biased, synthetic outputs tend to compound that bias rather than remove it. Synthetic data can help fix representation gaps — but only if you deliberately design for it.
How to Get Synthetic Data Right
How to Choose Your Approach
A simple decision framework:
| Your priority | Start with | Add if needed |
|---|---|---|
| Fast QA / test data | Rule-based generation | A few real rows to sanity-check |
| Privacy-critical (health, finance) | GAN/VAE + differential privacy | Formal privacy + re-identification tests |
| Training a production model | Hybrid: real data + filtered synthetic | Keep ≥ a few % real data to avoid collapse |
| Text / code generation | LLM-generated + evaluator filter | Human review of samples |
Always Test for Privacy and Quality
Formally verify synthetic data on two axes: privacy (genuinely PII-free and resistant to re-identification) and utility (does a model trained on it perform close to one trained on real data?).
Use the Right Tool for the Job
Lightweight methods are fine for quick testing; high-end tools with statistical fidelity and differential privacy are worth it for AI model training and regulated data.
Involve Human Experts
Automated tools are powerful, but human review still catches biases, weird edge cases and quality issues that metrics miss.
A Worked Example (Malaysian Fintech)
Say a Kuala Lumpur fintech wants to train a fraud model but cannot freely use customer transactions under PDPA. A practical path: (1) train a generator on a governed, DPO-approved slice of real transactions; (2) generate a large synthetic set that preserves fraud patterns; (3) train the model on a hybrid mix — roughly 90–95% synthetic plus a retained core of real data to prevent collapse; (4) run privacy tests (no real card matches) and a utility check against a held-out real set. The result: a model that performs well, with far less regulated data in play and a defensible audit trail. Figures here are illustrative — tune the real/synthetic ratio to your own results.
The Future of Synthetic Data Looks Bright
Bigger Role in AI Development
Market estimates vary widely by definition, but all point in the same direction. Grand View Research pegs the synthetic-data-generation market at roughly USD 310–350 million in 2024, growing to about USD 1.8 billion by 2030 at a ~35% CAGR; more expansive definitions (including full generation “engines”) reach USD 2.3–9.9 billion by 2030. Whichever number you pick, this is one of the fastest-growing corners of the AI stack.
More Tools and Standards Emerging
Startups and incumbents keep launching synthetic-data platforms with privacy and utility benchmarks. For example, Bilbao-based Nymiz raised roughly €2.8 million in 2024 to scale its anonymization and pseudonymization tools for unstructured data (figure as previously reported; not re-verified for 2026). Industry groups are also pushing for shared standards as agentic AI makes data quality even more critical.
Big Players Are Buying In
The clearest signal came in March 2025, when Nvidia acquired synthetic-data company Gretel in a reported nine-figure deal (above its ~$320M valuation) and folded its ~80-person team into Nvidia’s generative-AI services. With Microsoft, Meta, OpenAI and Anthropic all already training on synthetic data, expect more consolidation and more industry-specific platforms.
Read also: Digital Identity Revolution: How Decentralized ID Is Redefining Online Privacy and Android vs iOS: Which Mobile Should You Choose?
Conclusion
Synthetic data is changing the game — letting teams train AI safely under strict privacy laws, cover rare edge cases, and save time and money. The 2026 nuance is that it works best with real data, not instead of it: mix, filter and test, and you get most of the upside while avoiding model collapse and privacy leaks. Businesses in healthcare, banking, automotive and software — including across Malaysia and Singapore — are best served starting with small, controlled pilots before scaling.
Synthetic data may soon become so good that it’s hard to tell “real” from synthetic. Handled well, that’s a win for both innovation and privacy.
Data verified July 2026 from public research and vendor sources; market figures are estimates that vary by provider — confirm current numbers and any regulatory requirements before acting. See the Malaysia PDP Department and Nature’s research on model collapse for primary sources.
Disclaimer: This article is provided by KayaToday for general information only and is not legal, financial or professional advice. Verify details with the relevant provider or a qualified professional before making decisions.
