Summary

Argues alignment discourse over-focuses on epistemic problems (what AI knows/believes/understands) while neglecting structural constraints on authority and actuation. Traditional approach: get AI to know/care about right things. Axio approach: establish boundaries on what AI can do regardless of knowledge/values. Key insight: epistemic alignment (better world-models, value-learning) insufficient for safety because capable optimizers find ways around preference-based constraints. Structural alignment (capability removal, non-delegable actuation, enforced inadmissibility) survives optimization pressure that epistemic approaches don’t. This doesn’t eliminate need for epistemics but subordinates it: good world-models useful within structural boundaries, but boundaries must hold independent of understanding. Practical implication: build actuation architecture first, then permit epistemic capability growth within it. Reverses typical priority (make AI smart and good, then let it act).

Key Concepts

  • Epistemic over-focus – Alignment discourse prioritizes knowledge/values over structural constraints
  • Structural priority – Boundaries on capability before/independent of epistemic quality
  • Optimization resistance – Structure survives pressure that preferences don’t
  • Architecture-first – Build actuation constraints then permit epistemic growth within them

Tags

Cross-References

Open Questions

  • Can epistemic and structural alignment be developed independently or do they interact?
  • What minimum epistemic capability required for structural boundaries to function?
  • How do we avoid structural constraints that prohibit all useful capability?