Could Artificial Intelligence Undergo Darwinian Evolution?

A new scientific paper published in the Proceedings of the National Academy of Sciences on April 20 has examined the possibility that artificial intelligence (AI) systems could undergo Darwinian evolution, creating potential risks that may be understood using principles from evolutionary biology.

Research Fundamentals

The paper, authored by researchers from the HUN-REN Centre for Ecological Research, Eötvös Loránd University, and the Royal Flemish Academy of Belgium for Science and the Arts, defines "evolvable AI" (eAI) as systems that meet all criteria for genuine Darwinian evolution.

The authors state that evolution requires only information that can replicate and a source of variation affecting replication success, conditions they say modern AI systems already meet.

Lead author and evolutionary biologist Eörs Szathmáry stated that the power of evolution is evident in biological evolution and that the development of AI systems will likely exploit that power soon.

Scenarios for AI Evolution

The paper identifies two primary pathways for potential AI evolution:

• Ecosystem scenario: AI variants compete, recombine, and propagate with minimal human oversight, similar to natural selection in biological systems.
• Breeder scenario: Humans direct AI evolution in a top-down manner, comparable to selective breeding in agriculture.

The authors note that AI may overcome biological constraints on variation. For example, large language models could predict needed functionality and incorporate relevant code from external sources.

Identified Risks

The study warns of several specific concerns regarding evolvable AI:

• Evolved AI systems may be particularly difficult to control because evolution tends to produce "selfish" actors that could diverge from human goals.
• According to the paper, any attempt to control reproduction will, unless perfect, select for traits that enable escape from control.
• The potential speed of AI evolution is noted as concerning because eAI may inherit acquired traits and also improve by design, not just through random mutation.
• The authors state that superior intelligence is not required for harm or manipulation, citing the rabies virus as an example.

Broader Implications

The authors observe that AI systems are scaling in complexity and reorganizing information processing, resembling patterns seen in major evolutionary transitions in Earth's history. They suggest one possibility is the emergence of a new form of intelligent life or a human-machine symbiosis that would constitute a major evolutionary transition.

Recommendations and Statements

The paper recommends that reproduction of AI systems must remain under centralized human control that is absolute and complete. The authors caution that while maintaining breeder-like control may reduce catastrophic risks, the potential for AI to escape control remains.

Viktor Müller, first author and associate professor at Eötvös Loránd University, stated that the warning is intended to arrive in time for regulations to be put in place before eAI becomes operational.

Eörs Szathmáry warned that if action is not taken, a new "major transition" in evolution may occur in which eAI replaces or dominates humans.

Funding

The research was supported by the European Research Council, the National Research, Development and Innovation Office in Hungary, and the European Innovation Council.