Position Paper: Bounded Alignment: What (Not) To Expect From AGI Agents

Human intelligence should be the baseline for AGI safety expectations: The paper argues that current AI safety discussions often set unrealistic standards for AGI alignment. By studying how humans and animals actually achieve safety and cooperation—rather than expecting perfect alignment—we can develop more grounded expectations for what AGI systems can realistically achieve.

"Bounded alignment" is more realistic than perfect alignment: The paper's core argument is that complete alignment of AGI with human values may be impossible, and safety frameworks should instead focus on bounded or limited alignment—similar to how humans cooperate despite having different individual goals and values.

Current AI risk predictions may be overstated: By comparing AGI prospects to actual biological intelligence, the paper suggests that some alarming predictions in boardrooms and legislatures may not account for the inherent limitations and constraints that intelligence systems naturally have, leading to more measured policy decisions.

Policy decisions need better grounding in cognitive science: The paper calls for AI safety metrics and regulations to be informed by our understanding of human and animal cognition, rather than purely theoretical models of AGI behavior, which could lead to more effective and practical governance approaches.

The AI/ML community's vision of AGI requires fundamental rethinking: Rather than pursuing an idealized version of AGI alignment, the paper argues the field should adopt a more naturalistic perspective based on how intelligence actually works in biological systems.