AI Assistance Undermines Skill Formation
Judy Hanwen Shen and Alex Tamkin's preprint, discussed in a Psychology Today essay, argues that AI offloading has different consequences depending on whether the user already knows the task. In tests with developers learning a new asynchronous programming library, AI could produce working code but weakened conceptual understanding, code reading, and debugging, especially when people fully delegated tasks.
The real split here is not AI versus no AI, it's atrophy versus foreclosure: if you already know the skill, AI can dull it; if you're still forming it, AI can prevent it from ever solidifying.
- –Randomized experiments found no meaningful average efficiency gain, even though some fully delegated tasks did show productivity improvements.
- –The damage landed in the exact places that matter for oversight: conceptual understanding, reading code, and debugging the model's output.
- –The six interaction patterns are the actionable part of the paper, because they show some AI-assisted workflows can preserve learning if the user stays cognitively engaged.
- –The Psychology Today framing makes the policy lesson sharper: junior workers and students need workflows that scaffold judgment, not replace the reps that build it.
DISCOVERED
60d ago
2026-03-28
PUBLISHED
61d ago
2026-03-28
RELEVANCE
AUTHOR
ndr42