AI Literacy Lesson: What to Do When AI Gets It Wrong

It's well known that AI can hallucinate and sometimes just get things wrong. (Side note, have you met humans?) The question for educators is how do we teach that important AI literacy lesson to students?  

You can never be sure when an AI will fail in that way, so a predictable lesson that worked once may never work the same way again. It has to fall to the teachable moments. But beyond, "Hey, look, the AI failed us!" what do we offer our students as the next step? What should they do when the AI fails? Here is an example of the teachable moment that happened last week in my classroom. 

My seniors are reading Into the Wild. We are nearly done with the book, but my colleague just used her own powers of AI creativity to generate a fun activity where students have to decide what they would put in their own backpacks if they were headed into the wilderness. I liked it. My students liked it. It was a great way to connect with the book on a Friday afternoon. (Pack Like Chris Activity)

The catch was that after choosing their own supplies from the list, and keeping the weight under 35 pounds, my students needed to compare what they packed to what Chris McCandless packed. And, while the book mentions things he had with him many times, those moments are often spread across the book. Hunting down all the things he had when he started down the path would be a somewhat frustrating exercise. So I turned to Gemini. 

In short order, Gemini produced a comprehensive categorized list of all the things Chris had with him, from food, to gear, clothing, and books. We read through the list, agreed with most of it from our reading, and even noticed a few things we might have missed. 

That's when a student saw the first error. The book list contained War and Peace. As my student noted, Chris had gifted his copy of War and Peace to Wayne Westerberg. We looked at the book list more closely. It was also missing a book by Gogol that we knew Chris had with him because he wrote a note on a page from it. 

We returned to Gemini. "Tell it that it forgot the Gogol, and to take off War and Peace." They were so pleased with themselves at having found these errors. 

"No," I said. And then I began to type. "Check the list of books, some of those are mentioned in the book, but he did not take all of them to Alaska."

The next list of books was accurate, including removing some London books that had first been included. That was an error we had not noticed. 

Just for fun we decided to push back and argue about the London books. We were pretty sure Gemini was right about that one, but we wanted to be certain. Gemini responded with three sources to verify that Jack London books were not among the possessions recovered at the site. This was another lesson. When in doubt, ask the AI to defend its answer. That's still not a guarantee of accuracy, but it will often offer outside sources you can check yourself. 

The Lesson

Once we were happy with the book list, I asked my class why I chose not to tell Gemini the specifics of what was wrong with its first list of books. 

They paused, and then one of them said, "You knew if you told it exactly what was wrong then it might just tell you what it thought you wanted to see." Exactly. This is often called sycophancy, and it helps to name that for students. 

By withholding the specific information about the errors, I had a way to check the next output while also being confident that the AI wouldn't just feed me back information I gave it. 

My students learned the importance of being able to fact-check an AI, but also what steps to take when you know it is wrong and how to help it get to a more correct answer without compromising that answer. 

The Checklist

• Check the output against what you know (in this case the book).
• Tell the AI to recheck, without saying what’s wrong.
• Compare the new output to the errors you held back.
• Push back and ask for sources to see if it holds its ground.

How to deal with this inaccurate information was a small but important thing to get to teach them. Maybe something to keep in mind the next time you're using an AI tool with students and it gets the facts wrong.