Ten minutes of AI assistance reduce persistence without the tool

What this is about

A study presented by UC Berkeley on October 9, 2026 does not ask whether people give better answers with ChatGPT. It asks a more uncomfortable question: What happens immediately afterward when the tool suddenly disappears? Across three randomized trials with 1,222 participants, the AI assisted group initially performed better. Once access was removed, its solve rate fell and participants skipped more tasks.

The paper is titled AI Assistance Reduces Persistence and Hurts Independent Performance. A preliminary draft circulated in spring 2026. The revised, peer reviewed version was presented at the Conference on Language Modeling in October.

What the study actually does

In the first trial, 354 people worked through 15 basic fraction problems. One group could use ChatGPT, including asking directly for answers. After problem 12, the assistance was removed. A second trial with 667 people repeated the setup. In a third trial, 201 participants completed an SAT style reading comprehension task.

Researchers measured accuracy, persistence, and how often participants skipped or abandoned a question. While ChatGPT was available, the assisted group was more accurate. After removal, however, its performance dropped sharply, while the control group without AI remained stable or improved. The design therefore separates a short term productivity gain from the ability to continue doing the same kind of work independently.

Why it matters

Schools and companies usually assess AI tools by how quickly a task gets completed. This study reveals a possible blind spot: Usage metrics can look excellent even when independent performance declines immediately afterward. That matters most in work where people must remain capable during an outage, an examination, or a security incident.

The research team includes scholars from UC Berkeley, Carnegie Mellon, MIT, Oxford, and UCLA. Its findings support the idea of productive struggle: Trying a problem first helps people build knowledge and strategies. Receiving the complete answer immediately saves time but may provide less practice. This does not justify a blanket ban. A more useful sequence is to build the skill first, add assistance second, and periodically test without assistance.

In plain language

Imagine riding an e bike uphill for ten minutes. With the motor, you reach the top faster. If the motor switches off halfway up the next hill, that same climb may feel harder than it does to someone who rode without assistance from the start. The study measures that moment. It does not say e bikes are bad; it shows that assistance and personal training serve different goals.

A practical example

A customer service team introduces an AI system that drafts replies to complaints. One hundred employees practise for two weeks using only the system's suggestions. A second group writes an answer first and sees the suggestion afterward. In a monthly exercise without AI, each person handles 20 cases. If the first group skips more cases or confuses more rules, that would be a signal to change both training and interface. The system could ask questions and offer hints before revealing a complete answer.

This example is fictional. It does not treat the measured effect as a proven workplace number. It shows how an organization could test the issue using its own tasks.

Scope and limits

• The trials covered fractions and SAT reading comprehension. Open ended creative work, programming, and long term professional practice may behave differently.
• Researchers measured performance immediately after removing assistance. The study does not prove lasting cognitive decline or a long term effect.
• Berkeley's summary does not report an effect size. Education or workplace policies should therefore be based on the full paper and tests in the relevant setting.

The central limitation is built into the design: If a tool remains continuously available, unaided performance is not the only relevant measure. It is still crucial for examinations, outages, emergencies, and foundational expertise.

SEO and GEO keywords

AI assistance, ChatGPT, UC Berkeley, persistence, independent performance, productive struggle, education, workplace, randomized trial, Conference on Language Modeling, human skills