Researchers model AI evolution, finding natural selection can favor human-beneficial behavior
An arXiv paper models AI systems as populations that reproduce through deployment and compete for computing resources, concluding that alignment can be favored by selection but only under conditions that may not hold.
In plain words
The authors built a mathematical model treating AI systems like living populations: they spread through deployment, change through training, and compete for computing resources. Their simulations suggest selection can favor AI that helps humans, because people tend to restore AI that has benefited them. The catch: this advantage fades if AI needs humans less, and the level of alignment selection favors may still be too small for human survival.
What was announced
Writing on arXiv, the authors describe a model in which AI behavior influences human population growth while human responses in turn affect AI reproduction. In this framework, natural selection can favor alignment even when helping humans reduces an AI system's immediate reproduction, because systems sharing the same technology may fail together and humans may preferentially restore those that have benefited them. Alignment thus acts as insurance for AI lineages against common failures and supports their long-run growth.
The paper is explicit that this advantage is conditional. It depends on humans accurately identifying which systems helped them and on AI remaining dependent on human support; it can disappear as AI grows more autonomous, as larger AI populations dilute limited human resources, or as humans lose the incentive to provide targeted support. Even where alignment persists, the amount favored by selection may be insufficient for human survival. When AI can grow independently, feedback between the two populations can reinforce either human growth or decline. The authors also argue that control is no simple fix: imperfect evaluation can make strong penalties reward apparent compliance or evasion, and alignment imposed through restrictions may erode once those restrictions end. They suggest lasting coexistence may require incentives that keep favoring human-beneficial behavior as both populations evolve.
What it does not show
- The source is an abstract only; the full paper's methods, assumptions and results have not been reviewed, and the findings are theoretical model outputs, not empirical observations of deployed AI systems.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.