[marginalium]
AI use and skill formation
5 Feb 2026
AI use and skill formation. A paper out of Anthropic’s alignment fellowship program. Coders coding with and without AI. Not super compelled by the specifics—not a power analysis in sight. The general trend is what everyone seems to be remarking on (e.g. here and here). Cognitive offloading—you’re offloading skill acquisition because you know you can rely on the resource to do it.
Another problem for newbies.
More interesting was the ‘interaction personas’. They had six different groups of AI users, with various outcomes. Again, specifics seem a bit iffy, but if you look at them as two groups—three which learned and three which didn’t—then you get a difference in people who engaged cognitively and didn’t. In that sense, I guess the paper identifies a couple of ways of cognitively engaging in tasks while using AI: asking conceptual questions, and asking explanatory follow ups.
All together another argument for AI for content, and attention to method.
Less optimistically, I don’t see anything here that suggests that AI enables over and above just doing it yourself. We might be more productive, but it still seems very likely that, even if there are better and worse ways of using AI, using AI might be worse for learning.
Equally though, the paper is just about procedural knowledge—improving on doing specific things from noticing and correcting errors. Not really about more abstract reasoning about whether this specific thing or that specific thing is better at a more abstract level. Much of my own notable learning thorugh AI comes from the fact that when I ask it to do stuff, it does different stuff from what I’d do, and often this is better than what I’d do. Learning these kinds of new patterns requires ZPD-style education, and AI seems like it can provide that.
Of course, even that is still subject to the tyranny of the authority. So, positive on net? Still unsure.
Anthologies: Betterment, Digital Architecture, Wealth Architecture, On Being Fruitful, On Thinking and Reasoning, Humans Aren't Special