Paper Type

ERF

Abstract

Public-sector education increasingly relies on artificial intelligence (AI) deployed through cloud-centered large language model (LLM) architectures that externalize inference, alignment, and control beyond institutional boundaries. This creates a structural tension: institutions remain accountable for AI behaviour they do not govern. We address this problem by conceptualizing responsible human-AI collaboration as a question of architectural governability, defined as the extent to which AI behaviour can be directed, audited, and adapted within institutional settings. We identify four conditions: locality, modularity, governance autonomy, and socio-technical fit, that structure control over agentic behaviour. By contrasting cloud centered LLM architectures with local small language model (SLM)-based multi-agent architectures, we show how alternative designs redistribute control, dependence, and alignment capacity. The paper advances a structural account of responsibility, positioning architectural design as a primary determinant of whether AI systems can be governed in contexts where accountability cannot be delegated.

Paper Number

1944

Comments

SIG SVS

Download

Share

COinS
 
Aug 15th, 12:00 AM

Responsible Human-AI Collaboration as an Architectural Problem: Toward Architectural Governability

Public-sector education increasingly relies on artificial intelligence (AI) deployed through cloud-centered large language model (LLM) architectures that externalize inference, alignment, and control beyond institutional boundaries. This creates a structural tension: institutions remain accountable for AI behaviour they do not govern. We address this problem by conceptualizing responsible human-AI collaboration as a question of architectural governability, defined as the extent to which AI behaviour can be directed, audited, and adapted within institutional settings. We identify four conditions: locality, modularity, governance autonomy, and socio-technical fit, that structure control over agentic behaviour. By contrasting cloud centered LLM architectures with local small language model (SLM)-based multi-agent architectures, we show how alternative designs redistribute control, dependence, and alignment capacity. The paper advances a structural account of responsibility, positioning architectural design as a primary determinant of whether AI systems can be governed in contexts where accountability cannot be delegated.