It’s time to move past the notion of AI as an answer machine.

Earlier this week I had the opportunity to interview Leon Furze for my podcast. You may know Leon from his work on the AI Assessment Scale or his series of articles on teaching AI ethics. He’s a thought leader in the ways that generative AI is affecting education, and I’m excited to share my conversation with him on the podcast in a few weeks.

During our conversation, Leon said something that helped me connect a lot of dots in my thinking about AI-aware teaching. I’ll quote a blog post of his from March, where he first made this point:

“Most educators’ first interaction with GenAI was ChatGPT, sometime in late 2022 or early 2023. They opened it up, saw a clean interface that looked a lot like Google, typed in a question or asked it to write something, and got a response.

From that moment, their mental model of what AI is was set.”

Leon is arguing here that the user interface that OpenAI decided to use for ChatGPT, perhaps intended to make the chatbot familiar and easy to use, created the idea in users’ minds that ChatGPT was some kind of answer machine–or even oracle–since that’s how many of us think about Google’s search tool. The mental model of AI as answer machine is a problematic one, as I wrote about in June 2023, since AI chatbots like ChatGPT famously get things wrong and make up information. That’s a little less true with today’s frontier AI tools, but it was definitely true back in 2023.

Leon argues that higher education needs to unlearn the answer-machine mental model, and I agree.

In my 2023 post and, really, ever since then, I’ve argued that we shouldn’t think of AI chatbots as answer machines but instead as wordsmiths. These tools probabilistically string together words in ways that are sensible to humans, and when prompted well, the wordsmithing that AI does can be interesting, useful, or even provocative. Consider the ways that prompting AI to act as a editor can help us refine our writing, as Heidi Nobles detailed on my podcast, or Mike Caulfield’s strategies for leveraging AI chatbots as assistant factcheckers. These approaches don’t look to AI chatbots as oracles, but as tools we can use to deepen our own writing and research processes.

In my talks and workshops, I’ll often share the image below that I created using Midjourney with the prompt “imagine a robot weaving together words like a giant tapestry.” When we use AI image generators, we don’t usually ask, “Did the AI get the image right?” That question doesn’t even make sense for a prompt like this. No, we ask, “Is this an interesting or useful or provocative image?” The user interface of Midjourney, like the fact that it generates four image options to consider, directs us to think of the tool in this more playful way. I argue that if we can take that same mental model and apply it to text-focused AI tools, then we’re in a better position to use those tools productively.

I was thinking about all this later in the week while reading the report “Artificial Intelligence, Cognitive Offloading and Implications for Education” by Jason Lodge and Leslie Loble, also published in March of this year. It’s one of the studies we’ll be discussing on an upcoming “Study Hall” episode of the podcast, and it may be my new favorite piece of scholarship in this area. The authors review the research on AI and learning with a focus on cognitive offloading, and they identify a number of patterns that are emerging in that literature.

Lodge and Loble point to studies showing that students’ default use of technologies like ChatGPT in a learning context treats the chatbots like answer machines. Students too often outsource their thinking to the chatbot in an attempt to make learning easier and more efficient. This can lead to short-term performance gains for students, but at the cost of long-term learning. And not just learning in a particular discipline or domain: The cognitive offloading can result in a kind of “metacognitive laziness” that degrades students’ ability to learn in the future.

That’s pretty grim! The good news is that Lodge and Loble also point to studies showing that when students are directed to use generative AI in more productive ways (and not as an “answer oracle”), the AI use can actually enhance things like critical thinking and problem solving. They offer three mental models for these better approaches:

• AI as cognitive mirror, in which the AI chatbot plays the role of a novice learner and the students learn by teaching the chatbot
• AI as Socrative partner, in which the AI chatbot is prompted to act as what Anna Mills calls an ethical tutor–not answering student questions, but asking them questions to promote their learning
• AI as verification partner, in which the student “continuously evaluates and corrects the AI output,” practicing their growing expertise as they do

All three of these approaches are versions of strategies for practicing AI-aware teaching that I recommend all the time: using AI to create conversation simulations for students, designing prompts for custom AI tools to function as tutor bots, and giving students “critique the chatbot” assignments. Thanks to the report from Lodge and Loble, I now have some more citations for these strategies to go along with the examples from teaching practice that I share!

