I submit the following as the AI Agency Alignment Narrowing Effect:

  1. It is broadly recognized that aligning AI systems to human values is a critical, if not the critical, engineering problem in safe AI development.

  2. One of our values we seek to align to AI systems is as follows:

    We wish to expand human agency: people’s capacity to think and decide for themselves, to feel the power to improve some part of life, and take responsibility for their own decisions.

  3. When we utilize AI for anything beyond information-searching or pattern-matching – when we trust the opinions, recommendations, or suggestions of an AI – we also transfer agency from the human realm to the machine realm. This directly contradicts the value stated in Point 2.

  4. Therefore, AI systems should more closely resemble pure information-searching and pattern-matching systems as they become more aligned with Point 2. They should feel less like agents, and more like search engines.

I call it a “narrowing effect” because Point 4 reduces the capabilities of the AI system along some dimensions: art generation, AI therapy, group deliberations, event and travel planning. It also precludes much of the “AI Agents” capabilities that seem to be the next big thing from the big AI shops.