Help! AI stole my students’ brains!

If you’re like me, you’ve been reading with disquiet a growing number of studies that show how AI is already diminishing cognitive and creative abilities, in a process now called cognitive offloading. There are already so many, published in journals including Societies, Springer Nature, Science Direct, the Journal of Computer Information Systems and more. 

I can see the value of “keeping the human in humanities”, as Dr. Timothy Pettipiece so eloquently put it, when he argued that technology-free classrooms could help students acquire the very skills that would be necessary to use AI effectively. But, before we do that, we need to let students understand the limits of AI. Because right now, undergraduate students trust AI far too much. 

It’s difficult to overstate how much undergraduate students, fresh from COVID lockdowns and online learning in the middle school years and freely accessible AI in their high school years, have come to rely on this technology. While hardly definitive, a straw poll in my first-year ancient global history course found that every person in the class had used AI to complete at least one assignment in the last year. Every single person in a 45-person class. A recent KMPG survey discovered that in 2025, 73 per cent of Canadian high school students self-reported using generative AI to help with schoolwork.  

Did the students know, I wondered, that free versions of AI chatbots like Copilot and ChatGPT still regularly hallucinate references? Did they understand that much of the “literature” produced by AI chatbots is formulaic and shallow, lacking examples and context? That chatbots were designed to be eager to please, to encourage greater engagement and were not necessarily reliable in tasks as diverse as parsing Latin — Claude agreed with every different way I translated a Latin sentence — to analyzing the trends in Hammurabi’s Law Code—Copilot absolutely refused to write a story based on Hammurabi’s Law Code that followed the principle of lex talionis (reciprocal justice, or the “eye-for-an-eye” principle). Did they really grasp that AI could be used to spread Russian propaganda in the place of reliable information?  

To my horror, the answer was a resounding no. Here’s how I know.  

To help teach students AI literacy, my first-year premodern global history class had four different assignments designed around asking an AI tool to produce material, then analyzing the output. To do this we used Copilot, since our university had a Microsoft Enterprise licence that provided the students with basic data protection. The assignments ranged from having the students edit AI-produced essays, to analyzing Copilot’s ability to summarize articles, to tweaking prompts in Copilot until it created an accurate map. It was this last assignment that brought home how few students really questioned the accuracy of anything Copilot produced. I saw maps students accepted as accurate in which Africa was labeled as Australia; in which India was in Europe; in which landmasses were labeled as oceans and oceans as continents. It is hard for me to express just how inaccurate the maps produced by Copilot were. They were disasters. 

And yet, very few students, only about 25 per cent of those who attempted that assignment, realized that the maps produced were problematic. It wasn’t that the other 75 per cent did not know basic geography; it was that they didn’t even glance at what the AI produced. The absolute trust students put in AI was shocking. Even with an assignment that had warned that AI could generate error-ridden maps, most students just assumed that what the AI generated was accurate. 

It is this trust that should worry us educators the most. If students already so trust AI output that they can’t see errors like the continent of Africa being mislabeled, how can we expect them to see the subtle errors that AI introduces? Or disinformation that might be included? Learning how to scrutinize AI responses analytically may be the most important tool we teach for the foreseeable future. 

To reinforce the unreliability of AI map production, I added a question on the final exam that asked students to find six errors (both geographical and with respect to ancient trade routes) in a map Copilot produced in response to the prompt: “Please draw common trade routes across Afro-Eurasia circa 500 BCE.” The map is reproduced above; there are far more than six errors. This question was nearly uniformly answered well by the students. What this suggests is that the students could recognize bad geography when they were asked to notice it. But they had to be prompted to notice it. We badly need to instill the skill of always, automatically questioning AI production. We need to inculcate an attitude of constant skepticism and analysis.  

Right now, my students don’t believe that they are smarter than any AI. We need to prove to them that they are. And the only way to do this is to teach AI literacy at high schools and in university classrooms. And yes, even in humanities classrooms. The skills we already teach as humanists — analytical reading, constructions of knowledge and critical thinking — can pave the way for using this new tool responsibly and effectively.