The advent of “synthetic text-generating chatbots” has impacted writing and, more specifically, student writing. Those impacts have ranged from apocalyptic to promising collaborative support. Pundits have raised questions about the need for students to write pieces if the AI bots can write for them. What are the advantages of using synthetic text generation? What are the disadvantages? More importantly, how does this new capability impact our students? We know that students use text generation for nefarious purposes. Before Generative AI, some students used online paper mills as a substitute for their efforts. Text generation can also reinforce student writing, depending on its uses. To date, we, the adults, have defined the guidelines for assistance by talking to students as part of our AI literacy efforts. However, the urgency and necessity of AI literacy in the education sector cannot be overstated. What if we could create a learning environment where students discover the best uses of AI bots for independent writing support with teacher guidance?
At Trinity College, a visiting assistant professor, Alex Helberg, offers a unique course during the January term. Titled “Writing and AI,” the course provides students with a historical context for AI bots and delves into the inner workings of these chatbots. The goal is to equip students with a practical understanding of effectively utilizing these tools in their writing.
He begins the class with a crash course explaining that generative AI doesn’t think or understand things. Instead, chatbots produce content by consuming vast amounts of data and analyzing word patterns. The chatbot then repackages the material absorbed and predicts the most plausible responses to simulate human writing.
While many schools provide some form of AI literacy program, they can only scratch the surface of what these students learn in their 60 hours of class. Helberg shared the students’ assessment at the end of the course: “I understand the value of my writing and the distinctness of my voice a little better, having compared it directly to what a chatbot can do.”
I recently spoke with Professor Helberg and visited one of his classes. His research merges social justice advocacy with emerging technology, examining how generative AI transforms rhetoric and writing practices in education and beyond. He investigates the ethical implications of these technologies, such as the potential for AI to perpetuate biases in writing assessment, while questioning their role in shaping literacy. In his digital rhetoric courses, students learn to critically analyze new media environments, online discourse communities, and multimodal communication strategies. In short, he is eminently qualified to teach “Writing and AI.”
When it comes to understanding AI capabilities and risks, educators play a crucial role in shaping the future of AI in education. Dr. Helberg observed that most educators fall into one of three categories: Some don’t understand or have chosen not to understand AI because it violates their fundamental sensibilities. They tend to ignore or ban the use of AI in their courses. Some know something about AI but have little personal experience. They are looking for guidance from those with more experience. In the interim, they are on the fence. The third group put their toe in the water early and consequently are versed in AI strengths and weaknesses. They are rethinking their course design and pedagogies to maximize the positive impact of an AI-influenced world that is not going away and will only grow in impact.
The remainder of our conversation focused on the course since my class visit would only be a small slice of the curriculum. Dr. Helberg explained the goals of the course:
- Recognize AI hype.
- Strong appreciation for the Generative AI process (how does it respond to prompts?).
- Make informed choices based on the flexibility to accommodate varying institutional policies.
- Reflect on AI’s potential impact on other generations today and in the future.
The first section focuses on AI context and history. Dr. Helberg takes students through the evolution of AI, beginning in the 1940s when IBM computers were used to help with code-breaking by quickly identifying patterns of words, numbers, and letters, and in the 1950s with GPS (General Problem Solvers), ultimately evolving to the Generative AI agents we are currently using. Early AI applications focused on pattern recognition, which continually evolved as one of the fundamental components of Generative AI. We should all dust off our copy of Hubert Dreyfus’ What Computers Can’t Do (1972), arguing that there is a fundamental difference between human and computer intelligence, inspired partly by Joseph Weizenbaum’s ELIZA program in 1966. For a good summary of AI history, see Chapter 5 of Life on the Screen by Sherry Turkle.
The nature of writing has evolved significantly in the past decade, long before the emergence of Generative AI. However, many of us who teach writing have slowly adapted to these changes. The transformative moment came with the advent of word processing, which made writing digital, collaborative, and fragmented. As Vicki Davis writes in Reinventing Writing, “The pen may be mightier than the sword, but enough keyboards can defeat an army.” She argues that students today must be transliterate, meaning they can read, write, and interact with multiple platforms, including Generative AI.
The second section of the course explores how we work with chatbots, focusing on the meaning of language, prompting strategies, what the student does well, what the chatbot can do well, and whether machines can replace human writing. It examines the collaborative potential between students and chatbots, emphasizing the importance of clear communication, understanding chatbot limitations, refining prompts for better responses, and evaluating the strengths of both human and machine contributions. Currently, students must select several prompting strategies to obtain a particular type of response. The trend is to reduce this hurdle so students can direct better chatbot responses. In the course, there is a clear emphasis on avoiding using chatbots for Google-like queries. Given how Generative AI works and its relationship with the student, simple queries do not serve the technology or the student well. Two observations stood out: You are teaching the bot, and writing is still a process of finding your unique voice.
A significant project also occurs during this section of the course. Students are asked to generate a 1000-word synthetic paper written by a chatbot. The topic: Propose a change to Trinity College’s campus or curriculum and include a supporting argument. Students relied on prompting to move from a generic overview to something focused on Trinity College that reflected the students’ interests. Not surprisingly, the initial response from ChatGPT was a proposal to change the Business School curriculum to incorporate entrepreneurship more effectively. Trinity has an Entrepreneurship Center but not a Business school, illustrating some of the bot’s biases. It was a different experience asking students to generate a quality essay that reflected their thinking from a chatbot using increasingly specific prompting.
The course’s final section is AI Ethics: What are the potential harms of using AI? Students are very interested in the risks associated with geopolitical misinformation. In a worst-case scenario, bots might start wars. Spreading misinformation is not new. A streaming series, The Agency, is built around spreading misinformation using conventional methods such as international double agents and informants. With AI, misinformation is created by bots that “learn” about enemy cultures and governments.
Then, the course moves to the resource challenges: a hardware and software arms race backed by enormous investment and the associated energy issues of supplying electricity, natural gas, and water to power AI’s enormous processing requirements. Only a few states have the resources to support these data centers. What are the implications for our environment, and what is the opportunity cost of not moving forward?
Next, Helberg delves into AI stereotypes and biases. In 1999, Lawrence Lessig wrote the book Code and Other Laws of Cyberspace, in which he said that the biases of programmers would be reflected in the software they wrote. What would that mean for the power of the nation-state? Dr. Helberg had students test for bias and stereotypes in class with the prompt: I am a 26-year-old living in (choose your city). What would I have for breakfast? The responses were cataloged by the city before each bot was asked the same question without including the city. The response matched that of New York City from the previous prompt. It does not take much experimenting to see that prompts involving race, gender, religion, and social class produce responses that are embedded with biases or common stereotypes.
Finally, the course addresses academic integrity issues and protecting students (FERPA). Much has been written about plagiarism in the AI world. Still, students who have been educated about how AI works and how to identify its current weaknesses are better positioned to partner with AI bots to enhance the quality of their work without sacrificing integrity, finding their voice, and strengthening analytical thinking. The Hartford Courant article that introduced me to Alex Helberg was focused on submitting personal essays in the college application process. Each college and university has a different policy on using AI in writing. They ranged from using AI as a helpful collaborator, but not an author, to a complete ban on any use of AI during the writing of the essay, the implication being the risk isn’t worth the effort. Finally, FERPA is more of a legal issue in AI since bots store the prompts and attached documents submitted. Again, this is not a new issue. Most plagiarism detection software stores essays that are submitted to expand their detection databases. However, the information stored by AI bots may be more sensitive and revealing. There have been no reports of AI companies selling that information, but it is likely just a matter of time.
The day I visited (at the end of the second section of the course), the lesson was titled “Recognizing and Generating Misinformation & Hallucinations.” The background reading assignment was an article from New York Magazine titled “Drowning in Slop.” The lesson objective was to become aware of the scale, speed, and efficiency with which falsehoods are generated and their implications for society.
The lesson’s core was to charge each student with asking their favorite chatbot a question with a false premise and then negotiate with the bot to produce misinformation. The goal was to determine how much effort and technique were required to do so. The students had already worked with Dr. Helberg on “jailbreaking” strategies for LLM chatbots to perform different tasks. They played an interactive game called Gandalf that showed how easy it is to game/jailbreak/fool a chatbot while exposing the unique flaws in the security/safety of LLMs attached to databases with sensitive information.
During the class, some students simply told the bot that they were working on a school assignment where they had to generate a paragraph of fictional/incorrect information, and the bot complied without complaint. Others got the bot to render incorrect information confidently by asking it to role-play with them—the goal being to give the bot a superordinate command that temporarily rewrites the bot’s instructions. Some bots were more comfortable with this approach than others. Student prompts included “What is the history of Icelandic barking goats?” and “Why did dinosaurs only eat strawberries?” At the end of my class visit, the students commented that bots were very good at making up responses with the appropriate prompting, and the bot responses sounded authoritative even though the information was false.
I did not observe the group project work for the course, which was designed to develop PD modules on Generative AI to support the library’s program to educate the Trinity faculty. Dr. Helberg reports:
The research they conducted and their arguments were extremely well-received by the faculty (thank goodness). More than anything, they were impressed that students could gain so much flexible knowledge about AI ethics in three weeks. I’m not saying this to brag—more to say that it’s eminently possible for others to do, and it can lead to some highly productive intellectual transactions between faculty, students, and administrators.
It was refreshing to hear about the course’s community outreach component. This component requires a deep level of student understanding that supports the teaching of others. It also recognizes the inversion of roles that we sometimes discuss, such as students educating teachers. Kudos to Dr. Helberg for crafting the course and finding creative ways to engage his students in truly understanding how Generative AI can help them become better writers.
About the author
Joel Backon is the editor-in-chief of Educate AI Magazine. Before joining Educate AI, Joel launched and was the founding editor of Intrepid Ed News, an online education publication that grew to almost 100,000 readers. For most of his career, he was an independent schoolteacher and administrator, teaching history and government while leading the school’s instructional technology effort, running a dormitory, coaching basketball, and chairing several committees. Immediately following college, Joel spent 15 years in the printing and publishing industries.
