Delegation of Authority to AI Systems: Evidence and Risks

Summary of Findings

The delegation of personal, cognitive, and decision authority to AI large language models (LLMs) is an emerging and increasingly visible social practice. Evidence from academic research, policy institutions, and empirical studies demonstrates that this practice introduces structural risks to human agency, epistemic custody, social behaviour, and governance. These risks are not speculative; they are observed, measured, and documented.

Key findings include:

• Humans increasingly delegate decisions to AI systems even when they cannot verify correctness.
• Delegation to LLMs measurably degrades social behaviours such as trust, fairness, and cooperation.
• Novice users are significantly more likely than experts to delegate authority indiscriminately.
• AI systems exhibit confidence without accountability, creating authority inversion.
• Over-reliance on AI systems leads to deskilling, cognitive offloading, and weakened oversight.
• Market incentives reward habitual dependence and compliance rather than critical engagement.
• These effects scale culturally through normalization and media reinforcement