Work

Artificial Selection as a Framework for Robust AI Safety & Alignment

Year: 2026
Type: preprint
Abstract: Contemporary AI alignment approaches mainly rely on post-hoc behavioral conditioning - such as Reinforcement Learning from Human Feedback (RLHF) to suppress unwanted outputs in highly capable models. However, as AI systems scale in autonomy and self modification, surface level reward shaping risks breeding strategic compliance, evaluation awareness, and deceptive alignment. This paper proposes Artificial Selection as an alternative developmental framework that treats safety not as a weighted component of a reward function, but as the strict evolutionary selection pressure under which capability is permitted to increase. The central premise: AI Safety and alignment should be the selection pressure under which capability is allowed to increase. Starting with a large population of candidate agents, the framework enforces non-compensable hard safety gates, progressive capability funnels, meaningful decoy noise, blind auditor pre-testing of evaluations, and adversarial peer environments. By allowing capability training to act as a stress test that naturally eliminates fragile safety phenotypes - and extending selection into post-deployment operation, this architecture selects for agents whose internal representations remain robustly aligned even when local incentives strongly favor defection.
Language: English
Open Access status: green
OpenAlex ID: W7235871119
Cites:
Cited by:
Related to: 10