What it is
Synthetic data is made-up data that is realistic enough to learn from. Instead of collecting real examples from the world, such as real customer emails, real photos of roads, or real patient records, you generate examples that look and behave like the real thing. Today the generator is often an AI model: one model writes thousands of practice questions, conversations, or worked answers, and another model is trained on them.
The idea is easy to picture with a driving instructor. A learner can't wait for a child to run into the road to practise emergency stops. So the instructor sets up the situation on purpose, on a quiet lot, over and over. Synthetic data does the same thing for AI. It creates the examples you need, in the quantity you need, including the rare ones that barely appear in real life.
There are three broad ways to make it:
- Simulation. Build a virtual world with rules, such as a 3D city with physics, weather and traffic, and record what happens in it. Self-driving car companies and robotics labs have worked this way for years.
- Statistical copies. Study a real dataset, learn its patterns (how often each value appears and which values go together), and generate new rows that follow the same patterns without copying any real person. This is common in health care and finance, where the real data is private.
- Generation by a language model. Ask a strong language model to write examples, then use those examples to train or test another model. This is the kind that has exploded since 2023, and the kind this article mostly means.
A small example makes the third kind concrete. Say you want to build a customer-service bot for a bakery. You have 40 real customer emails, which is nowhere near enough to train on. So you ask a chatbot: "Write 500 different emails a bakery customer might send, covering orders, allergies, complaints, delivery and wedding cakes, in a mix of tones and writing styles." In a minute you have 500 examples. Some are good. Some are near-copies of each other. A few contain things no real bakery customer would ever say. What you do next, the checking and filtering, decides whether that pile is useful or harmful.
It's worth separating two jobs synthetic data can do. It can be training data, the examples a model learns from during fine-tuning or pretraining. Or it can be test data, the cases you use to check whether a model or a prompt works, as in evals. The risks differ. Bad training data quietly shapes the model. Bad test data gives you false confidence.
Why it matters
Real data is running out, or was never there. The largest models have already read most of the good public text on the internet. As the pretraining article explains, researchers at Epoch AI projected in 2024 that the stock of public human-written text could be used up some time between 2026 and 2032. At the other end of the scale, a small business may simply not have enough examples of its own. Synthetic data is how labs keep feeding models and how small teams get enough examples to start.
Privacy. Hospitals, banks and schools hold exactly the kind of data that would make useful AI, and they mostly can't share it, for good legal and ethical reasons. Synthetic records that keep the patterns but belong to nobody let developers build and test software without touching a real person's file. Whether a synthetic dataset really protects privacy depends on how it was made, since a careless generator can reproduce real records almost exactly, but done well it opens doors that real data keeps shut.
Rare cases. Real-world data is dominated by the ordinary. A dataset of dashcam footage is mostly empty roads in daylight. Fraud is a tiny fraction of transactions. A support inbox is mostly "where's my order?" Synthetic data lets you manufacture the rare, important situations (the cyclist at night in heavy rain, the unusual scam, the furious customer with a legal threat) in whatever quantity you need.
Cost and speed. Paying people to write and label examples is slow and expensive. The RLHF process, for instance, depends on humans comparing and rating answers. Having a model write the first draft of thousands of examples, with humans checking a sample, is far cheaper. In March 2023, Stanford's Alpaca project (covered in the fine-tuning article) generated its 52,000 training examples from 175 human-written seeds for a reported cost of under $500.
Control. With real data you get what the world gives you. With synthetic data you can decide the mix: equal numbers of each product, each tone of voice, each language, each difficulty level. That can reduce some kinds of bias in a dataset, though as we'll see it can also bake new ones in.
The catch is in the popup text for this element: it "needs checking to avoid compounding errors." A model can only generate what it already knows, in the way it tends to say it. Its mistakes, blind spots and habits go into the data, and from there into the next model. That is why synthetic data is an Advanced element: it is powerful, and it fails quietly.
