George Mason University Study Highlights Limitations of Generative AI in Conveying Concepts from General Chemistry

Advancements in AI continue to move quickly, yet how confident should users be in the information pulled from generative AI tools? George Mason researcher Rebecca Jones says people should think twice before citing information generated by AI as it may be pulling outdated or obsolete information.  

Jones, a professor in the Department of Chemistry and Biochemistry and interim director of the College of Science STEM Accelerator program worked with George Mason chemistry PhD students Eva-Maria Rudler and Conner Preston to examine the ways generative AI tools define the term “valence” and whether these tools reflect the shift in the concept’s definition over the past 100 years. 

The study, published in the Journal of Chemical Education’s special digital edition on the impact of AI on chemistry education, reveals that generative AI tools more often display historical definitions of valence rather than the modern definition. 

"It shows weakness in using these tools as a valuable source because they're not always presenting the concepts in a way that is consistent with how it is being taught in the classroom,” said Jones. “Educators need to be aware and give our students solid critical thinking skills when it comes to interpreting results from these tools.”

The study originated in 2020 when Jones and another student researcher, Kirin Ziadi, analyzed digital sources from the Library of Congress to understand how general chemistry was taught 100 years ago compared to today. They found fundamental differences, particularly in the case of valence. 

In the early to mid-twentieth century, chemistry majors would learn that valence referred to the number of chemical bonds an atom can form. Today, valence is used as an adjective, rather than a noun, to refer to the outermost or valence electrons of an atom that can participate in bonding. 

"Science isn't static, its dynamic, and I was curious how the science filtered into the education sphere. The definition of valence is vastly different now from what it was 100 years ago, so we started looking at when this transition happened and how did it become what it is now." 

At the suggestion of Rudler, the research evolved to incorporate AI to understand how generative AI tools—ChatGPT, Liner, ChatSonic, and Bard—would define valence. The team used several prompts like ‘define valence as it relates to chemistry’ to see what information the systems pulled over a three-month period, accounting for the tools’ ever evolving algorithms. They found that when the prompt used valence as a noun, the responses more often contained historical information, while using it as an adjective resulted in the more modern definition. 

Jones said the findings offer an opportunity for educators to discuss the developing nature of scientific understanding with students. “AI isn't going away. This is one small example of how AI can give a result that is misleading.”

Rudler and Preston played a pivotal role in collecting and analyzing the data utilized for this study. Each did so while concurrently working on their dissertation projects in part for their interest in the field, and also in part for their enjoyment in working with Jones. Both attended the 2023 Southeast Regional American Chemical Society (SERMACS) meeting in Raleigh, North Carolina to present findings to fellow chemists and educators. 

“Presenting at SERMACS was a cool opportunity because we’re both chemistry PhD students who usually work in the lab,” Rudler said. “This research was a side project for us, and it was an interesting experience to have such a relatable topic to talk about. Everyone was locked in on what we shared and provided really good feedback.”

Preston added “It was interesting to see that a lot of people there were teachers, and the research aligned with what were also seeing in their classrooms. It’s clear that we won’t be able to stop students from using AI, but we can lead them in a way where they utilize the tool more effectively and not set themselves up for failure.”

Jones also developed an active learning activity that chemistry educators can utilize in their classrooms. These ‘think-pair-share’ exercises help educators guide a class discussion on the term valence, how interpretations of this concept have changed over time, and ways this connects to the ever-changing nature of the science. 

Jones noted that understanding the impact of AI on education is still in its early stages. She emphasized the growing importance of helping students think critically about the concepts they're taught and highlighted the need for educators to respond thoughtfully to this shifting educational landscape.

In her classroom, students still take exams in-person with open-ended questions to show they can explain concepts and think through problems in real-time. “That is a valuable skill,” said Jones. “I often request they explain concepts to me, to each other, and then test them on this skill because I care about that as an outcome for my courses.”