Nature looks back at an experiment in which an AI reconstructed existing research concepts, and the harsh judgment made by the researchers behind this work.
The experiment is one of those that feeds the promise of an automated science: a system of agents responsible for starting from an open question and arriving at publishable concepts on its own. On two computer science articles targeted, he succeeded in doing so, in the sense that the ideas produced overlap with those of the original publications.
The verdict reported by Nature on August 13, 2026 is much less enthusiastic than the demonstration. When questioned, the authors of the articles in question were not impressed by what the machine had actually reconstructed, which shifts the discussion to more interesting territory than just performance: what is a rediscovery worth when the target is already in the training data, and who is in a position to judge the substance of a result?
What a rediscovery measures
The title chosen by the journal — AI is not ready to do research on itself — sums up the reservation. The methodological difficulty is real: evaluating an automated search system involves giving it problems for which the answer exists, and therefore potentially known to the model, or open problems for which no one knows how to judge the solution quickly. Neither option provides a convincing measure for the moment.
Read the full article on the original source
Go to the source →