Large language models (LLMs) and the institutionalization of misinformation

Area of Science:

• Psychological Science
• Artificial Intelligence Ethics

Background:

• Large language models (LLMs) like ChatGPT generate vast amounts of online information, blending truth with falsehood.
• The persuasive nature of LLM-generated content, informed by psychological science, can lead users to accept misinformation as fact.

Purpose of the Study:

• To examine the potential impact of LLM-generated misinformation on users' reality monitoring capabilities.
• To explore the feedback loop where user-generated misinformation is adopted by emerging LLMs, exacerbating the problem.

Main Methods:

• Analysis of LLM content generation techniques.
• Review of psychological principles influencing belief formation.
• Conceptual modeling of misinformation propagation through LLM networks.

Main Results:

• LLMs can disseminate convincing misinformation at scale.
• A feedback loop exists where misinformation can be amplified by LLMs.
• Erosion of reality monitoring may occur, leading to reliance on false information.

Conclusions:

• The proliferation of LLM-generated misinformation poses a significant threat to individual and societal functions.
• Loss of reality monitoring can lead to flawed decision-making, decreased trust in institutions, and societal fragmentation.