Real-world examples
Waymo's ten billion simulated miles (2019)
Synthetic data is older than chatbots. Self-driving car companies were among the first to depend on it, because real driving is slow, expensive and occasionally dangerous, and the situations that matter most are the rarest.
On 10 July 2019, at TechCrunch's Sessions: Mobility event, Waymo's chief technology officer Dmitri Dolgov said the company had driven more than 10 million miles on real roads and more than 10 billion miles in simulation. That is roughly a thousand simulated miles for every real one. In simulation, engineers can replay a tricky moment from a real drive and change it: make the pedestrian step out a second earlier, add rain, swap the truck for a bus. Each variation is new training and testing material that would be impossible or reckless to collect for real.
Dolgov's point, as TechCrunch reported it, was that simulation is only as good as its realism, meaning how well simulated drivers and pedestrians behave rather than how pretty the graphics are. That is the same lesson as every synthetic dataset since. A simulation that gets the physics or the behaviour wrong teaches the car the wrong thing, very efficiently. Real-world data is what keeps the simulated world honest. The same idea powered OpenAI's robot hand that learned to manipulate a Rubik's Cube entirely in simulation, described in the reinforcement learning article.
Synthea: a million patients who never existed
Health data is valuable and locked away. Software for hospitals still has to be built and tested, and researchers and students still need realistic records to practise on. MITRE, a not-for-profit organisation that runs federally funded research centres in the United States, built an answer called Synthea.
Synthea is free, open-source software that simulates the lives of patients from birth onward: illnesses, doctor visits, medications, allergies and more, following models of how common conditions actually progress. The team described the method in the Journal of the American Medical Informatics Association in March 2018. At that point it modelled the ten most frequent reasons for primary-care visits and the ten chronic conditions with the highest burden in the United States. The project describes its output as "realistic but not real" patient data, and offers downloadable sample sets of a million synthetic patients in standard health-record formats.
This kind of synthetic data isn't generated by a chatbot. It's generated by rules written by people who understand medicine, which makes it predictable and checkable. It also shows the limits: a synthetic patient only gets the diseases and complications someone thought to model. It's excellent for testing whether software handles a record correctly, and much weaker for discovering anything new about real patients.
Small models trained on AI-written textbooks (2023 to 2024)
In June 2023, Microsoft researchers published "Textbooks Are All You Need", introducing phi-1, a 1.3-billion-parameter coding model. Its training data was about 6 billion tokens of carefully filtered web material plus about 1 billion tokens of textbooks and exercises written by OpenAI's GPT-3.5. The paper reported strong results for such a small model, and the parameters article follows the family, now on this table as Phi.
What made this a turning point was the claim behind it: that well-made synthetic text could teach more per word than raw web text. Others tried to reproduce the idea in the open. On 20 March 2024, Hugging Face published Cosmopedia, a dataset of over 30 million synthetic textbooks, blog posts, stories and how-to articles, about 25 billion tokens in all, written by the open Mixtral-8x7B-Instruct model using more than 10,000 GPU hours. The team's write-up is candid about where the effort went: most of it was prompt engineering, varying the topic, the intended audience (children, students, researchers) and the style, because the same prompt run a million times gives a million near-identical textbooks. They reported getting duplicate content below 1 percent.
In June 2024, NVIDIA went further and released a family of models, Nemotron-4 340B, specifically so that other companies could generate synthetic training data, under a licence that allowed commercial use of the output. The set included a base model, an instruction-following model to write examples, and a reward model to grade them. NVIDIA's technical report says that over 98 percent of the data used to align its own instruct model (to make it helpful and follow instructions) was synthetically generated, with only about 20,000 human-written examples in the whole process. Synthetic data had gone from a trick to an industrial supply chain.
