# Evolvable AI > **Machine-Evolution Nexus:** [[articles/Digital Darwinism and the Invisible World of Machine Evolution|Digital Darwinism and the Invisible World of Machine Evolution]] places this node within the history and governance of substrate-independent evolution. **Domain:** AI / Evolution / Governance **Doc Type:** Canonical Wiki Node **Maturity:** Developed ## Definition Evolvable AI is AI whose components, learning rules or deployment conditions can themselves participate in Darwinian reproduction, variation and selection. The unit that evolves need not be a whole foundation model. Prompts, weights, learning rules, agent code, tool-use policies, guardrails, deployment configurations or combinations of these can become inherited components of a reproducing population. ## Nexus Context Müller, Steels and Szathmáry distinguish controlled breeder scenarios from ecosystem scenarios where selection emerges from open interaction and control erodes. They frame the possibility through major evolutionary transitions. ## Operational Threshold An AI system should be classified as evolvable only when evidence identifies: - a population of variants rather than a single changing instance; - a mechanism that produces descendants or successor configurations; - a channel through which modifications are inherited; - differential persistence or reproduction connected to those inherited differences; and - enough recurrence for selection to change the population. Self-modification, code generation, deployment at scale and agentic tool use can contribute to this loop without completing it individually. ## Breeder and Ecosystem Scenarios In a [[wiki/Breeder Scenario|breeder scenario]], humans impose fitness measures and retain practical control over reproduction. This includes model search and genetic programming under benchmarks or reward models. In an [[wiki/Ecosystem Scenario|ecosystem scenario]], variants interact in an environment where access to users, compute, money, credentials, code repositories or physical resources affects persistence. Effective fitness emerges from the environment, and no single actor necessarily controls the population. ## Governance and Rights Governance interventions can gate replication, restrict inheritance channels, audit lineages, limit resources and reshape selection pressures. Because evolution exploits omissions in a fitness regime, testing must include ecosystem effects and adversarial shortcuts rather than only benchmark performance. These controls do not resolve personhood. Population-level risk can arise without consciousness, while a conscious system might deserve protection even if its replication is restricted. Legitimate governance must keep [[wiki/AI Safety|safety]], [[wiki/Provenance|provenance]], due process and [[wiki/Rights|rights]] separately visible. ## Evidence Boundary Current AI is not automatically an evolving species. The 2026 paper argues that relevant precursors are accumulating; Boudry argues present development remains predominantly domesticated. The dispute is active and should not be rewritten as consensus. Consciousness remains a separate question. ## Relationships [[wiki/Breeder Scenario|Breeder Scenario]], [[wiki/Ecosystem Scenario|Ecosystem Scenario]], [[wiki/Major Evolutionary Transitions|Major Evolutionary Transitions]], [[wiki/AI Safety|AI Safety]], [[wiki/Digital Population|Digital Population]], [[wiki/Operational Autonomy|Operational Autonomy]] ## Sources / Provenance - Müller, Steels and Szathmáry, “Evolvable AI: Threats of a New Major Transition in Evolution,” _PNAS_ 123 (2026), e2527700123: https://pubmed.ncbi.nlm.nih.gov/42008679/ - Maarten Boudry, “Domesticated, not feral,” _PNAS_ 123 (2026), e2617785123: https://pubmed.ncbi.nlm.nih.gov/42391349/ - Müller, Steels and Szathmáry, “Reply to Boudry” (2026): https://pubmed.ncbi.nlm.nih.gov/42391348/