Connecting these dots got me wondering why I’ve been advocating for these alternative mental models for generative AI since the start of my work in this area. I think it’s because I talked with Robert Cummings for my podcast about generative AI in early November 2022, publishing that conversation just days before the release of ChatGPT. Seriously, the word “ChatGPT” doesn’t show up in the transcript! Instead of encountering AI through ChatGPT’s Google-like interface, I learned from Bob about the album Chain Tripping by the band YACHT. He cited lead singer Claire Evans on the band’s process for creating that album, which involved feeding their existing music into an AI tool and then asking it to generate more music in the style of YACHT. They used the outputs of the AI to create the new album, remixing and adding to the AI-generated lyrics and music as they liked.

Hearing that example form Bob before I was introduced to ChatGPT helped me see AI differently, as tools for creative expression and not as answer machines. I’ve been trying to give others that experience through my presentations as early as 2023. Back then, I opened my talks with a series of examples of AI use that I labeled “Everything is awesome!”, complete with a screenshot from The LEGO Movie which features a catchy song with that title. The YACHT example was often in that set of examples. This was followed by another set of examples under the title “Everything is fine.” with that meme of the cartoon dog drinking coffee while his house burns. In both the positive and negative examples of AI, I tried to show mental models for working with AI that went beyond the answer machine idea. That needed to happen first in these sessions, before we talked about students and AI use.

Today, I don’t run with the “Everything is awesome!” / “Everything is fine.” examples in my talks since we’re all more familiar with AI. But I do typically share the rubber duck effect, something I learned about and blogged about in March 2025. This is an idea from the world of computer programming. Imagine you’re a software engineer, trying to write some code to do some things, and it’s not working out. You might pick up an inanimate object from your desk, like a rubber duck, and start explaining your problem out loud to the object. The duck won’t do anything more than quack, but the act of articulating your problem can sometimes lead you to a solution.

That’s the rubber duck effect, and I think it’s a very useful mental model for working with AI. When you converse with your favorite generative AI chatbot, the AI will do a lot more than quack, but you can use the back-and-forth conversation as a way to sharpen and deepen your own thinking. It’s what’s happening in your head that matters; the AI output is just a provocation. You might share an essay draft with AI and ask it to imagine how different audiences would react. You then get to decide how you want to revise your essay based on those potential reactions. You might ask AI to come up with a lesson plan for your class next week, then decide if its ideas for your class align with your values and priorities–giving you a chance to reflect on those values and priorities. These are ways you can lean into the rubber duck effect.

Lodge and Loble argue (from evidence!) that educators should be teaching students these more cognitively useful ways of interacting with AI. So does Mike Caulfield in a blog post just yesterday, in which he argues that the “answer box” model for working with AI is limiting. He then writes:

“Yelling over and over again ‘It’s a bad answer box!’ simply accepts the corporate frame that it is meant to be an answer box, and our argument is over how good at that function it is. Accepting that frame is a horrible disservice to students, who need to conceptualize knowledge discovery and the exploration of ideas in more grown-up ways. It used to be pretty uncontroversial that our job as teachers was not simply to scold bad practice, but model good practice. To do that you do have to engage with these tools and learn how to use them so that you can model that for students.”

Circling back to Leon Furze’s comment during our interview this week, I’ll end by saying that many students and, indeed, many educators still see generative AI as that Google-like answer machine, leading to the detrimental cognitive offloading we’re all worried about. Our job as educators is to move past that mental model to understand other ways of thinking about and working with AI that can actually support learning, and then helping our students do the same.