An AI discovery can become human knowledge if people can learn and transmit it

Researchers used reward-based tasks in which the best strategy required accepting early losses. Participants rarely discovered that strategy unaided, but groups initially exposed to successful AI solutions often adopted and passed it onward.

The experiment involved 1,155 people and deliberately structured learning chains. It does not establish that AI will improve every real-world culture or institution. It identifies conditions under which a machine’s contribution can persist beyond the immediate interaction.

The authors emphasize three requirements: a useful strategy that humans would struggle to discover, an advantage people can recognize, and a method they can learn. The distinction is between receiving an answer and acquiring a practice that remains available after the original solver is gone.