People reach for all kinds of metaphors to describe their relationship to AI. For some, AI is like a mostly reliable intern. For others, it’s a virtual assistant. Increasingly, chatbots like ChatGPT are moving into the role of companion, therapist, even romantic partner. As a college writing professor, I’ve come to think of AI as a collaborator: an archive of knowledge that talks back. But as a sober alcoholic myself, I also can’t help but imagine it as a high-functioning drunk: It can sometimes sound brilliant even when it has no idea what it’s talking about.

I can tell you stories about the ways AI has come through when I needed it, saving me hours of time by doing mundane tasks, proofreading my writing, or conversing about my latest research obsessions. But then there are those other times when it lies with a cheery tone, when it seems to not understand a word I’m saying but just keeps talking rather than admit it’s wrong or that it doesn’t have an answer. Like a few weeks ago, when I asked ChatGPT to turn my written remarks for an academic conference into a slide deck. My talk was about literary journalism, and it proudly offered me a presentation about luxury travel in Brazil.

Off-the-rails incidents like that give me plenty of cautionary tales to share with my students. But even though I think AI undercuts some of the most important human reasons to write, not all kinds of writing are the same. To write, we often have to research first, and after we’ve written a draft we need critical feedback. Instead of taking a reactionary approach to AI, I want to explore with my students how it can be a useful collaborator in that process.

Chatting with the archive

So much of college writing is based on research and reading, a process that trains the mind to organize information and think logically. But using new technologies for that process doesn’t mean we’re not still doing critical mental work. Just in my lifetime, those technologies have changed radically: We’ve gone from library card catalogs and microfiche to online databases like JSTOR and Google Scholar. Those tools don’t require any less thinking—they just speed up some of the brainstorming and collecting information, and they expand the amount of knowledge we’re able to consider.

Because I had witnessed this rapid digitization of research and writing tools even before AI, I’m more inclined to imagine ways AI can be a collaborative research partner. In my field, for example, literary scholars spend hours combing through primary sources in libraries and archives. Digitization has already made these easier to access, and AI may make them easier to analyze. 

Lately, I’ve realized we could think of talking to an AI chatbot not like browsing an archive, but like conversing with one. Before we dive into more intensive work, we can have a research-orienting chat with a “mind” that at least has a general idea of what’s out there. A few weeks ago, I used my limited access to ChatGPT’s advanced voice function to ask if it thought that this idea of chatting with the archive was a reasonable way to understand what is happening when I converse with AI. It answered, “When you’re talking to an AI like me, you’re accessing a vast amount of information and patterns derived from human knowledge up to a certain point.” It also hedged a bit: “It’s important to remember that while I can provide information and insights based on that knowledge, I don’t possess human experiences or consciousness. So, while it might feel like conversing with a vast reservoir of knowledge, it’s always good to consider the human perspective and context as well.”

Still, as our conversation went on and my questions got more pointed, I could ask it to provide references and places I could go to do further reading. Since that first tentative conversation, this pre-writing conversation with AI is becoming part of my workflow. I’ve always found it easier to work out my ideas through dialogue, but not many people are interested in hearing my half-baked ideas. That is why I’ve found that talking through ideas is one of the best uses of AI for writers.

Creating your own mini-archive

While talking with AI has proven helpful for idea generation — and the fact that it keeps a transcript makes it easier to refer to later — there are a growing number of AI-based tools designed to help with the more intensive phases of research. At the end of the fall semester last year, a student sent me an email asking if I’d heard of Google’s NotebookLM. I hadn’t, but when I opened the link, I got the concept almost immediately. NotebookLM takes the idea of talking to the archive to the next level: The archive you chat with is one you assemble yourself with sources for a particular project, which the AI can also help you collect to get started.

Preparing for my recent conference talk, I dumped 25 PDFs that I had assembled and stored in Zotero, my favorite citation manager, into NotebookLM’s interface. It quickly “read” them and provided a summary that began, “These sources discuss ordinary language philosophy, primarily focusing on the work of Wittgenstein, Austin, and Cavell, and its relationship to other philosophical and literary movements like pragmatism, transcendentalism, and deconstruction.” Below the summary is a text entry field that encourages me to “Start typing…” and provides some suggested prompts like, “How does ordinary language philosophy challenge traditional philosophical approaches to meaning?”

On the right side of the page, in an area designated “Studio,” I’m invited to create an audio overview, which takes the form of a podcast, complete with two voices — one male, one female — bantering about my chosen topic. If I use Interactive Mode, I get treated like a caller on an old late night radio show. I get compliments for my great questions and responses based on the documents I provided. The podcast part isn’t great yet; it’s creepily pandering, but I can envision it getting better and becoming more useful. NotebookLM has other helpful features: It can create a “Mind Map,” study guide, briefing doc, FAQ, and timeline. I’ll continue to use it and suggest students do so as well.

A not-quite writing tutor

One more way that AI could prove helpful to student writers is in its ability to provide instant feedback on student writing. When I asked ChatGPT about this concept, it encouraged me to “think of AI like a writing tutor that’s available 24/7” with the caveat that it “lacks the personal touch and nuanced understanding of individual students that a human tutor provides.” 

I pasted in the full text of one of my previous Mashable stories and asked for suggestions. It seemed well-versed in what we refer to in classroom peer workshop sessions as the “compliment sandwich”: criticism folded in between two compliments. 

It told me, “This is a compelling, eloquently written piece…your voice is authentic and reflective,” before offering “some suggestions to elevate the piece further.” Again, it began with “Strengths to Keep,” followed by “Suggestions for Improvement,” including “tighten the opening,” “strengthen transitions,” and “consider a stronger conclusion.” It also had a few “minor style edits” to suggest. Finally, it provided an overall rating: 9/10.

Maybe it was all the compliments, but I got greedy. I pasted in another essay (9.5/10) and then the conference talk I was working on. The overall impression started off great: “Your paper presents a compelling argument for the value of literary journalism that focuses on the “ordinary” and ‘quotidian.’” That’s true, though I never used the word “quotidian.” But then — I should have expected — it went off the rails. “The references to foundational figures (e.g., Bateson, Becker, Carey, Geertz, Tuchman) and contemporary examples (e.g., Kiese Laymon, Eliza Griswold, E. Tammy Kim) help situate your argument within a well-informed scholarly framework.” I don’t reference any of those figures as foundational or otherwise. 

I called its attention to this, and it said I was “absolutely right” and thanked me for pointing it out. Its explanation, however, was still baffling: “I mistakenly based part of my response on assumptions or cached ideas from other academic discussions of literary journalism, not your specific paper.” I study literary journalism; the names ChatGPT dropped belong to writers, but they are not scholars in my field. Still, after I corrected it, we got back on track and it provided feedback, again utilizing the compliment sandwich. 

I’m not sure what to make of the fact that ChatGPT fared much better against my more journalistic writing as opposed to the academic, except that it provides yet another opportunity for me to urge caution when helping students think through appropriate uses of AI to complement — rather than replace — the writing process.

Ultimately, I love the notion of AI as conversant, albeit something that occasionally overindulges, leading it to overly flatter and outright lie. I’m all for the notion that, in talking with a chatbot, writers can approximate something like talking to a whole host of human knowledge, especially with a tool like NotebookLM that lets writers “teach” the AI about a topic before discussing it.

AI as a collaborator appeals to me, even if I have to approach it with a healthy sense of skepticism, always prepared for the next time it will let me down.