Michael Darius Eastwood . Independent AI alignment researcher

Published 29 July 2026

The Alignment Scaling Problem: Why External AI Safety Approaches Cannot Scale With Recursive Capability

The Alignment Scaling Problem: Why External AI Safety Approaches Cannot Scale With Recursive Capability: a research paper from the ARC-Eden programme by Michael Darius Eastwood. The paper suite is registered at the Open Science Framework (DOI 10.17605/OSF.IO/6C5XB).

This paper demonstrates that current AI alignment approaches produce alignment scaling exponents of approximately zero, meaning safety degrades relative to capability as recursive depth increases. If AI capability scales super-linearly through recursive self-correction (confirmed in 95.6% of tested This companion post provides context. For the full text and figures, see the canonical paper page.

The paper is one of 23 in the ARC-Eden programme. Each paper is dated and evidence-anchored. The programme priority claims are traceable through the priority claims ledger.

Paper suite

See the Master Table of Contents. See the Predictions Observatory for testable claims.

Paper companion. See the full suite or research hub.