Key Takeaways

  • Three randomized experiments found that AI assistance improved initial task accuracy but reduced persistence after access was withdrawn
  • The results raise workforce training concerns as ChatGPT, Microsoft Copilot, and Google Gemini become common workplace aids
  • Enterprises can reduce cognitive offloading by testing independent performance and keeping employees involved in reasoning

Berkeley News reports that short-term access to ChatGPT may leave users less willing or able to persist once the tool disappears. Across three randomized experiments involving 1,222 participants, people used ChatGPT for roughly 10 minutes before continuing without it. The AI-assisted groups initially solved tasks more accurately, but their accuracy and persistence fell sharply after access was removed.

The study does not establish permanent cognitive harm, nor does it prove that every use of generative AI weakens human capability. It does, however, identify a potential tradeoff that enterprises may overlook when measuring AI programs primarily through immediate productivity.

A worker who drafts a document with Microsoft Copilot has produced a useful result. But assessing whether that worker can explain the underlying reasoning, identify a subtle error, or complete the assignment during an outage requires different measures of performance.

Most enterprise AI evaluations focus on output while the assistant is available. Organizations track time saved, tickets closed, documents created, or code produced. Far fewer test what happens when employees need to perform the same work independently. The UC Berkeley findings suggest that this second measurement could reveal risks hidden by strong short-term results.

The pattern aligns with productive struggle, the educational concept that effort supports learning even when it slows immediate progress. Generative AI can remove that effort by supplying a polished answer before a user has built a mental model of the problem, obscuring the critical difference between substitution and assistance.

A separate randomized trial published in 2025 underscores that concern. Among 120 undergraduates, ChatGPT users scored 57.5% on a delayed knowledge-retention test, compared with 68.5% for students using traditional learning methods. That 11-point gap does not settle the broader debate, but it adds evidence that better performance during an AI-supported session may not translate into stronger retention.

Microsoft’s 2025 Work Trend Index surveyed 31,000 workers across 31 markets and reported that 75% used AI at work. According to the data, 46% viewed AI as a “thought partner.” That framing differentiates contributing to reasoning from substituting human effort entirely.

For business leaders, the practical response is not a blanket restriction on ChatGPT, Microsoft Copilot, or Google Gemini. A more measured approach focuses on designing workflows that preserve human judgment. Employees might draft an initial analysis before consulting AI, critique an AI response rather than accept it, or periodically complete tasks without assistance. Short explanations of why an answer is correct can also expose shallow understanding.

Training programs can apply similar principles. Instead of grading only AI-assisted output, instructors can test delayed recall, independent problem-solving, and the ability to detect fabricated or weakly supported claims.

The NIST AI Risk Management Framework gives organizations a structure for mapping and monitoring AI risks, while UNESCO’s guidance on generative AI in education and research emphasizes a human-centered approach. Neither framework treats AI deployment as a simple software installation.

The UC Berkeley research leaves important questions open, including whether the effect fades, varies by task, or changes as users gain experience. Broader organizational consequences also remain unproven. Even so, the operational lesson is concrete: enterprises should measure what employees can do with AI and what they retain without it. Productivity gained today may be less valuable if independent capability quietly erodes tomorrow.