AI-Resilient Society: arXiv Paper on Systemic Risks of Large Language Models
A recent study published on arXiv (1912.08786) emphasizes the necessity of creating an AI-resilient society, pointing out the systemic dangers associated with large language models (LLMs). The document outlines three software generations: explicit logic, neural networks, and LLMs that convert natural language into programming interfaces. Utilizing a forensic-psychology profiling approach, it identifies nine distinct characteristics of AI, such as hallucinations, bias, toxicity, sycophancy, echo chambers, fabrication, credulity, knowledge without understanding, discontinuity, inability to learn from experience, jagged intelligence, scaling limits, and shortcuts. The report underscores that these technological advancements influence not just computer science but also knowledge creation, decision-making, and governance. While GANs have led to deepfakes, LLMs introduce a new set of systemic risks. This report updates a paper first released in 2019, with the revised version noted as a replace-cross on arXiv. The authors urge society to enhance its resilience against these threats.
Key facts
- The report is available on arXiv with identifier 1912.08786.
- It describes three generations of software: explicit logic, neural networks, and large language models.
- The report uses a forensic-psychology profiling methodology to characterize AI.
- Nine documented features of AI are listed, including hallucinations, bias, toxicity, sycophancy, echo chambers, fabrication, credulity, knowledge without understanding, discontinuity, inability to learn from experience, jagged intelligence, scaling limits, and shortcuts.
- The report argues that AI shifts affect knowledge generation, decision-making, and governance.
- Generative adversarial networks (GANs) introduced deepfakes and synthetic media.
- Large language models add a new class of systemic risks.
- The paper was originally posted in 2019 and has been updated with a replace-cross announcement.
Entities
Institutions
- arXiv