Researchers automate the boring part of making fake training data — and it actually works better

A new method lets AI systems design the instructions for generating synthetic training data, instead of requiring experts to write them by hand. This means companies can now create usable training data in fields like medicine and law without expensive human curation, and the AI-designed instructions turn out to produce better results than handwritten ones.

For years, the bottleneck in training specialized AI models has been expensive expert annotation — a human needs to read documents and mark them up, which costs money and takes time. This paper shows you can replace that with a cheaper loop: generate fake data, measure how much it helps the target model learn, feed that measurement back into the system that generates the data, and iterate. The mechanism is boring but the effect is concrete: if this works as described, it collapses the cost of training domain-specific models in medicine, law, and finance from months of expert work to weeks of computation.