One of the things I appreciate about designing and facilitating workshops for faculty is that doing so gives me the opportunity to take a few ideas and examples that have been bouncing around my head and turn them into a new argument of some kind. Having led eight workshops for various audiences in the last ten days, I've had no lack of opportunities to do that synthesis work! What follows is one of the new arguments I made during my workshops recently.

Back in May, I wrote that it's time to move past the notion of AI as an answer machine. I credited podcast guest Leon Furze with the observation that the user interface of generative AI chatbots like ChatGPT and Copilot position the tools as perhaps all-knowing answer machines. "Explain quantum computing in simple terms. Got any creative ideas for a 10-year-old's birthday party?" read two of the sample prompts ChatGPT provided in its early days. These tools are, of course, unreliable answer machines, given how they operate, stringing words together in sensible but probabilistic ways. But they provide their answers with such great (simulated) confidence that it's tempting to see an AI chatbot as some kind of oracle.

This notion of AI-as-answer-machine is especially problematic in the learning context because when a student doesn't understand a concept, it's not always the case that a full and complete answer to a question is the most helpful teaching move. If a student is stuck on something, an ethical human tutor is more likely to ask questions to see how the student currently understands the topic at hand, then ask more questions to guide the student to a deeper understanding. And they certainly wouldn't do the problem set or write the essay for a student asking for help on an assignment! ChatGPT will do so and enthusiastically, if asked.

That latter use of AI is what I like to call the forklift-in-the-gym. If you use a forklift in the gym, you can move the weights around, but you won't gain any strength in the process. Similarly, if a student uses AI to do their problem set or write their essay, they'll get a product they can turn in for their assignment, but they won't develop any of the skills the assignment was designed to foster. And there's a growing body of literature that students' default uses of AI in their learning tends to be the along the lines of forklift-in-the-gym or AI-as-answer-machine. To be clear, these uses of AI are a kind of cognitive offloading that's not great for learning.

In my workshops, I'll pair the forklift-in-the-gym analogy with another one featuring gym equipment: the treadmill. I'm a runner (occasionally these days) and when it's 95 degrees and 100 percent humidity here in Nashville, I don't go running in my neighborhood. I go the gym and hit the treadmill. It's a technology that doesn't do the running for me, but instead creates conditions under which I can engage in the kind of physical practice I'm after. AI can certainly function like a forklift for our students, but it can also function like a treadmill, providing new learning experiences for them.

"The good news is," as Lodge and Loble write in their literature review, "that research studies also suggest these harmful effects [of cognitive offloading] can be counteracted through purposeful teaching and learning strategies and effective design of AI education technology." I'll save my thoughts on the design of AI-powered educational technologies for another day. What I want to argue here is that if our students are using AI as an answer machine or a forklift in a gym, we can and should provide them with better mental models of working with--and learning with!--AI, turning AI into something more like a treadmill.

During my workshops last week, I tried out a new term to describe a particular way of working with generative AI that I find myself using often: AI as a possibility machine. And I argued that we can introduce this mental model to our students through thoughtful assignments.

What do I mean by AI as a possibility machine? Here's an example: For one of my workshops, I wanted an image for my opening slide that captured the main idea of the workshop, that we can use AI syllabus statements as invitations for ongoing conversations with students about AI and learning. I gave ChatGPT my workshop title and description and asked it to come up with multiple prompts I could use with Midjourney (an AI image generator) that might visualize metaphorically that main idea. ChatGPT came back with eight metaphors (the open gate, the conversation table, the trail map, the jazz ensemble, and so on) along with Midjourney prompts for each. As a huge fan of national parks, I liked the trail map angle, so I asked Chat for three different prompts using that metaphor. I took one of those prompts and copy-pasted it into Midjourney, which generated four images that were all in the ballpark of what I wanted. I picked one I liked and asked for more variations, eventually settling on the image below, which I used in the title slide for my workshop on AI syllabus statements. The AI tools generated a range of possibilities at every step of the process, and I selected the option that best aligned with my goals for the task.

I'm not the only one that uses AI in this way, of course. You might use AI in similar ways.

Here's a second example: In my summer "study hall" episode of Intentional Teaching, the panelists and I discussed a 2024 study by Samangi Wadinambiarachchi and colleagues titled "The Effects of Generative AI on Design Fixation and Divergent Thinking." In their study, they asked a number of student volunteers to design the avatar for a new helpful AI assistant. They wanted to see how varied the student designs were when students had different sources of help in the design process. Some students had nothing but pen and paper, while others had access to Google image search, while a third group could use Midjourney. The researchers quantified the extent to which a student's designs resembled each other and the extent to which they resembled the sample design given to all students--a fairly classic cute robot design. Consistent with the Lodge and Loble findings, most students who used AI as part of their design process showed a kind of design fixation, typically producing avatar designs that looked basically like cute robots.

However, one of the students took a different approach to using AI. They started with several prompts not related to the sample avatar (e.g. "dragon" and "mother love"), spent a little time playing with the notion of intelligence (and drawing an Einstein-inspired avatar), then detouring into more abstract concepts (e.g. "singularity" and "consciousness") resulting in a non-objective avatar design that looks nothing like the classic cute robot. "The participant has considered a range of alternatives," the researchers wrote, "and has produced a seemingly useful design that bears no resemblance to the given example. This is perhaps indicative of what we could consider to be effective AI-supported ideation." I might say that the student used AI as a possibility machine, and I have a hunch that the student had some experience brainstorming in the design context. Expert use of AI requires expertise, and I can imagine someone who is already good at this kind of visual brainstorming using AI to supplement their creativity like this.

In the Wadinambiarachchi study, one of the students happened to approach the task using a more useful mental model of AI, but the Lodge and Loble findings would suggest that teachers can guide students toward similarly useful mental models through intentional assignment design. During my workshop, I shared an example of an assignment that directs students to use AI as a possibility machine from my University of Virginia colleague Andrew Simon. Here's what I wrote about Andrew's assignment on my blog last year when I was reporting some experiments in AI from a fall 2025 faculty learning community:

If you're interested, you can read Andrew's full assignment description here.

I think by comparing three sets of decisions about how to condense a long academic paper into a short policy brief--the authors' decisions, the student's own decisions, and the AI chatbot's decisions--Andrew provided his students with a useful mental model of working with AI, that is, treating AI as a way to generate possibilities that you then critically evaluate against the task at hand.

I hope all this gets you thinking about the ways that you interact with generative AI and how you direct your students to use AI in support of their learning. When you include some AI component in student work, what mental model of AI does it assume? And what kind of assignment or activity could you give your students that frames AI as a possibility machine? If you have an example of such a thing from you're teaching, I would love to hear about it! And if you have a better term than "possibility machine" for this mental model, please let me know. I find that term a little fantastic, but "option generator" isn't as catchy.