MIT-Oxford Study Documents Immediate Drop in Task Persistence After 10 Minutes of AI Assistance

The paper titled AI Assistance Reduces Persistence and Hurts Independent Performance, presented at the 2025 Conference on Language Modeling, divided subjects into AI-assisted and unaided cohorts across three trials. After the AI was withdrawn, the assisted cohort solved fewer fraction problems correctly and quit more often, while the control cohort maintained higher completion rates. The design isolates the effect of brief exposure rather than long-term habituation.

Prior cognitive-decline research on smartphones tracked changes over years; this study registers measurable shifts inside a single session. State education ministries and defense research agencies therefore face an incentive to track whether reliance on commercial models erodes domestic human capital faster than rival programs can be scaled. No primary document yet records policy responses from Beijing or Washington.

The authors note that current systems optimize for immediate query resolution without regard to downstream user autonomy. Governments that treat AI as a productivity multiplier must now weigh the documented cost to independent problem-solving capacity when setting procurement or classroom standards.

Next data points will come from longitudinal follow-ups scheduled for 2026 and from any national workforce assessments that begin disaggregating AI-exposed cohorts.