How to Leverage the Power of AI in Differentiation

In the sci-fi story “The Fun They Had,” Isaac Asimov paints a dreary picture of what school will look like in the year 2155. Children “go to school” by moving to a room in the house outfitted with a “mechanical teacher”. The characters are surprised to learn that teachers were once humans. By now, we are all aware that we have the technology described in the story — chatbots that greatly echo the abilities of people. But a social environment is necessary for children’s development. Instead of fearing artificial intelligence, we could use it to design better learning experiences for our students, braiding a tapestry that combines machine efficiency and human affection, but ultimately still relying on our judgment as these models are not transparent.

As it stands, AI has a bad reputation among educators. It is the Swiss army knife that students might use to write essays, type up code, or translate text. Indeed, some might call OpenAI’s flagship product “CheatGPT.” But where its potential lies for us educators is in its ability to 1) Explain concepts, and 2) Design alternative lesson plans. As previously mentioned, tailoring our message to different kinds of learners is key. Back in 1996, Joseph Beck described that Intelligent Tutoring Systems should consist of compartments related to the domain knowledge, pedagogy, communications, and most importantly, the student. The AI model reworks its output according to the student’s needs and preferences. *
But even when it comes to hard facts, AI is not always reliable. These machines are capable of stringing along information and dishing it out as veritable. This is known as “hallucinating.” To understand why a model makes an error, we should understand the model’s procedures. Mathematically, we base predictions (output) on our reality (input) and we combine it with patterns we learned from observing previous data, using the coefficient and intercept (model weights and biases). What is different about artificial intelligence, is that we do not know what the features of the input are. How does Dall-E, the image generator, define a sad image? Humans would consider lighting and color. But those AI models take into account thousands if not billions of parameters.
So, writing a good medical text, in ChatGPT’s eyes, might rely less on the actual information and more on grammar and using words like “hence” and “diagnosis.” Remember, the whole idea behind these models is for them to generate output that is similar to the data they were trained on. Training them on data is akin to showing a child a bunch of different iterations of the same thing so that they build schemata of concepts, like a dictionary of ideas. If you have a dictionary, does that mean you have an intuitive understanding of the language?
A well-known thought experiment among AI ethicists is the Chinese Room. If you translate text using a dictionary and give the translation someone on the outside without them seeing you using it, you are decidedly still not a Chinese speaker. As previously explained, these are mathematical models. The neural network crunches the numbers, and out comes a result. Thus AI models do have maps of ideas, but they do not have the ideas themselves. They seem to understand that sunny weather is good, because in the texts they were trained on, passages where the word “sunny” appears also had positive emotions described – and so on.

As such, we cannot put our brains on autopilot and let those models do all our work for us. What we can use AI for is aiding us with refining concepts and working out the granular details of planning a project in a streamlined manner. Even though we are seeing great strides in AI rapidly, we will always need the human element. Everyone has a teacher that made them try harder and believe in themselves. Writers do more than string letters together, and educators do more than recite lessons. What stuck more with audiences, the many documentaries about Dunkirk or Christopher Nolan’s motion picture?

فاطمة الورداني