Training on model output degrades models unless real data stays in the mix: the research behind buyers' demand for human-made data
A Nature paper (Shumailov and others, July 2024) found that "indiscriminate use of model-generated content in training causes irreversible defects in the resulting models", which it calls model collapse, and concluded that data about genuine human interactions "will be increasingly valuable" as generated text fills the web. A follow-up study (Gerstgrasser and others, April 2024) found that replacing real data with each generation's synthetic data tends toward collapse, while keeping synthetic data alongside the original real data avoids it. Generated data is therefore usable in a mix, but the real data underneath is what a lab cannot make for itself.
Dated 24 Jul 2024Source checked 30 Oct 2026Basis: indicative
“We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear. ... Indeed, the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of LLM-generated content in data crawled from the Internet.”
It is shown on its own, not counted into a figure.
Cited in this guide