How we read the signal

Analysis frame

Evidence level

Primary-source evidence

Analytical lens

Distinguish experimental short-term assistance from later unaided performance, and ask whether a system optimizing completion displaces the practice required for durable skill.

Affected groups
  • Students and workers learning reasoning-intensive tasks
  • Teachers, employers and developers choosing how to measure progress
What remains unknown
  • The trials do not establish long-term effects of months or years of AI use
  • The subgroup pattern for direct-answer use is not a randomized causal comparison
Second-order effects to watch
  • AI-assisted productivity metrics may reward output while hiding later deskilling
  • Hint-first product designs could preserve learning but initially appear less helpful

What the experiments actually tested

Participants used AI for brief arithmetic or reading tasks, then completed comparable problems without it. The randomized assignment supports a causal conclusion about those experimental conditions, not a universal claim about every kind of AI use.

The missing measure

An assisted score is not the same as retained competence. The next evaluation should compare direct answers with hints, Socratic prompts and delayed help over longer periods, then test unaided transfer.

Primary trail

Go to the source

Read the evidence behind this analysis. External links open in a new tab.

arXiv — AI Assistance Reduces Persistence and Hurts Independent Performance arXiv — full paper, version 5