Capability Is Not Authority: Why the Danger of Superintelligence Runs Through Human Arrangements (Paperback) (ISBN-13: 9798237495775)

Capability Is Not Authority examines how increasingly capable artificial intelligence acquires consequential power and what it takes for people to retain meaningful control.

Connecting contemporary computer science with institutional analysis, the book explores transformer architectures, reasoning during inference, autonomous agents, recursive improvement, preference learning, mechanistic interpretability, and the physical infrastructure behind computation. It distinguishes demonstrated capabilities from theoretical possibilities and explains the scope and limits of mathematical arguments about superintelligence.

Its central proposal is revisable delegation, a framework for assessing which legitimate changes of direction remain achievable as a deployment accumulates commitments and dependencies. Two systems can have identical permissions and comparable performance while leaving their institutions with very different abilities to correct decisions, replace suppliers, or fulfill changing public obligations.

Through documented incidents, clearly identified hypothetical scenarios, mathematical models, and worked comparisons, the book examines risks across critical infrastructure and autonomous scientific systems. It develops practical controls alongside a research program for testing whether the proposed framework improves real decisions.

Written for technically curious readers, researchers, engineers, policymakers, and institutional leaders, Capability Is Not Authority offers a rigorous account of artificial intelligence and institutional power grounded in evidence, explicit assumptions, and accountability.

The governing question is whether those affected by consequential decisions will retain practical ways to challenge and revise them.