Research Summary: Why we need an AI-resilient society

The evolution of Artificial Intelligence (AI) through three distinct generations—from explicit logic to neural networks and now large language models (LLMs)—is profoundly transforming societal knowledge generation, decision-making, and governance. LLMs, in particular, introduce new systemic risks beyond those previously seen with deepfakes and synthetic media. This report characterizes AI's current state using a forensic-psychology profiling methodology, focusing on nine identified features, including hallucinations, bias, toxicity, and sycophancy, highlighting the urgent need for an AI-resilient society.

Why it matters

The rapid advancement of AI, particularly large language models, is fundamentally reshaping societal structures, including knowledge production and governance. Understanding and mitigating the new systemic risks associated with these technologies, such as bias and hallucinations, is crucial for maintaining stability and trust in information and decision-making processes.

Key insights

• AI development has progressed through three generations: explicit logic programming, neural networks learning from data, and large language models using natural language as an interface.
• These advancements extend beyond computer science, significantly impacting how societies create knowledge, make decisions, and self-govern.
• While generative adversarial networks (GANs) introduced risks like deepfakes, large language models (LLMs) present a novel class of systemic risks.
• A forensic-psychology profiling methodology has been applied to characterize AI, identifying nine documented features such as hallucinations, bias, toxicity, and sycophancy.
• There is a recognized need to develop societal resilience against the challenges posed by evolving AI technologies.

Source

arXiv — Computers and Society — https://arxiv.org/abs/1912.08786