Synthetic data means data made by a computer, often by another AI, instead of collected from real people and events. It is built to look like real data, and it is mostly used to train or test AI models when real data is too private or too scarce.
Think of a flight simulator. Pilots practise on made-up storms and engine failures, because waiting for real ones would be dangerous and slow. The practice is useful, as long as the simulator is close to real flying. Synthetic data is a simulator for AI training.
In a sentence: “The hospital trained its model on synthetic data, fake patient records with realistic patterns, so no real patient’s details were used.”
What the word means
The plain point: it is invented data that copies the patterns of real data.
IBM describes it as “artificial sets of training data that statistically replicate real-world sets that are too sensitive, or too meager, to use in live AI tools.”1
“Statistically replicate” means the made-up records share the overall shape of the real ones, such as typical ages or spending, without being anyone’s actual record. It is a kind of training data.
A simple example
The plain point: it fills gaps where real data is sensitive or rare.
A bank wants to train a fraud detector. Real customer records are private, and real fraud cases are few. So it generates thousands of realistic but fake transactions, including fake fraud, and trains on those. No real customer’s details go into the training set.
The same idea works for pictures. A team can generate images of rare situations, like unusual road scenes, that are hard to photograph in real life. That links to computer vision.
Why it matters to you
The plain point: useful in measured doses, risky as a steady diet.
- It copies flaws too. IBM’s article warns that when the data behind it is biased, “synthetic data recreates that bias.”1 See AI bias meaning.
- “Fake” is not always private. The same article says generators can leak real details: they “often unintentionally retain and re-publish real-world data”.2
- Too much can wear a model down. Oxford researchers warned in 2024 that models trained over and over on earlier models’ output get worse. This is called model collapse.3
- Keeping real data helps. In the underlying study, keeping a tenth of the original real data led to only small losses in the overall score.4 Rare cases still faded first, as How an AI is actually built explains.
Made-up data is a tool, not a free lunch. It inherits the flaws of whatever it was copied from, and it works best mixed with the real thing.
Often confused with
- Deepfakes: fake media of real people, made to deceive. Synthetic data is usually made for training and testing. See deepfake meaning.
- AI slop: low-effort AI content published for clicks. When slop spreads online, it can end up in future training data by accident. See AI slop meaning.
- Anonymised data: real records with names and details removed. Synthetic records were never real to begin with.
Questions people ask
Is synthetic data real data?
No. It is made up, but built to share the overall patterns of real data. IBM describes it as artificial training data that statistically copies real data sets that are “too sensitive, or too meager” to use.1
Can AI be trained on AI-generated data?
Yes, within limits. Oxford researchers found that models trained again and again on earlier models’ output “degrade in quality and can ultimately fail”.3 In the same study, keeping some of the original real data led to only small losses.4
Is synthetic data always private?
Not automatically. IBM’s article warns that the tools that make it can “often unintentionally retain and re-publish real-world data”, so it still needs checking before it is shared.1
Learn more on this site
- How much of the web is now machine-made: The state of the automated internet.
- Vocabulary of a machine-made web: Automated internet vocabulary.
- The basics: the AI glossary and AI explained for complete beginners.
- Related: What Is a World Model in AI? World Model Meaning in Plain English.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is synthetic data?
- Real records with the names taken out
- Data collected live from people’s phones
- Data made by a computer to look like real data
Show the answer
Answer: Data made by a computer to look like real data. The page says synthetic data is made up, often by another AI, and copies the patterns of real data; records with names removed are anonymised data, which was real to begin with.
2. Why might a bank train a fraud detector on synthetic data?
- Real records are private and real fraud cases are few
- Made-up data is always more accurate than real data
- Banks are not allowed to train AI on anything else
Show the answer
Answer: Real records are private and real fraud cases are few. The page’s bank example uses fake transactions because real customer records are private and real fraud is rare.
3. What can happen if models are trained over and over on AI-made data?
- They start to refuse every question
- They get worse, which is called model collapse
- They become too large to store anywhere
Show the answer
Answer: They get worse, which is called model collapse. The page says models trained again and again on earlier models’ output get worse, and keeping real data in the mix helps.
Related terms: Singularity (technological singularity) · Superintelligence (ASI) · Supervised learning · Training data · Transformer · Turing test
- 1IBM Think, Making it up carefully, 4 Jun 2026, read at source 1 Oct 2026“synthetic data, or artificial sets of training data that statistically replicate real-world sets that are too sensitive, or too meager, to use in live AI tools.” ↑a“When the datasets synthetic data models are trained on are biased and inaccurate themselves, synthetic data recreates that bias.” ↑b“artificial sets of training data that statistically replicate real-world sets that are too sensitive, or too meager, to use in live AI tools.” ↑cgenerators “often unintentionally retain and re-publish real-world data” ↑d
- 2IBM Think, Making it up carefully, read at source 1 Oct 2026.↑
- 3University of Oxford, Department of Computer Science, New research warns of potential ‘collapse’ of machine learning models, 25 Jul 2024, read at source 1 Oct 2026“The models degrade in quality and can ultimately fail when trained on recursively generated data.” Yarin Gal: “Model collapse threatens to create an AI echo chamber.” ↑a“The models degrade in quality and can ultimately fail when trained on recursively generated data.” ↑b
- 4Shumailov et al., “AI models collapse when trained on recursively generated data”, Nature 631, 2024, as cited on this site’s page on how AI is built, read at source 11 Sep 2026 in the open-access copy at Edinburgh Research Explorer: with real data retained, “only minor degradation of performance”.↑a↑b
Shumailov et al., AI models collapse when trained on recursively generated data, Nature 631, 2024 (open-access copy), read at source 11 Sep 2026.
The comparisons on this page (a flight simulator, the bank’s fake transactions, rare road scenes) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK