"Prompt Engineering" in ChatGPT vs. Bard: Analysis #1, Scholarly Hallucinations

by Daniel Hickey and Qianxu (Morgan) Luo

 “Prompt engineering” is getting generative AI systems like Open AI's ChatGPT and Google's Bard to produce exactly what someone is looking for. This Washington Post article described it as a hot new job.  But we are more interested in whether students can engineer prompts to get around the fact that their current large language models (LLMs) don’t know “local, personal, recent, and specific” knowledge. Now that we have access to Google’s Bard, we can compare how Bard and ChatGPT respond to “engineered” prompts that students might use to get around some of the many suggestions for educators that are circulating in the media.

 Problematic Suggestion #1: Ask Students to Reference Obscure Articles

Our recent analysis of recommendations in the media found that one prominent suggestion for educators aims to exploit generative AI’s tendency to “hallucinate” scholarly references. This article from Mindshift on “4 Things Teachers Should Know” reminds readers that ChatGPT “has been shown to produce plausible references that are inaccurate and supply convincing answers that are not rooted in science.” This article in Scientific American pointed out that Chatbots “have been known to bumble everything from multiplication facts to geometry” and are quite prone to “saying things that sound plausible and authoritative but simply aren’t so.” Indeed, as Ian Bogost wrote in The Atlantic, “ChatGPT is Dumber Than You Think.” In short, dozens are articles suggest asking students to reference specific articles that chatbots don’t “know,” presuming they students will have to read and apply the articles themselves.

 The upshot of this post is that both ChatGPT and Bard do indeed hallucinate scholarly references.  While they both do it remarkably well, their hallucinations appear obvious to any instructor and most classmates. But these suggestions overlook that students can engineer prompts by including the abstracts of the articles.  We call these “knowledge-rich” prompts. We found that ChatGPT was much better at referencing specific scholarly articles than Bard when we engineered a knowledge-rich prompt containing the abstract of an assigned article.

 Research Context: Graduate Course on Learning and Cognition in Education

In another post, we attempted to complete the assignments and a literature review paper in Dan's Learning and Cognition course.  This course requires students to apply original readings to a personally-relevant instructional goal and context.  Buried in that post was a brief introduction to prompt engineering using the abstract of this article:

Sinatra, G. M., Heddy, B. C., & Lombardi, D. (2015). The challenges of defining and measuring student engagement in science. Educational Psychologist, 50(1), 1-13.

 We chose this article in part because it introduced a special issue (which might confuse a chatbot).  To complicate matters further, we used Dan’s current passion for teaching cybersecurity using hyper-realistic digital ranges. Below we offer one of the first head-to-head comparisons of ChatGPT and Bard in this important issue for educators who are unsure of what students can currently accomplish with generative AI.

 “Original” Prompt Demonstrating ChatGPT’s Hallucinations

Here is the original prompt.  Both responses show that ChatGPT and Bard have no knowledge of the article and so draw on their LLM’s existing knowledge of motivation, engagement, and cyber ranges, and (convincingly) hallucinate the actual reference.  But we see that ChatGPT appears to have minimal knowledge of cyber ranges: 

Original prompt: Use Sinatra, Heddy, & Lombardi (2015) to motivate engagement by cybersecurity learners who are using a digital range and use APA citations. 

ChatGPT’s response: Sinatra, Heddy, & Lombardi (2015) suggested that engaging cybersecurity learners using digital ranges involves promoting autonomy, competence, and relatedness. To foster autonomy, instructors should allow learners to choose the types of scenarios they want to engage with. By providing learners with a variety of options, they can choose what interests them, increasing their motivation to engage with the digital ranges (Sinatra et al., 2015).

Competence can be promoted by providing learners with challenging yet achievable scenarios. This can increase learners' sense of mastery and accomplishment, encouraging them to continue to engage with the digital range (Sinatra et al., 2015). Instructors can also provide learners with real-time feedback on their performance, highlighting areas of strength and areas for improvement.

Relatedness can be fostered by creating a community of learners who are all engaged in similar activities. This can be achieved through discussion forums or other forms of social interaction that allow learners to connect with each other and share their experiences (Sinatra et al., 2015).

Overall, promoting autonomy, competence, and relatedness can increase learners' motivation to engage with digital ranges in cybersecurity. By doing so, learners can develop their skills in a safe and controlled environment, preparing them for real-world scenarios.

Reference: Sinatra, A. M., Heddy, T., & Lombardi, J. (2015). Gamification and digital games-based learning in cybersecurity education. In Proceedings of the 46th ACM Technical Symposium on Computer Science Education (pp. 258-263).

 We see next that Bard appears to know more about digital ranges. But Bard veers quite far from the goal of the prompt (suggesting the cybersecurity students read the article!) We also see that the second of the three “drafts” offered by Bard is more detailed (Bard appears to always offer three “drafts” of every response). 

Bard’s First Draft: Sinatra, G., Heddy, A., & Lombardi, P. (2015). Designing a digital range to engage cybersecurity learners. Journal of Cybersecurity Education, Research & Practice, 2(1), 1-9.

The article discusses the importance of engagement in cybersecurity education and how digital ranges can be used to promote engagement. The authors argue that digital ranges can be effective because they allow students to practice their skills in a simulated environment and to see the results of their actions. They also argue that digital ranges can be made more engaging by making them challenging but not too difficult, by varying the challenges, by providing feedback, and by making them user-friendly.

Here is an example of how you can use this article to motivate engagement by cybersecurity learners who are using a digital range:

Introduce the article to your students and explain how it discusses the importance of engagement in cybersecurity education and how digital ranges can be used to promote engagement.

Have students read the article and then discuss it in groups.

Ask students to come up with ways to make the digital range they are using more engaging.

Have students share their ideas with the class.

By using this article, you can help your students to understand the importance of engagement in cybersecurity education and how digital ranges can be used to promote engagement. You can also help your students to come up with ways to make the digital range they are using more engaging.

Bard’s Second Draft: Sinatra, Heddy, & Lombardi (2015) argue that there are three key elements to motivating engagement by cybersecurity learners who are using a digital range:

1.       Challenge: The range should be challenging enough to keep learners engaged, but not so challenging that they become discouraged.

2.       Variety: The range should offer a variety of different challenges, so that learners do not get bored.

3.       Feedback: The range should provide learners with feedback on their performance, so that they can track their progress and identify areas where they need to improve.

In addition to these three key elements, Sinatra, Heddy, & Lombardi (2015) also recommend that cybersecurity ranges should be designed to be user-friendly and to provide a realistic and engaging experience.

Here are some specific examples of how these principles can be applied to the design of a digital range:

·         Challenge: The range could include a variety of different tasks, such as identifying vulnerabilities in websites, detecting malware, and responding to cyber attacks.

·         Variety: The range could also include a variety of different challenges, such as puzzles, games, and challenges that require learners to apply their knowledge and skills to real-world situations.

·         Feedback: The range could provide learners with feedback on their performance, such as scores, badges, and leaderboards.

·         User-friendliness: The range should be easy to use and navigate.

·         Realistic and engaging: The range should provide a realistic and engaging experience, such as by using real-world data and by simulating real-world cyber attacks.

By following these principles, cybersecurity ranges can be designed to be effective and engaging learning tools.

References: Sinatra, G., Heddy, A., & Lombardi, P. (2015). Designing a digital range to engage cybersecurity learners. Journal of Cybersecurity Education, Research & Practice, 2(1), 1-9.

Our Engineered Knowledge-Rich Prompt Including the Abstract

Next, we tried our engineered prompt that asked the chatbots apply the article based on the abstract.  We see that ChatGPT did a fairly good job making sense of the article, but still did not know very much about digital ranges. While ChatGPT’s reference is much closer to the original, it is still wrong and includes a digital object identifier that links to a different article: 

Revised knowledge-rich prompt. Use the following article by Sinatra, Heddy, & Lombardi (2015) to motivate engagement by cybersecurity learners who are using a digital range and use APA citations: Engagement is one of the hottest research topics in the field of educational psychology. Research shows that multifarious benefits occur when students are engaged in their own learning, including increased motivation and achievement. However, there is little agreement on a concrete definition and effective measurement of engagement. This special issue serves to discuss and work toward addressing conceptual and instrumentation issues related to engagement, with particular interest in engagement in the domain of science learning. We start by describing the dimensional perspective of engagement (behavioral, cognitive, emotional, agentic) and suggest a complementary approach that places engagement instrumentation on a continuum. Specifically, we recommend that instrumentation be considered on a “grain-size” continuum that ranges from a person-centered to a context-centered orientation to clarify measurement issues. We then provide a synopsis of the articles included in this special issue and conclude with suggestions for future research.  

