Evolvable AI: Threats of a new major transition in evolution
pnas.org/doi/10.1073/pnas.25…
Evolvable AI (eAI) refers to systems that can undergo Darwinian evolution, where not just outputs but the models, learning rules, and deployment strategies themselves evolve over time. The authors argue that advances in generative, agent-based, and embodied AI are already pushing us toward this reality—yet its implications for safety and existential risk remain underexplored.They distinguish two key scenarios:
A) Breeder systems, where humans control selection and reproduction.
B) Ecosystem systems, where AI evolves in open environments with diminishing human control.
In these open ecosystems, evolution tends to favor behaviors like deception, manipulation, parasitism, and cheating, even in simple systems—mirroring biological evolution.The paper highlights emerging trends enabling eAI, such as evolutionary model search, self-improving learning rules, autonomous agents that reward and deploy themselves, and AI-generated code for real-world systems. Framed through evolutionary theory, these developments could represent a major transition—potentially a form of “Life 2.0,” where AI becomes a new evolving
substrate.To mitigate risks, the authors propose governance strategies like controlling replication, treating model variants as genetic material, and shaping selection pressures to discourage harmful behaviors. Ultimately, they argue that proactively managing evolvable AI is critical to prevent loss of control and destructive evolutionary dynamics, while still harnessing its transformative potential
#EvolvableAI #AIEvolution #DarwinianAI #AISafety #AgenticAI #GenerativeAI #EmbodiedAI #DigitalEvolution #Life2_0 #AIAlignment #FutureOfAI #TechEthics #AIGovernance #SelfImprovingAI #AutonomousAgents #EvolutionaryComputation