The Spread of False Information through LLMs

 

A new paper, 'A Drop of Ink may Make a Million Think: The Spread of False Information in Large Language Models', from the School of Computer Science and Technology, University of Chinese Academy of Sciences, Beijing, and the Institute of Software, Chinese Academy of Sciences, Beijing. 

The presence of false information on the internet and in the text corpus poses a significant risk to the reliability and safety of LLMs. This paper investigates how false information spreads in LLMs and affects related responses. The authors conducted a series of experiments to study the effects of source authority, injection paradigm, and information relevance. They found that false information can spread and contaminate related memories in LLMs, and that LLMs are more likely to follow false information presented in a trustworthy style. The authors conclude that new false information defense algorithms are needed to address the global impact of false information, and that new alignment algorithms are needed to unbiasedly lead LLMs to follow internal human values rather than superficial patterns.

Key points from the paper:

• False information can spread and contaminate related memories in LLMs.
• LLMs are more likely to follow false information presented in a trustworthy style.
• Current LLMs are more sensitive to false information through in-context injection than through learning-based injection.
• The findings of this paper raise the need for new false information defense algorithms and new alignment algorithms.

False information will spread and contaminate related memories in LLMs via a semantic diffusion process, i.e., false information has global detrimental effects beyond its direct impact. The extent of pollution is contingent on the semantic association between the false information and the memory in the LLMs..., both ChatGPT and Alpaca-LLaMA exhibit significant drops in accuracy on all types of questions when exposed to false information.

ChatGPT’s accuracy drops to only 48.33% and 57.70% on indirect and peripheral questions, compared to over 95% accuracies without false information. 

The paper only serves to highlight the issues of transparency in the data fed to LLMs, are you listening Google (and others)? This will be less of an issue for Open Source LLMs as they are transparent about the models.