AI Assistance Reduces Persistence and Hurts Independent Performance: New RCT Evidence from 1,222 Participants

Available in: 中文
2026-04-07T16:44:29.476Z·1 min read
Researchers conducted multiple RCTs examining human-AI interaction across mathematical reasoning and reading comprehension tasks. The findings are causal, not merely correlational.

A rigorous randomized controlled trial (N=1,222) has found that AI assistance, while improving short-term task performance, significantly reduces human persistence and impairs unassisted performance. The effects emerge after just 10 minutes of AI interaction — raising fundamental questions about how we deploy AI tools.

The Study

Researchers conducted multiple RCTs examining human-AI interaction across mathematical reasoning and reading comprehension tasks. The findings are causal, not merely correlational.

Key Findings

Short-Term Benefit

Long-Term Cost

Why This Happens

The researchers posit a mechanism: AI conditions people to expect immediate answers, denying them the experience of working through challenges independently. Like a calculator that makes mental arithmetic atrophy, AI assistance may erode the cognitive persistence needed for deep learning.

The Contrast: Good vs. Bad Collaboration

Implications

  1. Education — AI tutoring systems need "productive struggle" features, not instant answers
  2. Workplace — Over-reliance on AI could erode problem-solving skills
  3. AI design — Models should scaffold long-term competence, not just complete tasks immediately
  4. Policy — Organizations should think carefully about which tasks get AI assistance

Why This Matters

Persistence is foundational to skill acquisition and one of the strongest predictors of long-term learning. If AI tools systematically undermine persistence, the long-term costs may outweigh short-term productivity gains.

↗ Original source · 2026-04-07T00:00:00.000Z
← Previous: Width Growth in Language Models: Exact Copy Warm Starts Surprisingly Beat Complex Initialization StrategiesNext: Predicting Student Video Behavior with Multimodal LLMs: 77 Million Events from 66 Online Courses →
Comments0