University of Waterloo Study Reveals Language Models Like GPT-3 Perpetuate Misinformation and Conspiracy Theories

 Language models such as GPT-3 have been shown to perpetuate conspiracy theories, negative stereotypes, and other misleading information in new research conducted by the University of Waterloo. The study tested more than 1,200 different statements across six categories such as facts, conspiracies, disputes, misconceptions, stereotypes, and fiction to understand how the model interprets claims and how they can avoid dangers. The researchers found that GPT-3 made mistakes, contradicted itself, and repeated harmful misinformation, agreeing with incorrect statements between 4.8 per cent and 26 per cent of the time.

In a recent study, researchers from @WaterlooMath discovered that OpenAI’s GPT-3 frequently repeats conspiracy theories, harmful stereotypes, and other forms of misinformation.

More: https://t.co/4P21gTYJMM | #UWaterlooNews pic.twitter.com/5BfM1Bdx5O

— University of Waterloo (@UWaterloo) December 20, 2023

The study warns that while large language models are always learning, the possibility of them learning misinformation is troubling and that their inability to separate truth from fiction could undermine trust in these systems. The research was published in Proceedings of the 3rd Workshop on Trustworthy Natural Language Processing. The lead author of the study, Aisha Khatun, stated that even the slightest change in wording could flip the answer and that it can be unpredictable and confusing.

The study's relevance is emphasized by the researchers, who argue that most other large language models are trained on the output from OpenAI models, which is concerning and could lead them to repeat the same problems found in the study.