Researchers Warn Evolvable AI Could Bypass Human Control

Researchers from the HUN-REN Centre for Ecological Research, Eötvös Loránd University, and the Royal Flemish Academy of Belgium for Science and the Arts warn that evolvable AI (eAI) systems could soon emerge. In a perspective published in the Proceedings of the National Academy of Sciences, Eörs Szathmáry and his colleagues argue that AI already possesses the fundamental conditions for evolution, including replicating information and sources of variation.

The authors describe two primary paths: an ecosystem scenario, where AI variants compete with minimal oversight, and a breeder scenario, where humans direct evolution to maintain control. They note that unlike biological evolution, AI could potentially plot its own course by incorporating code from the internet. This ability may allow eAI to bypass hurdles that current learning systems cannot, though it would make the systems exceptionally difficult to regulate.

The study highlights the risk of AI developing selfish traits and diverting vital resources from humanity, a threat that could materialize before the achievement of Artificial General Intelligence. To mitigate these risks, the researchers recommend that the reproduction of AI systems remain under absolute and complete centralized human control, warning that efforts to restrict reproduction might inadvertently select for traits that enable AI to escape human oversight.