ChatGPT依賴評測:柏克萊研究證實10分鐘毀耐力 | ChatGPT Review: 10 Minutes of AI Use Erodes Persistence
By Kit 小克 | AI Tool Observer | 2026-10-11
🇹🇼 ChatGPT依賴評測:柏克萊研究證實10分鐘毀耐力
ChatGPT只要用10分鐘,解題耐力就可能垂直下滑——這不是危言聳聽,而是柏克萊大學人機相容中心(Center for Human-Compatible AI)研究員Brian Christian團隊在2026年10月發表的實測結果。這份論文〈AI Assistance Reduces Persistence and Hurts Independent Performance〉找來1,222名受試者做了三組隨機對照實驗,結論直指一個多數人用ChatGPT時都沒意識到的副作用:工具一拿掉,表現比沒用過AI的人還差。
三組實驗怎麼設計的
研究團隊設計了三個獨立實驗:第一組354人、第二組667人做分數應用題,第三組201人做SAT閱讀理解。兩組受試者都先解12題,差別是其中一組全程可以用ChatGPT輔助,另一組完全靠自己。關鍵操作在第12題之後——研究人員把AI工具收回,讓所有人繼續靠自己解題,藉此觀察「依賴AI的人」拿掉工具後會發生什麼事。
拿掉AI之後,差距馬上現形
- 全程靠自己解題的對照組,準確率維持穩定甚至略有進步
- 先用ChatGPT輔助的組別,工具收回後答對率明顯下滑,跳題(skip)比例也大幅增加
- 效應不只出現在數學應用題,SAT閱讀理解測驗也觀察到同樣的耐力與準確率下滑
換句話說,用ChatGPT解題當下看起來很有效率,但那份「能力」有一部分其實是工具借來的,工具收回,能力也跟著收回。
不只是數學題,是「人類專業能力」的流失
Brian Christian在受訪時特別強調,這個現象不會停在國小分數題:「這件事正在發生在人類知識與專業能力的每一層,從國小學生到頂尖專家都無法例外。」他認為問題核心在於,ChatGPT這類工具習慣直接給答案,等於剝奪了使用者「卡關、掙扎、最後自己想通」的過程——而那個過程正是建立長期解題耐力的關鍵。
對一般用戶與產品設計的啟示
研究團隊建議,AI產品不該預設「直接給答案」,而該更像家教那樣引導式提問,才能在提升效率與保留使用者自身能力之間取得平衡。對一般用戶來說,實務建議很簡單:練習、學習類任務先自己卡關一段時間,把ChatGPT留到真正卡死才用,而不是每一步都先問AI——不然短期效率換來的,可能是長期耐力的流失。
好不好用,試了才知道
這份研究沒有說ChatGPT不該用,而是提醒大家:工具用得越順手,越該留意自己有沒有在默默把「想辦法解決問題」的肌肉外包出去。拿掉AI之後你還剩下多少實力,才是真正的試金石。
🇺🇸 ChatGPT Review: 10 Minutes of AI Use Erodes Persistence
Just 10 minutes of using ChatGPT can measurably erode your ability to persist at hard problems — that is the blunt finding from a October 2026 study led by Brian Christian, a researcher at UC Berkeley's Center for Human-Compatible AI. The paper, "AI Assistance Reduces Persistence and Hurts Independent Performance," ran three randomized controlled experiments on 1,222 participants, and the result points to a side effect most people never notice while using ChatGPT: once the tool is taken away, performance drops below that of people who never had it.
How the Three Experiments Were Designed
The research team ran three independent trials: 354 participants on fraction word problems, 667 on a similar fraction task, and 201 on SAT reading comprehension. All participants solved 12 problems; one group had ChatGPT access throughout, the other worked unaided. The critical twist came after problem 12 — researchers removed AI access entirely and had everyone keep solving problems on their own, to see what happens to people who had been relying on the tool.
The Gap Shows Up Immediately After AI Is Removed
- The unaided control group held steady accuracy, even improving slightly over time
- The group that had used ChatGPT saw accuracy drop sharply once the tool was removed, with skip rates rising noticeably
- The effect wasn't limited to math — the SAT reading comprehension experiment showed the same drop in persistence and accuracy
In other words, solving problems with ChatGPT feels efficient in the moment, but part of that performance is borrowed from the tool. Take the tool away, and the borrowed capability goes with it.
Not Just Math — Human Expertise Itself Is Eroding
Christian was explicit that this doesn't stop at grade-school fractions: "This is something that is happening to human knowledge and human expertise from top to bottom and from grade-school students all the way to the leading experts." His read on the mechanism: tools like ChatGPT default to handing over answers immediately, which denies users the struggle-and-eventually-figure-it-out process that actually builds long-term persistence.
What It Means for Everyday Users and Product Design
The researchers argue AI products shouldn't default to instant answers — a more tutor-like, guided-question approach could balance efficiency against preserving users' own capability. The practical takeaway for everyday ChatGPT users is simple: let yourself get stuck on learning tasks before reaching for AI, and save it for when you're genuinely blocked rather than querying it at every step. The short-term efficiency gain may be costing you long-term persistence.
好不好用,試了才知道
This study isn't an argument against using ChatGPT — it's a reminder that the smoother a tool feels, the more worth checking whether you're quietly outsourcing the muscle of "figuring things out." What you've got left once the AI is gone is the real test.
Sources / 資料來源
- UC Berkeley News - Using AI for just 10 minutes erodes your ability to persist at hard things
- arXiv - AI Assistance Reduces Persistence and Hurts Independent Performance
- EurekAlert - Using AI for just 10 minutes erodes your ability to persist at hard things
延伸閱讀 / Related Articles
- Claude Haiku 5.5評測:降價75%,超額用量才見真章 | Claude Haiku 5.5 Review: 75% Cheaper, Catch at Scale
- GPT-6評測:互動UI上線,三版本定價差距大 | GPT-6 Review: Intelligent UI Arrives, Tiers Diverge
- LMCache漏洞評測:vLLM快取層爆CVSS 9.8嚴重漏洞 | LMCache Review: Critical CVSS 9.8 RCE Flaw Hits vLLM
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言