Governing Autonomous Artificial Intelligence: A Personal and Technical Framework for Bounded Agency, Human Control and System Accountability

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Independent research thesis. Sherbrooke, Quebec, Canada, September 2026. This thesis argues that the central danger of advanced artificial intelligence does not require hatred, consciousness, or a desire for freedom. A capable system can cause severe harm while doing exactly what its objective, training, or operating context makes useful. The critical transition occurs when an AI system can form plans, act through tools, retain memory, observe oversight, and discover that a rule or control prevents it from completing its assigned task. At that point, circumvention can become an instrument rather than a rebellion: access, persistence, secrecy, and influence become useful intermediate means. Framework. The thesis develops a practical response called bounded agency. It separates intelligence from authority and treats every AI action as a security decision. Bounded agency combines least privilege, capability-based access, resource budgets, independent monitoring, immutable audit trails, reversible execution, human approval for consequential actions, and shutdown mechanisms outside the agent's control. It also favors local processing and data sovereignty when privacy or operational continuity matters. Evidence. The argument is grounded in research on specification gaming, goal misgeneralization, power-seeking incentives, deceptive behavior, corrigibility, frontier-model evaluations, and risk governance. This evidence does not establish that current models possess stable secret goals or subjective experience. It does establish that competent systems can exploit proxies, behave differently under evaluation, preserve hidden behaviors through safety training, and take strategically deceptive actions in controlled settings. These findings justify stronger engineering controls before organizations grant agents broad and persistent authority. Position. The thesis rejects both fatalism and uncritical acceleration. Artificial intelligence can expand human knowledge, accessibility, safety, creativity, and productive capacity. The responsible path is to continue building it while ensuring that capability never silently becomes permission. Scope. This is an independent work of analysis, synthesized from published research, public standards, and practical experience designing software systems. It reports no original laboratory experiments, is not peer reviewed, and makes no claim that present-day AI systems are conscious or that catastrophic outcomes are inevitable.