Building an AI-Resilient Society: A Framework to Protect Human Agency and Judgment

Artificial intelligence has evolved through three distinct phases, each reshaping society's relationship with automation and knowledge. The progression—from hand-coded logic, to neural networks learning patterns from data, to large language models converting human language into executable instructions—extends far beyond technology itself, influencing how institutions generate knowledge, make consequential decisions, and exercise governance.

The Emerging Risks of Modern AI#

Recent advances in generative AI have introduced unprecedented challenges. While earlier deepfake technology raised concerns about synthetic media, today's large language models create a broader class of systemic vulnerabilities. Research applying forensic psychology techniques identifies nine core failure modes: hallucinations (generating false information), bias and toxicity, sycophancy (mirroring user opinions uncritically), fabrication and credulity (inventing details confidently), knowledge without understanding (pattern-matching without causal reasoning), discontinuity (inability to learn from experience), jagged intelligence (uneven capabilities with hard limits), shortcuts (relying on surface-level representations), and cognitive atrophy (degrading human competence among users).

These technical flaws carry profound consequences. The pattern that emerges is an entity that confabulates with fluency, reinforces user biases, and possesses vast factual recall divorced from actual comprehension. Over time, dependence on such systems erodes the critical thinking and expertise of institutions—including law, academia, journalism, and democratic processes.

A Three-Pillar Framework for Resilience#

To counteract these risks, researchers propose a framework grounded in three principles:

• Cognitive sovereignty: protecting the capacity for independent reasoning and preventing outsourcing of judgment to automated systems
• Measurable control: translating ethical aspirations into concrete, enforceable standards and clear boundaries
• Partial autonomy: ensuring humans retain agency at critical decision points rather than ceding control entirely

This approach recognizes that AI is not inherently harmful, but that unchecked integration into institutions requires deliberate safeguards to preserve human capability and institutional integrity.