What is synthetic data? the AI ingredient nobody talks about

TL;DR: Synthetic data is training data generated by a model rather than collected from humans. It is now standard practice, it solves genuine problems around privacy and scarcity, and its main risk is a slow drift toward blandness when models learn mostly from other models.

Why anyone does this

Three reasons, all practical:

• Scarcity. High-quality human text on specialised topics is finite and largely already used. Generating more is the obvious lever.
• Privacy. You cannot train freely on real medical or financial records. A synthetic dataset with the same statistical shape and no real people in it sidesteps that.
• Control. Real data is imbalanced — thousands of common cases, three of the rare one that matters. Synthetic generation lets you manufacture the edge cases deliberately.

How it is actually made

Usually a strong model generates candidate examples, then a filtering stage — automated checks, verification against known answers, or a second model acting as judge — throws away the bad ones. The filtering is the whole trick. Unfiltered synthetic data makes models worse; well-filtered synthetic data reliably makes them better, which is how smaller models keep catching up to larger ones.

The failure mode: model collapse

If models train on model output, which trains the next models, distributions narrow. Rare phrasings, unusual perspectives and long-tail facts get sampled out generation by generation. Everything converges on the confident middle. Researchers call this model collapse; you experience it as AI writing that all sounds identical.

This is not hypothetical — it is a large part of why so much AI text has a recognisable flavour, and why undifferentiated AI content struggles to rank. Sameness is the product of averaging.

What it means for you

1. Original input is the scarce resource. Your own experiments, numbers, screenshots, client stories and mistakes are the things no model can generate. That is the durable differentiator for anyone publishing.
2. Verify against primary sources. As more of the web becomes synthetic, “three sources agree” means less. Check whether those three sources have a common origin.
3. Use it deliberately yourself. Synthetic data is genuinely useful at small scale: generate fifty realistic test records for a spreadsheet, or a hundred sample support tickets for a prototype, without touching customer data.

The practical version

When you need fake-but-realistic data, ask for it explicitly: specify the fields, the distribution, the edge cases you want represented, and the number of rows. Models are good at this and it is one of the least risky things you can ask for — nothing is being asserted as true. If you are cleaning or building datasets regularly, the ChatGPT plus Sheets workflow shows the same technique applied to real data.

And if you write for a living, the strategic conclusion is uncomfortable but simple: the more of the internet becomes synthetic, the more valuable it is to have actually done the thing you are writing about.