Problems aren't problems anymore
Historically engineering education has been what I am calling "calculation forward." We went to lecture and listened to conceptual discussions and hopefully took a few notes, but the real learning happened while working out our homework.
The work was generally referred to as "problem solving" and the assignment itself was probably called a "problem set." However, these were actually just well-defined calculation requests that were only problems for us because we had limited information to solve them. They weren't "Capital-P Problems" because the answers were already well known. They were toy problems with a specific learning objective attached. The equivalent of scales for a musician, or drills for an athlete.
When I took kinematics for example, my only resource was my paper textbook and my notes, and so it was easy for the professor to give a "problem" that was not represented in that content. This forced me to wrestle with concepts and apply them in a way that was novel to me. I would argue that was the actual learning experience. By going through that exercise I got the systems thinking, the dimensional reasoning, the ability to just get myself unstuck, and everything that we generally associate with problem solving skills.
The ubiquitous access to information, AI and otherwise, has cut the legs out from under this tool. The classical calculation requests that we would depend on to teach concepts are well documented online, so even without AI students can find the answer to most teaching problems. So many of our classic subjects are well represented on YouTube. Google will handle the unit conversions. AI has drastically accelerated this and removed the last bit of friction, but the issue has existed for some time. It's now totally possible to get the calculation correct and learn no meaningful problem solving skills. The worst part is it feels like learning when you go look up the solution.
We have to find a way to teach dimensional reasoning, systems thinking, critical thinking, and getting unstuck without relying on take-home calculation/derivation problems. The Problem (Capital P) that we have is that there is no one alive in a teaching role who learned under this paradigm. Everyone to some extent learned in the calculation forward model. There are people with ideas, and people who were better at drilling into concepts than others, but at some point we all learned in a "now go do it" framework.
LLM Behavior
To understand this problem fully we have to understand mechanistically what these tools are actually doing. Transformer models work by analyzing connections between words. More context leads to more connections and more nuanced and rigorous output. The reason these tools are so good at solving homework assignments is that the assignment itself is an exercise in supplying context. We never just tell a student "prove to me that you can design a distillation column." Instead we supply specific context that narrows the scope and gives the student a framework to work in. That same context is exactly what the LLM needs to work from.
When problems get more nebulous, it's up to the person prompting the LLM to supply context. This can take two forms. In the first, the person generally understands the context and supplies it outright. The equivalent of writing a homework problem statement. In the second, the human is not aware of all the required context but has some idea of the big picture, and can step back and use the LLM for both planning and execution. Basically treating it as a collaborator. I believe this is the highest impact application of these tools. However it is not a rote process and ultimately requires a well-developed knowledge and experience framework. The open question for society is whether it's possible for new learners to build that framework in the presence of these tools when so much low-level friction has been removed. It is a neuroscience fact that friction and negative consequences lead to stronger neural connections.
The Human Element
A large percentage of faculty and administration seem to treat the students as a given. However there is already a meme percolating through society that college is not worth the money. STEM usually gets a pass with some sort of "...unless you study engineering" disclaimer. But our current students are less capable, not more. I think it will only take a few years of underemployment among graduates for that qualifier to start to erode. The vast majority of the university budget is undergrad tuition and there is likely an enrollment decrease coming due to the demographic issues within the state. There will be a massive space for competition among universities, and there will likely be clear winners and losers.
There is also a mindset among some of the faculty that it is the students' responsibility to learn and "just not use" the tools available to them. This is both disingenuous and lazy. It is disingenuous because it rests on the assertion that we were somehow making better decisions when we were 19 years old. Most of us were not. We were just doing what was required of us. If we are going to take their money then we have an obligation to create an environment that demands that they rise to the occasion. That's what was given to us. The fact that we can't accomplish this with business as usual is simply one of our generation's burdens to bear.
Solutions
I have heard folks talking about custom chatbots as one possible solution. There are a couple of obvious pitfalls here. The first is that no one is about to compete with the frontier LLMs. Ed Tech does not have the capital for either the manpower or the hardware. Whatever it is that people think they are about to do, Anthropic, OpenAI, DeepSeek, etc. are already doing better. I personally think these tools have a valuable place in the future of education, but students have to be taught how to learn in order to use them. If you already know how to learn and have some knowledge framework then these tools are amazing teachers. But it cannot be overstated that self-directed learning requires a significant skill set.
There are also faculty who simply want to revert back to more in-class exams. This is definitely part of the solution and I am personally weighting my exams higher and giving more of them (closed book, closed note). But this alone does not address the underlying learning issues. It might combat grade inflation and offer some motivation to a certain number of students, but it will not address the core critical thinking skills gap that we are seeing. And our ultimate goal is not gatekeeping grades but developing students who can make an impact in the work force.
If our goal is to produce students who can reason and leverage these powerful tools, I believe we need a wholesale restructuring of the way we teach. What I am actually attempting in my current courses is using AI not as an administrative assistant or a grading tool, but as a collaboration partner and content quality multiplier. I am taking all the pedagogical ideas about experiential learning and flipped classroom and conceptual reasoning and leveraging AI to generate comprehensive course material that (hopefully) addresses some of these issues.
The workflow looks roughly like this:
Phase 1 (Context) A long discussion about the overall curriculum: where the course fits, what the learning outcomes are (official and aspirational), what state students are in when they arrive, where the problems are, and what my vision is for the course. No content creation at this stage, just discussion. You really can't supply too much context here.
Phase 2 (Story arc) Now we plan the big picture. What gets covered each week, specifically, and how it all fits together. Pedagogical decisions about how content fits together happen here.
Phase 3 (Weekly architecture) Now we go down a level. How exactly do the pieces fit together in a given week? What's the tempo of the learning? What kinds of activities are in the studio? No specific content yet, just a very detailed outline of how all the parts will fit together.
Phase 4 (Content generation) Now that all that context is in place, we start generating actual content. Studios and homework with detailed references for TAs. These studios can include engineered sticking points, conceptual explorations, and real time problem solving. It all fits together and is anchored in the work done in the earlier phases.
The important distinction is that this is not really saving time in the sense of spending less time on a particular task. The quality of what comes out simply exceeds anything I would be capable of producing alone. A faculty member with deep subject matter expertise and a clear pedagogical vision, working this way, can build learning experiences that are more coherent, more contextually grounded, and more responsive to where their students actually are than anything the traditional model could produce at a reasonable time investment.
The open question is how you create the conditions for faculty to work this way at scale. The LLMs are powerful enough to adapt to any individual's teaching style and course content. The bottleneck is not the technology.
There are two distinct constraints worth separating. The first is willingness and time. The front-loaded investment is real, and most faculty are already stretched between teaching, research, and administration. That's an incentive and support structure problem, and it seems solvable to me but it requires deliberate institutional effort.
The second constraint is harder. This workflow draws heavily on pedagogical depth. Not just subject matter expertise, but the accumulated intuition that comes from watching students struggle with the same concepts across dozens of courses, learning where the misconceptions are, developing a feel for pacing and sequencing and where things break down. The AI is an amplifier for that experience rather than an adequate substitute. Garbage in, garbage out.
This means the intervention we need isn't as simple as "here's how to use AI for course development." It has to grapple with faculty development in a more fundamental way. Longer runways for new faculty, intentional pairing of new and experienced teachers, or some other mechanism for transferring pedagogical depth that currently lives only in individuals. That is a much bigger and more interesting problem than tool adoption, and I don't think many people in this conversation have named it yet.