Synthetic data is data that was made by a program instead of collected from the real world. In AI, it usually means text, code, images or records that another model wrote. If you have heard that "AI is now training on AI", this is what people mean.


It sounds circular, and in some cases it is a problem. In others it is a practical tool. This post explains the difference as of November 2026, and what it means if you run a business and have your own records.


What synthetic data is, in plain words


Think of the practice problems in a textbook. A teacher can write them by hand, or generate variations of a good problem: same idea, new numbers. The variations are not real exam questions from last year, but they can still teach the idea.


Synthetic data works the same way. A model is asked to write examples: questions with answers, short lessons, customer emails, made-up invoices. Those examples are then used to train or test a model.


How AI labs use it


To teach a skill with clean examples. A 2023 Microsoft Research paper, "Textbooks Are All You Need", trained a small code model called phi-1 with 1.3 billion parameters. Its training data was about 6 billion tokens of web data selected for "textbook quality", plus about 1 billion tokens of textbooks and exercises generated with GPT-3.5. The authors trained it for four days on eight GPUs and reported 50.6% on HumanEval, a coding benchmark. The point was not the score. It was that carefully chosen and generated data could do a lot with a small model. These are the authors' own results on a public benchmark, and benchmark results do not always carry over to your work.


To fill gaps. Real data is uneven. There may be plenty of common cases and very few rare ones. Generated examples can add rare cases, such as an unusual refund request or a tricky tax situation, so a model is tested on them.


To make evaluations. Labs and companies use models to write test questions, which are then checked by people or by code.


To avoid real data. When the real thing is private or regulated, a team may make look-alike records for development and demos. This is common practice in software, long before AI.


The risks


Model collapse


In 2023, researchers including Ilia Shumailov and Ross Anderson published "The Curse of Recursion: Training on Generated Data Makes Models Forget" (a version later appeared in Nature in 2024). They found that when models are trained on content that earlier models generated, the unusual cases in the original data, the "tails", disappear over generations, and the models degrade in ways that are hard to reverse. They called it model collapse. They showed it in several kinds of generative models, including language models, and suggested that genuine human data becomes more valuable as generated content spreads online.


Two honest notes. First, the research describes what happens when models are trained on generated content, including repeatedly. It does not say every use of synthetic data fails; labs mix, filter and check what they generate, and the phi-1 paper above is an example of generated data helping. Second, how much collapse matters in any specific lab's practice is something researchers continue to study. We would not claim more than that.


Bias and blind spots


A model writes what it knows how to write. If it has a blind spot, such as a region, a trade, a dialect or an accent, the examples it generates share that blind spot. Training on them can strengthen it. This is reasoning from how the method works, not a measured figure, but it is why generated data needs a check from outside the model.


Looking real without being real


Synthetic examples are plausible. That is their purpose, and also their risk. A fake customer email will not contain the strange phrasing your actual customers use. A fake invoice will not contain the mistakes your suppliers actually make. A system tested only on tidy synthetic cases can look better than it is.


Privacy is not automatic


Records generated from real ones can still reveal things about real people if they are made carelessly. "Synthetic" is not a legal category that makes personal information safe. If you are building look-alike data from customer records in Québec, Canada or the US, ask your privacy lead or lawyer how it is made and checked.


When your own data matters more


For most small and mid-sized businesses, the question is not "how do we get more training data?" You are probably not training a model. The question is "what does the model know about us when it answers?"


Models know the public world. They do not know:


  • your price list and what you really charge for a rush job;
  • which customers get net-30 and which pay up front;
  • how you describe a job, what your warranty covers, what you will not do;
  • the history of a customer: what they bought, what went wrong, who spoke to them last.

No generated data can supply that, because it was never written down anywhere a model could read. That is the real advantage of a small business: facts about its own customers and work that no one else has. When a model works from those facts at the moment it answers, results tend to be more specific than when it guesses.


This also explains why the Shumailov paper matters for you. As more of the public web is machine-written, real records with a clear origin become a more valuable thing to keep clean.


What this means for your business


1. Use synthetic data for testing and demos, and label it. Make fake customers and invoices to try a workflow. Mark them clearly so they never mix with real ones.

2. Do not tune a decision on synthetic data alone. If you are checking whether an agent handles your real work, test on real, past examples, with the answers a person already knows.

3. Keep your real records clean and in one place. Names, prices, policies, job history. This is the part a model cannot invent for you.

4. Ask vendors where their data came from. For anything trained on your behalf: how much of it is generated, how was it checked, and was any real customer data used?

5. Be careful with look-alike records made from real ones. Check with whoever handles privacy for you before you share them.


How this shows up at Humonix


Humonix does not train models, so there is no synthetic training data behind our product. What we do use is sample data for demos, and we label it. The worked examples on our site are marked as made up, and our demo environments use labelled sample data, not client records.


The real work is the other direction: Know is the one record that a client's facts live in, and Hue answers from it and shows where each answer came from, with a person approving anything that goes out. Each feature on a module page is labelled Live, Early access or Coming, so you can see what runs today. Nothing here promises a particular result.


For the bigger picture, read The Agentic Enterprise.


Sources



This post is general information, not technical or legal advice.