Aug 23, 2026
We can all see the potential of AI in shaping education. As any powerful and versatile tool, the effect of it strongly depends on how it's used. Used well, it's an amazing tool and a huge amplifier for learning; used poorly it causes cognitive atrophy and sometimes worse, the illusion that one learns or understands but in fact only convinced of AI's plausible and persusive words.
There are many reports and research that with the advent of advanced AI student's take home assignment/exams got vastly improved scores whereas closed book in-person exams scores got noticably worse; here's one example and I reproduce a chart from it.

There is credible randomized field experiments (PNAS, Bastani et al, 2025) on ~1000 high school students learning math. The standard ChatGPT (GPT base) assisted group of students have large 48% improvements in grades, and a specialized designed safeguarded GPT Tutor assisted group has a whopping 127% improvements of grades, compared to control group (no AI access group). However all these gains in grades are gone for both the GPT base group and GPT Tutor group after AI access is removed subsequently. GPT base group performed 17% worse than control group; and GPT Tutor group essentially performed the same as control group. There are also large scale recent studies coming from China that have similar conclusions.
My experience also largely align with the research and reports that indiscrinmate use of AI is net negative. The key reason is that the most tempting use of AI is the most harmful one--using AI to do the "hard work", the heavy lifting--which turns out to the point of learning! Without overcoming difficulty, the learning is shallow, the mental muscle is not exercised thus gets weaker, and any learning or understanding is delusion.
On the other hand, used well AI is of great help, at the scale and extent we have never seen before. Not to mention that being able to work well with AI is essential skill for knowledge workers. For example one can use AI as a challenger, a sparring partner, which does not take over the exercise itself, but rather helps the user to do it more efficiently and safely (if you lift too much you likely hurt yourself; the analogy for trying too hard to solve too difficulty problem for too long). In my experience, the trick is to never offload understaning (often through hard work, trial and error, blood and sweat so to speak) to AI. Physical exercise is an excellent analogy here: you don't want your trainer to do the lifting for you. Another good use of AI is to help clear misunderstandings that typically requires back and forth with an experienced teacher; now AI is good enough to be able to understand your frame of thinking and patiently point out your misunderstanding and doubts.
So here's the rules for this class regarding AI use: