When ChatGPT debuted a year and a half ago, many professors immediately worried that students would use it instead of doing their own written assignments — clicking the chatbot's buttons instead of doing the thinking required to answer an essay.
But two English professors at Carnegie Mellon University initially reacted differently: They saw the new technology as a way to show their students how to improve their writing.
To be clear, these professors (Professors Taku Ishizaki and David Kaufer) were also concerned that generative AI tools could be easily misused by students. And that's still a concern.
But they have ideas for how to set up their own guardrails to create a new kind of educational tool that allows students to incorporate more of their ideas into assignments and spend less time thinking about formatting their writing. did.
“When everyone was afraid that AI would take away students' writing ability, we said, 'If we can suppress AI, it will prevent students from really developing their writing ability. This can reduce the amount of remedial work required. [looking] To see what’s going on with their writing.”
Professors call this approach “constrained generative AI,” and have already built a prototype software tool called myScribe to try out in the classroom, which will be piloted in 10 university courses this semester. ing.
Kaufer and Ishizaki were in a unique position. They had been building tools together to help teach writing for decades. A previous system they built, DocuScope, used algorithms to identify patterns in students' writing and then presented those patterns visually to students.
A key feature of the new tool is called “Notes to Prose,” which interfaces with ChatGPT to turn student-typed bullet points and prose into draft sentences and paragraphs.
“The bottleneck in writing is sentence generation, turning ideas into sentences,” Ishizaki says. “It's a big job. This part is very costly in terms of cognitive load.”
In other words, it can be difficult for a novice writer to simultaneously think of a new idea and remember all the rules for constructing a sentence, just as it can be difficult for a novice driver to grasp both their surroundings on the road and how driving works.
“We thought, 'Can we really use generative AI to take that burden off our shoulders?'” he says.
Kaufer says that novice writers tend to organize fragmented ideas into carefully crafted sentences early in the writing process, only to later abandon them because the ideas don't fit with the final argument or essay. He added that he may end up deleting the text.
“They start really honing too early,” Kaufer said. “So what we're trying to do is with AI. When you're prototyping quality of thinking, you have the tools to quickly prototype languages.”
He says the concept is based on writing research from the 1980s that found experienced writers spend about 80 percent of their early writing time thinking about planning and structuring the whole piece, rather than the sentence itself.
Taming the chatbot
According to the professors, building the “notes to prose” feature took a lot of effort.
In early experiments with ChatGPT, when we asked users to input a few snippets to create a sentence, “we found that the text started adding a lot of new ideas,” Ishizaki says. In other words, the tool tended to make the essays even more complete by adding other information from the vast amount of training data.
“So we prepared a very long series of prompts to make sure there were no new ideas or new concepts,” Ishizaki added.
This approach differs from other efforts focused on using AI in education in that the only source the myScribe bot draws on is student notes, rather than a broader dataset.
Stacey Rohrbach, an associate professor and graduate dean at Carnegie Mellon University's School of Design, sees potential in tools like the one her colleagues created.
“We have long encouraged our students to always create a solid outline and say, ‘What am I trying to say in each sentence? ” approach will help in that effort.
And she says she's already seen student writers abusing ChatGPT, so she thinks some restraint is needed.
“This is the first year that I've seen a lot of AI-generated text,” she says, “and ideas are lost. The sentences are well-constructed, but they become gibberish.”
John Warner, an author and education consultant who is working on a book on AI and writing, said he wondered whether the myScribe tool would be able to completely prevent “hallucinations” caused by AI chatbots, or cases where the tool inserts false information.
“The people I talk to think that's probably not possible,” he says. “Hallucinations are a feature of how large language models work. Large language models are not judgemental. They may not be able to escape from making something up, because what does it know?”
Kaufer says the tests have gone well so far. In a follow-up interview over email, he wrote: “It's important to note that 'annotating prose' works within the confines of a paragraph unit, meaning that if it crosses the boundaries of the annotation (or 'hallucins' as you say), it's immediately obvious and easy to identify. The concern that the AI will hallucinate would be magnified if we were talking about larger units of discourse.”
However, Ishizaki acknowledged that his company's tools may not be able to completely eliminate AI illusions. “But we want to constrain or guide the AI enough to minimize 'hallucinations' and inaccurate or unintended information, and to allow writers to correct them during the review/revise process.” Masu.”
He explained that his company's tools are not just one-off systems, but a “vision” of how technology will evolve. “We are setting our sights on the direction in which lighting technology should advance,” he says. “In other words, the concept of notes to prose is essential to our vision for the future of writing.”
But even as a vision, Warner says he has different dreams for his writing future.
He says a technology writer recently pointed out that ChatGPT is like having 1,000 interns.
“On the one hand, it's 'wow,'” Warner says. “On the other hand, 1,000 interns will make a lot of mistakes. An intern will spend more time than it saves early on, but as time goes on, the person will have less supervision.” ” But with AI, “monitoring doesn't necessarily improve the underlying product,” he says.
Doing so, he argues, turns AI chatbots into “very powerful tools that require a huge amount of human oversight.”
And he argues that converting notes into text is a crucial process of actual human writing and should be preserved.
“A lot of these tools try to make processes efficient that don't need to be efficient,” he says. “A big thing happens when you go from notes to draft. This isn't just a translation. These are my ideas, and I want to get them on the page. Rather, these are my ideas, and my ideas are taking shape as I'm writing.”
Kaufer is sympathetic to that argument. “The bottom line is that AI is here to stay and it's not going away,” he says. “There will be fights over how it's used, and we're fighting for responsible use.”