Model collapse: what happens when models eat their own output (2024)
On 24 July 2024, the journal Nature published "AI models collapse when trained on recursively generated data" by Ilia Shumailov and colleagues from the University of Oxford, the University of Cambridge and other institutions. They asked what happens if each generation of a model is trained on text produced by the previous generation, as could happen when AI-written text fills the internet that future models are trained on.
In one experiment they fine-tuned a small language model on Wikipedia-style text, generated new text with it, trained the next model on that, and repeated. The starting prompt was about church architecture. By the ninth generation, the model's output had drifted into a repetitive list of jackrabbits with various coloured tails. The authors describe two stages. First the model forgets rare things, the "tails" of the data, meaning the unusual cases that appear only occasionally. Later it converges on bland, repetitive output with little variety left. They warn that the rare cases matter, including for fairness, because they often describe minority groups.
The study also showed the way out. When they kept 10 percent of the original human-written data in each generation's training, the degradation was minor. A separate 2024 study led by Matthias Gerstgrasser, "Is Model Collapse Inevitable?", found that when synthetic data is added alongside real data rather than replacing it, collapse did not occur in their experiments. The practical reading is not "synthetic data is poison". It's that synthetic data needs a steady anchor of real data, and that careless recycling degrades models in ways that are easy to miss because the output still looks fluent.
Distillation and the DeepSeek dispute (2025 and 2026)
Training a smaller model on a larger model's answers has a name: distillation. It's a normal, widely used technique when you own both models. It becomes contentious when the larger model belongs to someone else, because most AI companies' terms of use forbid using their output to build competing models.
In January 2025, soon after the Chinese company DeepSeek released its R1 reasoning model to wide attention, OpenAI told the Financial Times it was reviewing indications that DeepSeek may have "inappropriately distilled" its models. Microsoft was reported by Bloomberg to be investigating whether a group linked to DeepSeek had obtained OpenAI data. These were claims, and OpenAI did not make its evidence public at the time. In February 2026, OpenAI repeated the accusation in a memo to US lawmakers, saying it had seen accounts linked to DeepSeek employees trying to get around its access restrictions. At the time of writing, no court has ruled on the dispute.
The episode made a quiet issue public. Anyone who generates training data with a commercial AI tool is bound by that tool's terms. Stanford's Alpaca team in 2023 limited their model to research use partly for this reason. If you plan to use AI-generated examples to train a model you'll sell, read the current terms of the tool that generated them first.
Try it yourself
This exercise uses synthetic data the way many small businesses will: to create test questions for a help page or FAQ. You'll generate ten synthetic customer questions, then audit them. Any free chatbot works, such as ChatGPT, Claude, Gemini or Copilot. Allow 30 minutes.
- Pick a small business. Your own, or an imaginary one you know well: a dog groomer, a bike repair shop, an online candle store. Write two or three sentences about it, including what you sell, your prices if you're happy to share them, and your main policies (returns, delivery, booking).
- Generate the questions. Paste this into a new chat, with your description in the brackets:
You are helping me test a customer FAQ. Here is my business:
[two or three sentences about the business and its policies]
Write 10 questions real customers might send us by email or chat.
Make them varied: different topics, different lengths, different moods
(polite, confused, rushed, annoyed), and at least two with spelling
mistakes or missing details. Number them. Do not answer them.
- Read them all the way through. Before judging any single one, read the ten as a set. Ask yourself whether they sound like your actual customers.
- Audit for sameness. Count how many start the same way ("Hi, I was wondering…"), cover the same topic, or share the same tone. AI-generated sets are usually more alike than they first appear. Put a mark next to any near-duplicates.
- Audit for errors. Mark any question that assumes something untrue about your business: a service you don't offer, a policy you don't have, a product that doesn't exist. These are small hallucinations, and in training data they would teach a model things that aren't true.
- Audit for what's missing. Write down two or three questions your real customers ask that the AI didn't think of. Those gaps are the rare cases, the "tails", and they are exactly what synthetic data tends to lose.
