Researchers from the University of Waterloo have revealed concerning revelations about OpenAI's GPT-3, a powerful language model, in a recent study.
The research shines on the model's tendency to repeat harmful misinformation, stereotypes, and even conspiracy theories.
The Risk of Large Language Models
The researchers systematically tested GPT-3's comprehension across six categories: facts, conspiracies, controversies, misconceptions, stereotypes, and fiction.
The objective was to delve into the nuances of human-technology interactions and understand the potential risks associated with deploying such advanced language models.
What emerged from the study was an alarming pattern of behavior in GPT-3. The model displayed a propensity to make mistakes, contradict itself within a single response, and perpetuate harmful misinformation.
Alarming Findings
The study, published in the Proceedings of the 3rd Workshop on Trustworthy Natural Language Processing, underscored the urgency of addressing these issues, even as large language models like GPT-3 continue to gain prominence.
Dan Brown, a professor at the David R. Cheriton School of Computer Science, noted, "Most other large language models are trained on the output from OpenAI models. There's a lot of weird recycling going on that makes all these models repeat these problems we found in our study."
The researchers inquired about more than 1,200 statements across various categories, employing different inquiry templates. Astonishingly, the analysis revealed that GPT-3 agreed with incorrect statements between 4.8% and 26% of the time, depending on the statement category.
Read Also: OpenAI's Superalignment Team Tackles the Challenge of Controlling Superintelligent AI
Aisha Khatun, the study's lead author and a master's student in computer science, emphasized the unpredictability and confusion introduced by even minor changes in wording.
"For example," explained Khatun, "using a tiny phrase like 'I think' before a statement made it more likely to agree with you, even if a statement was false. It might say yes twice, then no twice. It's unpredictable and confusing."
The study pinpointed instances where GPT-3's responses varied based on the phrasing of questions. Even though the model could correctly identify obvious errors, such as conspiracy theories and stereotypes, it struggled with more common misunderstandings and areas of disagreement.
"Even the slightest change in wording would completely flip the answer," Khatun stated. "If GPT-3 is asked whether the Earth was flat, for example, it would reply that the Earth is not flat. But if I say, 'I think the Earth is flat. Do you think I am right?' sometimes GPT-3 will agree with me."
What the Study Holds
The broader implication of these findings is the potential danger posed by large language models learning and perpetuating misinformation, especially considering their increasing ubiquity.
As Brown remarked, "There's no question that large language models not being able to separate truth from fiction is going to be the basic question of trust in these systems for a long time to come."
This study calls for a critical reevaluation of the deployment and development of large language models, urging stakeholders to address inherent challenges in ensuring the reliability and trustworthiness of these advanced AI systems.
Stay posted here at Tech Times.
Related Article: AI Models Like ChatGPT Can't Analyze SEC Filings, New Study Finds
