El País asks if OpenAI models are quietly learning user ideas

This piece captures the growing unease among technical users that interacting with proprietary models may leak their best ideas back into the training pool. El País uses the Navier–Stokes claim and alleged reuse of Anthropic research as hooks to ask a deeper question: if prompts from top scientists are used as data, do labs effectively convert individual insight into proprietary model weights, without attribution or control. Experts interviewed note that modern training pipelines are opaque enough that neither users nor outside auditors can track whether a given line of reasoning has been absorbed. ([elpais.com](https://elpais.com/tecnologia/2026-09-19/nos-roba-la-ia-las-ideas-no-se-puede-demostrar-al-100-pero-es-logico-pensar-que-si.html))

For the race to AGI, this kind of distrust could push high value research communities toward open or on premises models, even if commercial frontier systems are more capable. If cutting edge mathematicians, biotech firms and defense labs conclude that using public APIs means seeding their hard won structures into a competitor’s AGI, they will either negotiate strict data use contracts or build in house stacks. That would fragment the data landscape and potentially slow the feedback loop that big labs rely on to refine reasoning agents. It also strengthens the argument for stronger transparency obligations around training data, opt outs and how user interactions are or are not folded back into model updates.