Alignment researcher proposes 'pre-aligned' AI to reverse safety-capability trade-off

The researcher begins by referencing prior work on 'explicit strong generalisation' (linking to two earlier posts) and states that once such generalisation is working, the dream would be to build 'pre-aligned generalising AIs.' The core argument pivots on the standard narrative of AI safety: that there is a fundamental conflict between alignment (doing the right thing) and capabilities (doing the easy, powerful thing). This narrative is typically told as a story where good people are at a constant disadvantage—they must carefully plan every AI advance and watch for dangers, while those without scruples can pursue raw power without responsibility. The researcher acknowledges nuance exists in this story but explicitly rejects taking that nuance route. Instead, the goal is to invert the entire dynamic. The vision is to create AIs such that the good people can 'YOLO and reap the rewards of increased AI capabilities,' while bad actors have to carefully plan, limit their AIs, and constantly restrict what their systems can do. A pre-aligned AI is then introduced as an AI whose morality is bound to empirical concepts. The post cuts off before fully developing this concept, but the framing suggests that by embedding moral reasoning into the empirical nature of the system (rather than treating it as an external constraint), it becomes possible to align capability-seeking with ethical behavior rather than oppose them. This would be a fundamental inversion of the current incentive structure in AI development.

The post addresses a foundational asymmetry in AI safety discourse: the idea that virtue is expensive and vice is free. Standard narratives depict responsible AI developers as perpetually on guard, required to scrutinize every capability gain for potential misuse, while malicious actors face no such friction—they can maximize power-seeking without restraint. The researcher frames pre-aligned AI as a solution to this structural problem. Rather than treating alignment and capability as opposing forces (a zero-sum constraint model), pre-alignment would embed moral reasoning into the empirical substrate of the model itself, making ethical and powerful behavior aligned rather than opposed. This is positioned as part of a broader research agenda on 'value generalisation'—the ability to reliably transfer learned values to new domains. The proposal is speculative and incomplete in this post, but its goal is clear: to create a regime where the safety-conscious actor is the one free to pursue aggressive capability gains, and the bad actor is forced into a costly posture of constant limitation and restriction.

No comments yet. Be the first to share your thoughts!