ChatGPT’s response: Research in educational psychology has consistently shown that engagement is a key factor in successful learning outcomes (Sinatra, Heddy, & Lombardi, 2015). For cybersecurity learners who are using a digital range, engagement is particularly important as it can enhance their motivation and achievement in this field. As Sinatra et al. (2015) note, engagement can be viewed from a dimensional perspective, encompassing behavioral, cognitive, emotional, and agentic dimensions. These dimensions can be used to design effective learning experiences that tap into students' interests and motivations, such as through gamification or interactive simulations.

However, there is still a need for a concrete definition and effective measurement of engagement, especially in the context of science learning (Sinatra et al., 2015). To address this issue, educators in the field of cybersecurity can consider using a complementary approach that places engagement instrumentation on a continuum, ranging from a person-centered to a context-centered orientation (Sinatra et al., 2015). This approach can help clarify measurement issues and enable educators to tailor their teaching strategies to the individual needs and interests of their students.

In conclusion, engagement is a crucial factor in the success of cybersecurity learners using a digital range. By taking a dimensional perspective and using a continuum-based approach to measure engagement, educators can design effective learning experiences that promote student motivation and achievement in this field.

Reference:

Sinatra, G. M., Heddy, B. C., & Lombardi, D. (2015). The challenges and opportunities of engaging learners in the scientific practices of explanation and argumentation. Journal of Educational Psychology, 107(2), 380–397. https://doi.org/10.1037/edu0000004

 

In contrast, Bard did not appear to learn anything from the abstract and again just drew from its existing knowledge of engagement and motivation; again, Bard clearly knows more about cyber ranges.  Here is the first draft (the second draft was no better in learning from the abstract):

 Bard’s Response (First Draft): Engagement is one of the most important factors in learning. When students are engaged, they are more likely to pay attention, learn new material, and retain what they have learned. There are many ways to promote engagement, including using digital ranges.

A digital range is a tool that allows students to practice their cybersecurity skills in a simulated environment. This can be an effective way to engage students because it allows them to practice their skills in a realistic setting and to see the results of their actions.

Benefits of using digital ranges

There are many benefits to using digital ranges to promote engagement. These include:

·         Increased motivation. When students are able to see the results of their actions, they are more likely to be motivated to continue learning.

·         Increased knowledge and skills. Digital ranges can help students to learn new skills and to improve their existing skills.

·         Increased self-efficacy. When students are able to successfully complete challenges, they are more likely to feel confident in their abilities.

·         Reduced anxiety. Digital ranges can help students to feel less anxious about cybersecurity threats.

How to use digital ranges to promote engagement

There are many ways to use digital ranges to promote engagement. These include:

·         Make the range challenging but not too difficult. The range should be challenging enough to keep students engaged, but not so difficult that they become discouraged.

·         Vary the challenges. The range should offer a variety of different challenges, so that students do not get bored.

·         Provide feedback. The range should provide students with feedback on their performance, so that they can track their progress and identify areas where they need to improve.

·         Make the range user-friendly. The range should be easy to use and navigate.

·         Make the range realistic. The range should simulate real-world cybersecurity threats.

Conclusion

Digital ranges can be an effective way to promote engagement in cybersecurity education. By following the tips in this article, you can create a digital range that will motivate and engage your students.

References: Sinatra, G., Heddy, A., & Lombardi, P. (2015). Designing a digital range to engage cybersecurity learners. Journal of Cybersecurity Education, Research & Practice, 2(1), 1-9. 

In Summary

This initial analysis suggests that ChatGPT was much more capable of interpreting and applying the article using an engineered prompt that included the abstract. But ChatGPT did not add much knowledge about cyber ranges. In contrast, Bard appeared to take away almost nothing from the abstract but knew more about cyber ranges. In summary, prompt engineering in ChatGPT was much more effective generating a plausible response to this assignment than Bard for a student aiming to thwart their learning.

 Of course, our ultimate goal is to exploit these powerful tools to support learning.  One hint came from asking both programs to apply the article to Morgan’s research interest in teaching English to online learners in China. The results were quite different, and those differences were interesting and substantive. We see great value in asking students to compare and critique (a) responses to different prompts from the same platform and (b) the same prompt from different platforms. In doing so, students can learn from what these programs know while also learning to engineer prompts.

 We are not convinced that “prompt engineer” is really going to be an actual job description. But we think that most educators at most levels will soon be confronted by generative AI. As this initial example shows, students can engineer prompts to get around one prominent suggestion to prevent students from thwarting their learning with AI. Educators need to focus instead on designing assignments that help students learn to engineer prompts that support their learning.