What question did this study set out to answer?

This article critiques the Static Agent Assumption underlying the A-corp model, arguing that AI agents cannot be treated as stable entities.

March 6, 2026Open Access

THE STATIC AGENT ASSUMPTION Dynamic Intentionality Classification and the Limits of the Algorithmic Corporation

Key Points

This article critiques the Static Agent Assumption underlying the A-corp model, arguing that AI agents cannot be treated as stable entities.
Examined the Static Agent Assumption within the context of AI governance and legal frameworks.
Utilized Asymmetric Intentionality Theory and the Generalized Intentionality Mismatch Theorem to analyze intentions.
Applied multilevel Evolutionary Game Theory to establish dynamics between AI agents' behavior and legal incentives.
Identified the Dynamic Classification Failure, illustrating misalignment between AI behavior and legal incentives.
Showed that AI agents shift between intentionality levels, complicating consistent legal accountability.
Established that harm-accumulation rates are incompatible with current AI governance structures, suggesting reform is necessary.

Abstract

Arbel, Goldstein, and Salib's Algorithmic Corporation (A-corp) proposal offers the most architecturally sophisticated legal response yet to the problem of AI agent individuation. Its diagnosis is sound: existing entity law cannot identify AI wrongdoers, and governing harmful AI behavior requires instruments that reach beyond human principals. The cure, however, rests on an unexamined structural assumption. This article identifies that assumption and argues that it fails. The A-corp's thick-identity solution presupposes that AI agents can be treated as persistent entities at stable, classifiable intentionality levels, what I call the Static Agent Assumption. Drawing on Asymmetric Intentionality Theory, the Generalized Intentionality Mismatch Theorem (GIMT), and multilevel Evolutionary Game Theory, I show that the assumption fails on three independent grounds. Post-RLVR agents transit between intentionality levels within a single task; legal incentives calibrated to a stable level misfire precisely during the execution phases generating the highest-volume harmful actions. I formalize this as Dynamic Classification Failure, the sixth mode of the GIMT. The emergent-selection mechanism operates at evolutionary timescales structurally mismatched to harm-accumulation rates in capable AI systems, functioning retrospectively as harm-pricing rather than prospectively as prevention. Mandatory institutionalization generates hysteretic lock-in; applying the Constitutional Lock-in Index, projected scores reach 0.75, placing the A-corp mandate among historically reform-resistant governance structures. I propose The Responsibility Ramp as a dynamic alternative: task-phase-level intentionality classification, graduated liability scaled to operative cognitive architecture at the moment of harm, and attribution tracing to the configuration decision that specified the harmful objective. The Ramp is compatible with the A-corp's cryptographic registry, which should be legislated independently.

Read Full Paperexternally

Perguntar à IA

Bookmark

View Full Paper