Notes by Lex Toumbourou

Something I’m increasingly worried about is that the more mental labour we offload to AI, the more our own cognitive capacity starts to dwindle.

In a popular paper from 2025 (AI Meets the Classroom: When Does ChatGPT Harm Learning?), they found that students in a programming course who used LLMs to generate answers to questions (substitution) saw a significant decrease in topic understanding, even though they perceived an increase Lehmann et al. (2025).

Learning is effortful: you need to exert mental energy to learn. However, AI can give you the illusion of learning without any mental exertion at all.

The study also found that students who used AI to complement learning activities, like asking for explanations rather than answers, actually saw an increase in understanding. But when given a free choice, most students defaulted to substitution and found that nearly half of all solution requests came without the student making a single attempt first. Clearly, our brains tend to prefer to conserve energy. They also found that students with weaker foundations learned less with AI access, while stronger students benefited more.

For a software developer, it’s impossible to ignore the upside from modern agentic coding workflows; they are clearly here to stay. But, at the same time, for complicated projects beyond what can be vibe-coded, a pre-developed intuition for software is necessary to guide those agents safely.

So the natural question arises for a software developer, and a knowledge worker in general: how do we find the balance between making the most of AI while avoiding letting our skills, intuition and understanding decay?

Just as many people leading sedentary lifestyles have to make a deliberate effort to exercise, because inactivity is really bad for our bodies, I think we’re going to realise that a similar process is necessary for our minds.

My hypothesis is that you need to deliberately spend time every day exercising your skills, operating near your cognitive limits - performing deep System 2 Thinking - if you want to avoid the kind of atrophy that will make you worse at wielding AI agents. The idea of Progressive Overload in strength training is to gradually increase resistance via heavier weights, building overall capacity. The same paradigm applies to cognitive capacity.

One thing I've been experimenting with is making sure I allocate at least 30 minutes a day on something mentally difficult, like writing code by hand for an unfamiliar problem or reading difficult papers. Writing these notes is another task that clearly feels like a form of this exercise.

Cover by simon follin on Unsplash