This article was reviewed by Caltech a faculty member.
For years, Susanne Hall has watched computer scientists develop AI tools that write, or at least automatically generate, text.
As a teaching professor of writing at Caltech and director of the Institute's Hixon Writing Center, she teaches undergraduate writing courses and helps support and improve Caltech's writing curriculum. Her interest in publication ethics led her to join the first major study of text recycling, the reuse of one's own writing, as a co-principal investigator. She shared her thoughts about the changes to come in teaching and writing as text-generating AI evolves.
How did generative AI tools come to your attention?
I've been asking Caltech students to write about artificial intelligence for a decade. That is how I first learned that large language models existed and computer scientists were exploring how to use them to train algorithms to generate text. As soon as I knew that, I was fascinated.
What surprises you about the tools we have now?
I am surprised how quickly the people building these tools are making technological decisions that have economic and social implications, seemingly in response to news coverage or other kinds of limited feedback. The tools are being changed very rapidly in response to their uses in the real world, and my impression is that this is more improvisational than it would ideally be.
How might ChatGPT and other AI tools change how educators teach writing?
As faculty, we have to remind ourselves of our learning goals for students. As the tools develop, can they promote that learning? Might they impede it?
One of the richest ways to figure out the uses and limits of the tools is to work with them in a controlled context—in-class activities or short low-stakes assignments. For example, let's say the 15 people in my class ask a tool to summarize an essay I assigned. What did it say? When the students read the essay, what did they think of those summaries? At this point, our consensus is that the summaries have obvious limits and are a poor substitute for those written by humans who have carefully read the essay. It's helpful for students to reach these insights collaboratively.
Are there other ways generative AI tools could benefit writing instruction?
I'm curious about how the tools could increase accessibility. Some students with ADHD or who are on the autism spectrum are reporting that the tools can help with blocks they face as writers. Multilingual writers are finding that the tools could help them avoid sentence-level errors that trigger biased reactions to their research writing.
Writing can feel overwhelming for students. Writing teachers and tutors help student writers understand that many challenges they experience are normal and even necessary. It may be that dialogue with a chatbot—especially in the middle of the night, when some students are still working—can offer some positive feedback to keep writers motivated and confident in their work.
How may these AI tools alter students' preparation for college?
Access is a concern—students will have different levels of exposure to these tools.
Then, too, some K–12 administrators may prohibit these tools broadly, wanting to make sure students do the important work of learning. But that might result in students only being allowed to complete writing assignments during class. Already, a lot of students' writing is timed because instruction is cued toward standardized tests. Most texts we write from college forward are not produced that way; timed writing does not cultivate the most relevant skill set.
How can faculty prevent the use of AI tools to cheat? What about plagiarism?
We can make a direct, persuasive case to students for why we ask them to undertake an assignment. We have to say, here is why it's worth your time in 2023 to read a 500-page novel or a particularly difficult essay or technical research paper. Once they see a good reason to do something, most students will do the work.
As a community, what incentives and barriers do we create for students?
In my experience, students cheat because they are desperate. It's the middle of the night; they haven't left enough time, and they make bad decisions. Students create these situations, but faculty sometimes make them more likely by assigning overwhelming amounts of work. Any one professor might notice that and make changes, but we need a broader conversation.
What are your thoughts about plagiarism and intellectual property?
These tools raise serious questions about intellectual property. Most have been trained on a wide array of texts or images that are under copyright. I think we should seriously consider what is owed to the writers and artists who created the works on which these tools are trained.
As for plagiarism, students now have easy access to sentences, paragraphs, and outlines that were not directly written by another person. Norms for that have not yet been established. Chat tools challenge the long tradition that the language in any academic text was originated by one of its listed authors.
What other drawbacks should writers consider when using generative AI?
Chat tools are wrong a lot, and they fabricate information, and yet they seem like reliable interlocutors. Because these tools are configured in a conversational context, it is challenging for humans to remember that the tools are not sentient and don't know what they are saying.
In teaching, I help students assess the credibility of texts they encounter. I can teach students how to see that a website is written for propaganda purposes. But with this technology, the source of the text is hidden in the black box of the algorithms producing it. Not even the developers of the tools know exactly how the tools generate their outputs. It's harder to develop critical media literacy in this context.
We must also be actively looking for bias in the text and images these tools produce. Large language models trained on datasets that contain bias will reproduce that bias in their outputs unless developers actively guide the tools not to do so.
Are there pitfalls specifically for faculty and teachers?
There are companies selling products that promise to respond to student work on our behalf with automated scoring and feedback. These tools aren't reading; they are doing something else. Given the unmanageable amounts of work assigned to many K-12 teachers and college instructors, we could see students using AI tools to write and professors using them to respond—it becomes tape recorders talking to each other (a classic montage in the 1985 film Real Genius).
What is the best way to use tools like ChatGPT in writing research papers?
For one, machine learning could help us write scientific research papers' methods sections more clearly and consistently. We need consistency because in science, the hope is that if I describe clearly to you how I did something and then you do the same thing, you'll get the same results. Better written methods sections could help us assess the reproducibility of our results.
Do AI ethics complicate these new abilities?
I worry about how much concern is going into thinking through the impacts that exist beyond corporate shareholders. The people who will experience the most climate impacts from these [energy-intensive] technologies have more limited access to these tools. And all residents of democratic countries are going to be negatively affected if the internet becomes overrun with AI-generated misinformation, such that finding credible information about key topics becomes more and more difficult.
At Caltech, we are educating the next generation of people who will build these tools. I hope we lead them to care about ethical considerations.
The tools have so much positive creative and social possibility. They can lead in a utopian or a dystopian direction. Both avenues are quite open right now.
Will you change the way you teach in response to new AI tools?
Teaching critical reading strategies will continue to be important. Generative AI tools may become good at summarizing text, which is useful in some contexts. In others, reading critically is important. It cannot be outsourced to a tool that doesn't think and isn't actually reading.
Teaching rhetoric will be even more important than before. These tools can't fully understand the specifics of a rhetorical situation. We are in command of that. That is what good writers and editors do: finesse texts so that they can accomplish the goals they have with the audience they're meant for.
Reader question: [How do] experts such as yourselves view the authenticity of text that has been generated by ChatGPT? Is ChatGPT the author of the text it generates or is the author the person inputting the query? Or are they co-authors?
My sense is that faculty in higher ed are still developing ways of thinking about authorship and generative AI. What seems important to me is that we should not consider ChatGPT and similar tools "authors," because the idea of authorship has long been associated with intentionality and responsibility. While ChatGPT can produce cogent prose, it is doing so algorithmically and lacks any unique intention for its outputs. Are the programmers the authors? Programmers who design the tools may train them toward certain types of outputs and away from others, but at the level of an individual output, we cannot say that the programmers are collective authors of a particular text they never see or endorse.
There seems to be a consensus emerging that human authors are to be held responsible for the prose they present to others, and if they include outputs from ChatGPT that are incorrect or offensive in their writing, it is the human author and not the text generating technology that is responsible. Many classroom policies I'm seeing indicate to students that they are responsible for the accuracy and content of the writing they submit. This would suggest that the authorship lies with the human.
Conversely, many people still seem to see the act of taking a block of writing from ChatGPT and putting it into an academic paper as a kind of plagiarism, which would suggest that the tool is an author. I'd suggest it's better to consider it source misuse (in contexts where there is a policy or agreement that the tool shouldn't be used in that way). ChatGPT outputs represent a new kind of source that has never existed before.