- Ask for a better set. Paste this into the same chat:
Here is my audit of your 10 questions:
- Too similar: [numbers]
- Wrong about my business: [numbers and what was wrong]
- Missing topics: [your list]
Write 10 new questions that fix these problems. Make each one start
differently. Include the missing topics. Only use facts I gave you.
- Mix in real ones. Take three to five genuine questions from your inbox and add them to the set. Real examples keep the synthetic ones honest, which is the lesson of the model-collapse research in miniature.
- Use the set. You now have a small, checked test set. Try your FAQ page, your chatbot, or a new prompt against it, and see which questions it handles badly. The evals article shows how to turn this into a repeatable test.
Keep a note of what the audit found. Most people are surprised by step 4. The sameness is the most common flaw in synthetic data, and it's the hardest to see one example at a time.
Common mistakes and misunderstandings
Thinking synthetic means fake and therefore useless. "Made-up" isn't the same as "wrong". A well-designed simulation or a carefully checked generated dataset can teach a model skills it could never get from the limited real data available. The question is always how it was made and checked, not whether it was generated.
Thinking it's free information. A model can't generate knowledge it doesn't have. Synthetic data rearranges and re-expresses what the generating model already knows. It's excellent for practice, variety and format, and poor for adding new facts about the world. If the generator is wrong about something, the synthetic data is wrong about it in thousands of places.
Copying the teacher's flaws. A student model trained on a teacher model's output inherits the teacher's habits: its phrasing, its blind spots, its biases, its made-up facts. The Stanford Alpaca team said openly that their model, trained on another model's output, still made things up and repeated stereotypes. If you wouldn't trust the generator's answers directly, don't train on them without checking.
Ignoring sameness. Ask a model for a thousand examples and you'll often get a hundred ideas written ten ways. Models favour their most likely answers, which squeezes out the unusual. The Cosmopedia team spent most of their effort on this one problem. Deliberately vary topics, audiences, tones and formats, and measure how many examples are near-duplicates.
Replacing real data instead of adding to it. Model-collapse research shows the danger comes from letting synthetic data push out human data, generation after generation. Keep real examples in the mix, and keep collecting them.
Grading synthetic data with the same model that made it. If one model writes the examples and the same model judges them, its blind spots are invisible to both steps. Use a different model, a rule-based check, or a human in the loop for at least a sample.
Letting test data leak into training data. If synthetic examples accidentally resemble the questions on a public benchmark, a model trained on them can score well on that test without really being better. The Cosmopedia team checked for overlap with standard benchmarks for this reason. The same applies to your own tests: keep the examples you test with separate from the ones you train with.
Assuming synthetic data is automatically private. A generator trained on real records can sometimes reproduce a real person's details. Privacy has to be checked, not assumed, particularly with health, financial or personal data. See privacy.
Forgetting the terms of use. Most commercial AI tools restrict using their output to build competing models. The rules differ between companies and change over time. Check the current terms of the tool you generate with before training anything you plan to sell.
Related elements
- Datasets: the collections of examples synthetic data adds to, and the quality rules that apply to both.
- Pretraining: where the shortage of human-written text is felt most, and one reason labs turned to synthetic data.
- Fine-tuning: the most common use of synthetic examples, including Stanford's Alpaca.
- RLHF: human feedback that synthetic preference data increasingly supplements.
- Reinforcement Learning: agents trained in simulation, one of the oldest forms of synthetic data.
- Evals: where synthetic test cases are useful, as long as real ones sit alongside them.
- Hallucination: the made-up facts a generator can scatter through a synthetic dataset.
- Bias: how a generator's slant is copied into every example it writes.
- Privacy: the main reason synthetic records exist in health care and finance, and a risk if done carelessly.
- Phi: Microsoft's small models, built on textbook-style synthetic data.
- DeepSeek: the company at the centre of the 2025 distillation dispute.
- World Models: AI that simulates environments, a future source of synthetic experience for robots and agents.