Insights · 5 min read

Real, synthetic or scraped: what each kind of AI training data teaches

Different data teaches different things. Know which lesson you are paying for.

Bearish AI · October 21, 2026

Scraped public data

Huge, cheap and real. But it’s the finished version of everything, and anything public may already be in a model’s training set, answers included. Research on a popular coding benchmark built from public repositories found strong signs that models had memorized parts of it.

Synthetic data

Fast to make in any volume. Useful for drilling known patterns. But a model can only generate what it already knows, and repeated training on generated data erodes the rare cases, a failure documented in Nature as model collapse.

Real process data

The record of how real work actually happened: decisions, arguments, mistakes, reviews, failures and fixes, linked together. It’s scarce, because most companies never keep it in usable form, and hard to clean, because it involves real people. That scarcity is why it’s valuable.

Use them together

The strongest training mixes all three: public data for breadth, synthetic data for drills, and real process data for judgment. Bearish AI supplies the third, and labels anything synthetic plainly, including the examples on our own samples page.