Synthetic data for Indian startups: when it helps, when it hurts, and the DPDP question
By Vishal · published 25 October 2026
Synthetic data helps Indian startups in three situations: when a real dataset is too small or too skewed to train on, when real data cannot be shared with a vendor or partner, and when systems must be tested against cases that have not happened yet. It hurts when it copies the bias of the original or is treated as anonymous without testing. Under the DPDP Act it reduces what you hold and share, but only if generation and privacy are measured.
Where it helps
Too little data. A lending startup has 40,000 loans and 600 defaults; a manufacturer has thousands of good parts and a few dozen photographed defects. Models trained on that learn to say "fine" and be right 98% of the time while missing what matters. Synthetic examples of the rare class, generated from the real ones and checked by someone who knows what a real default or defect looks like, rebalance the training set.
Cannot share. A health-tech company wants an analytics vendor to build dashboards; an HR-tech firm wants a contractor to debug a pipeline. Sharing real records means consents, agreements and risk. A synthetic copy with the same shape and none of the people removes most of that friction.
Testing the future. Festival-season volumes you have not reached yet, a new city's address formats, a pricing change's effect on order mix. Synthetic scenarios let you test systems and train staff before the real thing arrives.
Development and demos. Realistic staging data stops engineers testing on production and stops sales showing prospects a customer's actual dashboard.
Where it hurts
Copying bias. If the real approvals data under-represents a region or a gender, the synthetic set will too, only larger and with the appearance of objectivity. Synthetic data is a mirror, not a correction, unless you deliberately rebalance and say so.
Assuming anonymity. A generator that overfits reproduces real records with small changes. The dataset is then personal data with a different label. Nearest-neighbour and membership-inference tests exist to catch this; skipping them is the most common mistake we see.
Replacing collection. No generator knows about the variable you never logged. If the real problem is that the data is not being captured, the fix is instrumentation, and we say so in the feasibility note.
How to tell good from bad
| Question | Test | Pass looks like |
|---|---|---|
| Does it look like the real data? | Per-column distributions and pairwise correlations compared | Small distances; business rules hold (city matches PIN, totals add up) |
| Does it work like the real data? | Train on synthetic, test on real | Within a few points of the real-trained model on your metric |
| Is anyone in it? | Nearest-neighbour distance, membership inference, rare-combination check | No synthetic row is a near copy; attacker accuracy near chance |
Ask any vendor, including us, for these three numbers. A scorecard is the difference between a dataset and a claim.
The DPDP question
India's Digital Personal Data Protection Act, 2023 applies to digital personal data — data about an identifiable individual. Data that cannot be linked back to a person falls outside that definition. Synthetic data generated and tested properly is a practical way to reduce the personal data a startup holds and the number of parties it shares real records with, which is data minimisation in substance. Two cautions. First, the real data used to train the generator is still personal data and needs its own lawful basis and protection. Second, "synthetic" is not a legal category; what matters is whether a specific dataset is identifiable, which is why the privacy tests above should be documented for your data-protection officer or counsel. This is engineering guidance, not legal advice.
How to start
- Pick one table or one text type where a small-data or sharing problem is costing you now.
- Define the downstream metric: model accuracy, vendor turnaround, test coverage.
- Generate, score with the three tests, and compare on that metric before generating anything else.
That is a two-week pilot, and it is how we run synthetic data projects.
Questions people ask
Is synthetic data legal to use in India?
Yes. There is no prohibition on generating or using synthetic data. The legal question is whether a given synthetic dataset still counts as personal data because it can be linked to a real person, which depends on how it was generated and tested.
Can we sell or share synthetic data with partners?
If it is properly non-identifiable, sharing it does not require the consents that sharing personal data would. Document the method and the privacy tests, and have your counsel confirm for your situation.
Will a model trained on synthetic data work on real data?
Only if the synthetic data is faithful. The standard check is train-on-synthetic, test-on-real: if the model scores close to one trained on real data, the synthetic set is good enough for that task.
Tell us the problem, not the model.
A 30-minute call and a one-page feasibility note, free. If a two-week pilot makes sense, we scope it with a fixed price. If it does not, we say so.
Related