Synthetic data
Synthetic data is text or other data made by AI models and used to train other models, alongside or in place of data written by people.
Synthetic data is data that an AI model made, used to train another model. Instead of collecting more text written by people, a company asks a model to write examples: questions and answers, conversations, summaries. Then it trains on those.
For a marketer, the point is where a model's knowledge comes from. Part of what the next models learn about your category may come from text that earlier models wrote, mistakes included.
Why AI companies use it
It's cheaper and faster than collecting new human writing. It can fill gaps where there isn't much real text, like rare questions or less common languages. And it can be made to order, to teach a model a particular skill or style of answer.
What it means for brands
A model that learns from another model's output can inherit its picture of the world. If an earlier model was confused about something, that confusion can be passed on.
For example, say an older model believes Harrow & Pine only makes oak furniture, when it has made walnut tables for years. If that model writes thousands of example answers about furniture brands, and they're used for training, the mistake can carry into a newer model, even though no person ever wrote it down.
Researchers have also shown that models trained over and over on their own output get worse, a problem known as model collapse. Clear, original writing by people, about things you know first-hand, is the material that keeps models anchored.
What you can do
You can't see what a company trained on. What you can do is make the true version easy to find: say it plainly on your own site, and make sure the places others describe you say it too. When a platform searches the web before answering, those pages can correct what its training got wrong.