![Learning [with] AI](https://substackcdn.com/image/fetch/$s_!LVvT!,w_36,h_36,c_fill,f_webp,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcbb26443-cd3e-4aa2-829b-ce038d59d370_756x756.png)
Image Generators in the Classroom: Beyond the Magic
How can we make AI processes visible so students understand how AI works and build agency around these new tools? AI image generators might be the answer!
Here's a little story.
One day in September 2024 (the date matters), I came home from work and found a doodle on the kitchen table.
On a whim, I took a picture of it, put it in chatGPT and asked the bot to replicate it.
As always, ChatGPT politely answered my prompt and did what I asked it to do.
Or… did it?
This result was surprising to me. While the prompt seemed simple — after all, it was a simple line doodle, which seemed relatively easy for a bot that could come up with hyper realistic pictures or a mock studio Ghibli drawing in a few seconds — it proved to be challenging for chatGPT. With a little bit of care and tracing paper, a human could have probably traced the doodle in a minute or two. What was easy for a human became mission impossible for smarty-pants ChatGPT.
What it illuminated to me was twofold.
First, the fact that AI had not been trained to reproduce doodles. When you think of it, it makes sense, as there are probably not too many doodles on the Internet. It’s a good reminder that AI models are trained on specific datasets — and these are often whatever happens to exist on the Internet.
In essence, it shows AI’s limitations: even though its datasets are vast, they definitely don’t represent all human knowledge or creativity. It showed one of AI’s blind spots.
Which led me to a second realization: that AI is by essence an approximation. That’s why the AI-generated version of the doodle looked so different: smoother, curvier, more stylish. AI uses computational algorithms and statistical models to generate the most probable response to a given prompt. The statistical model it’s using has been trained on tons of examples. That “most probable response” is made visible really well when you look at the two images side by side. The differences between the harsh and irregular lines of the doodle and the smoothness and polish of the lines of the AI-generated image are striking.
In my experience, even now, most people (kids and grown-ups alike) don’t consider that AI relies on statistical models to generate likely outputs. It is concerning to me, even shocking, because in my opinion, this is the FIRST thing we should teach people — and students — about AI.
AI is predictive by nature and simply generates the most probable answer to a human request.
Why is this important?
In the educational world, it is important to demystify AI: the “magic” people see behind it, and how it is sold to us. We all know magic is not real — it is just tricks. Yet when we go to a magic show, we want to be wowed. We don’t want to know the magician’s secrets. Why? Because knowing the tricks kills the magic.
I think we need to kill the magic of AI.
In AI-world, like in magic shows, people don’t want to know the mechanics under the surface.
But in the educational world, we want teachers and students to understand how the magic show works and what the behind-the-scenes look like.
When students understand how AI works, they understand when it’s helpful to use AI and when it’s not. Intentionally choosing AI (or not) develops learner agency: the ability to understand when to seek help, support or inspiration from AI, and when not to. For example, AI may be most helpful when you’re okay with an approximate answer. It may not be the best when you need a lot of precision, or when the answer really matters. And sometimes, it is best to skip AI entirely — maybe the answer is too exact, the prompt too fiddly, or it takes more time or energy to have AI do something for you instead of doing it yourself.
I am aware that models and tools are getting better, and that the margin of error is shrinking. The downside is that it is becoming less and less obvious to people that AI is “just” a predictive machine.
I ran the same request a few days ago (about a year after the first request) and here is the doodle that chatGPT came up with:
When you look at the result side-by-side, the similarities are striking and the differences are getting minimal. One day in the near future, AI tools for the general public will be so good that the gap between ”probable” and “accurate” will be non-existent.
So now what? Can we still teach the predictive nature of AI, even though the tools are getting better?
Absolutely. .
We still need students to learn about AI while they learn with AI!
Learning with AI includes learning about AI and how it works. We want students to understand how the magic show of AI works, even if this means they will be disappointed. Because this is how they will develop agency toward AI so one day, they can run their own show, whether it be magic or not.
Resources for Teachers
We can teach the predictive nature of AI to students using slow looking experiences of AI-generated images. Slow looking is at the core of many different Project Zero projects such as the Visible Thinking project, Agency by Design or the Cultures of Thinking project, among many more. The book Slow Looking: The Art and Practice of Learning Through Observation by Project Zero principal investigator Shari Tishman offers a compelling argument for the power of slow looking in learning environments to inform teaching and learning. Lots of routines and frameworks to choose from!
Here are some routines to look closely at an AI-generated picture:
You can also use the following to look at two pictures side by side:
The Maker Moves (“Looking closely” and “Exploring complexity”)
You can look side by side at a human-generated drawing, doodle or painting and its “replication” using an AI tool, like we did in this post.
You can also look at famous works of art and their “copycat” using AI tools.
Here is what I did with the Mona Lisa, arguably the most famous painting of all times.
My prompt was the following:
Here is the image chatGPT came up with:
The side-by-side comparison of the real Mona Lisa and its AI counterpart is an excellent way to look closely at AI and explore its complexity using slow-looking routines and strategies.
Here is a slidedeck, “Looking Closely at AI” I made with learning experiences for students. Feel free to copy, modify and use as you wish, and message me on Substack to let me know what you did and how it went with your students, or if you would like me to add additional resources to it!