Learning and AI: Don’t let AI do the hard parts

Strong learning requires that we don't let AI replace the often difficult steps that help us build powerful mental models
I am beginning to teach my undergraduate “Introduction to synthetic biology” class this semester and have been thinking a lot about how I want students to use AI. We read a fair number of scientific papers and do associated writing in the course, and there are quite a few AI tools out there that students may be thinking about using.
While still in the early days, AI is showing its potential to have a positive impact in addressing complex problems in various fields. However, in education, I think we are really struggling. We don’t actually know how to best incorporate AI into the classroom. How should students use it or not use it? In what way do we make sure students are actually learning — altering their own brain circuitry and gaining lasting knowledge and skills — when AI tools can take away the parts that are essential for real learning? How do we capture the good capabilities that come with AI without damaging our innate strengths and capabilities?
AI is only going to get better and more pervasive, so figuring out how it fits into learning and education is important for all of us.
Synthesis and mental models
A major component of learning, regardless of the field, is synthesis. I’ll define it here as pulling out the major concepts, themes, and ideas established across a range of previous work and putting them back together into your own mental model of how it all works or what it all means. With this model, you can now answer a range of arbitrary questions, including the answer of “I don’t know” that lets you know when your model is incomplete. This mental model also gives you the ability to make “what if” predictions and generate hypotheses and insights about what could be. This is a powerful thing to have for yourself and well worth the effort to build it. (I am focusing on more “knowledge work” areas, though the same thing applies to trades and craftspeople. They also build mental models of their domains that enable them to do novel and significant work in their fields — e.g., metal working, music, choreography, photography, etc.)
If we are boiling it down to money, I’d argue that mental models are fundamentally the key “value” you bring to the conversation. It is the diversity, depth, and richness of your mental models, along with your ability to communicate them, that determine how much people are willing to pay.
Building your initial mental models
In knowledge work, a simplistic view of how you start to synthesize and build these models starts with you reading multiple books, papers, or other work, and then identifying and summarizing the major points, themes, and ideas. This is hard.
You then have to spend further time and effort integrating and synthesizing this information across all the other work you have been looking at and incorporating it with your previous knowledge. One of the best ways to do this second part is through written elaboration. This is very hard.
Keep repeating this cycle and incorporate retrieval practice, testing, and a few other practices, and you have started on your path to mastery. All hard.[1]
AI summarization
AI writing is its own topic, so I’ll look at just the first step of this process for now.[2] One of the earliest mainstream tasks AI tools have been able to do reasonably well is summarizing writing. This summarization ability is built into numerous applications, where articles and transcripts from videos and other media can be read by a model, with the major points extracted and served up to you. For scientific papers, this ability to summarize was initially weak, but has improved rapidly both in identifying the key points and broader accuracy. Reading and summarizing papers and then discussing the key points with others is a major component of learning in and out of the classroom.
So a major question here is, if you don’t go through the effort of reading and summarizing an article for yourself and let an AI do it, did you learn anything?
Offhand, I’d say no. Coming up with the summary yourself is a big part of learning the material. Even if you go back and read the full paper yourself, there is the additional concern that, having seen the AI summary, you are biased and going to tend to just look for and find the same major points. Avoiding this bias can be done, but you would have to be extremely self-aware to make sure you weren’t doing this.
But what if I use AI summarization to scan and identify many papers as an initial screen? I could use it to summarize these articles and then only look at those that, I think, will be most valuable to read myself. How about that? It is already impossible to read the huge number of papers that are produced in any given field, so such a tool would be really useful. (In full disclosure, I have used AI systems in this way, to find papers on a given topic along with generated summaries so that I can figure out which ones I want to read further. I have found this extremely useful.)
However, when I have done this, the question I start to have is how sure am I that I am not missing something significant? Is it actually finding the “right” papers? Are the summaries accurately pulling out the key points? The “right” points? Does reading the “less interesting” or relevant papers actually give you essential insight into why the good and important papers are good and important?
This last part is intriguing and analogous to the training of these same AI systems. Training with negative data is a key factor in improving the performance of AI models. They need to see both positive and negative data to have the best performance, and not just be limited to only learning from the “correct” or “best” parts.
I think students need the same thing — to see both the good and bad to help better understand what distinguishes high-quality science from bad. Reading, summarizing, and comparing the good and the bad is a massive part of strong learning. Active comparison and discussion of these summaries is particularly significant for learning, and it is really difficult to do this if you haven’t developed the summary yourself and didn’t build the justification for them on your own. This early (and difficult and often boring) work is what creates the foundation for your own mental models.
Maybe when you already have significant training in an area, it is easier to avoid these pitfalls, but I’m not so sure. The ease of these tools and the fact that they do remove many of the difficult aspects makes them incredibly appealing.
But fundamentally, the greatest challenge is that learning that produces a long-lasting impact on our brains requires doing things that are hard and difficult. It is just so easy for us to take the easier path first and then rationalize after the fact that nothing was lost as a result.
When starting off learning a topic, not using AI summaries may likely be the best way to begin.
What next?
My students will be reading various involved and scientifically complicated papers this semester. I want them to understand the methods and ideas involved, as well as potentially come up with their own opinions about the next steps that should be taken to advance the field. They should come up with these insights based on the mental models they are starting to build for themselves. I don’t want them to have the false sense that they learned something when they likely never did.
For right now, students can use AI however they want, but they do have to point out all the instances of when and how they use it. As a first pass, I am hoping this can provide regular conversations about when AI may be helping them along with when it may be hurting them. This will hopefully start to build a self-awareness that will help them know how to tell the difference.
We will see how this goes, but I’d love to hear if others have ideas on how to make AI work in teaching and learning more generally!
Extending these models goes even further. For example, in scientific research we also incorporate experiments of various types that allow us to test these models, where they are working and where they aren’t. ↩︎
Nothing here was written by an AI, so any errors and poor communication are all my own human ones. ↩︎