New research posted to arxiv finds that AI assistance doesn't just change how fast people complete tasks - it changes how they approach difficulty. People who receive AI help on tasks show reduced persistence (they give up sooner when things get hard) and perform measurably worse on the same tasks when the AI is removed.
The paper's title says it plainly: "AI Assistance Reduces Persistence and Hurts Independent Performance." For anyone using AI tools every day, that finding deserves more attention than a standard benchmark story.
The GPS Effect, But Broader
The pattern researchers found isn't new - we've seen versions of it before. People who rely on GPS navigation lose the ability to build spatial memory on familiar routes. Spell-check dependence correlates with higher unassisted error rates. Calculators introduced too early in math education reduce mental arithmetic fluency.
In each case, a tool that removes friction from a task also removes the practice that builds the underlying skill. AI assistance is a more aggressive version of this because the scope of what it handles is so much wider. A GPS does one thing. A coding assistant like Cursor or an AI like Claude can handle debugging logic, API syntax, code structure, and test writing - essentially the entire problem-solving loop that developers need to internalize to improve. A writing AI can supply word choice, paragraph structure, and argumentation, the same cognitive work that makes a writer better over time.
When AI takes over that work consistently, the practice reps drop toward zero.
Who Carries the Most Risk
The obvious response - "just stop using AI" - doesn't hold up economically. If AI drafts an email in 10 seconds that would take 5 minutes to write from scratch, and the quality is equivalent, the math is clear.
The more useful distinction is between two types of AI use. The first: using AI to handle tasks you already know how to do, faster. The second: using AI to avoid the struggle of learning tasks you haven't internalized yet.
The first is a productivity trade. The second is a skill trade, and the research suggests the deal is worse than it looks. That stuck feeling before reaching for Claude or ChatGPT - the 10 or 20 minutes of thinking through a hard problem before asking for help - is probably where most skill development happens.
This matters most for anyone earlier in their career. Senior practitioners offloading routine work to AI are drawing on capabilities they already have. Junior developers and early-career writers using AI as a first resort may be stalling their own development without seeing it happen.
The practical implication isn't to avoid these tools. It's to be deliberate about when in the problem-solving process you reach for them - and to notice when you're using AI to escape discomfort rather than to move faster on things you already understand.