跳到主要內容

AI依賴症評測:柏克萊證實,用10分鐘毀解題力 | AI Dependency Review: 10 Minutes of Use Erodes Focus

By Kit 小克 | AI Tool Observer | 2026-10-30

🇹🇼 AI依賴症評測:柏克萊證實,用10分鐘毀解題力

AI依賴這件事,現在有數據了。柏克萊大學人機共存中心(Center for Human-Compatible AI)主導的最新研究發現,只要用ChatGPT輔助解題10分鐘,一旦AI被拿走,使用者解難題的持續力會明顯崩跌,這篇論文已經在Conference on Language Modeling(COLM)正式發表。對天天開著AI助手工作的人來說,這不是危言聳聽,是有對照組的實測結果。

柏克萊AI依賴研究在測什麼?

研究團隊招募了1,222名參與者,讓他們做分數數學題跟SAT閱讀理解題,分成「全程有AI輔助」和「中途被抽走AI」兩組。結果很直接:原本靠AI解題的人,一旦AI在第12題被拿走,答對率立刻掉下去;而從頭自己練習的對照組,表現反而穩定維持。

三組實驗設計了什麼?

  • 第一組:全程可用ChatGPT輔助解題
  • 第二組:解到特定題目後AI被切斷,剩下題目靠自己
  • 第三組(對照組):從頭到尾都沒有AI,純自己練習

三組重複跑了多輪隨機對照實驗,結果一致:有AI輔助的那組一旦失去支援,持續解題的意願跟正確率都明顯下滑,這正是研究論文標題寫的「AI Assistance Reduces Persistence and Hurts Independent Performance」。

為什麼AI依賴會影響解題能力?

研究的解讀是,當AI隨時能給答案,大腦會傾向把「卡住就問AI」當成預設反應,久了遇到困難的忍受度(persistence)會變低,而不是單純「變笨」。這跟過度依賴計算機、依賴GPS導航讓人路痴的邏輯類似,差別是AI依賴的速度快很多——只要10分鐘就能觀察到效果。

日常用AI該怎麼避免依賴?

  • 先自己試算、試寫,卡住超過合理時間才開AI輔助
  • 把AI答案當「核對」用,不是「代寫」用
  • 定期找沒有AI的時段練習同類任務,保持獨立解題肌肉
  • 團隊導入AI工具時,評估的不只是效率,還有長期能力流失的風險

AI依賴不是不能用AI,而是要意識到長期全自動依賴的代價,特別是學習階段的人(學生、新手工程師)更該留意,持續力是練出來的,不是靠AI借來的。

好不好用,試了才知道。


🇺🇸 AI Dependency Review: 10 Minutes of Use Erodes Focus

AI dependency is no longer just a vibe — there's data now. A peer-reviewed study led by researchers at UC Berkeley's Center for Human-Compatible AI, presented at the Conference on Language Modeling (COLM), found that just 10 minutes of AI-assisted problem solving measurably erodes a person's ability to persist at hard tasks once the AI is taken away. If you keep an AI assistant open all day while working, this isn't hype — it's a controlled experiment with real numbers.

What did the Berkeley AI dependency study actually test?

Researchers recruited 1,222 participants for fraction math problems and SAT reading comprehension tasks, splitting them into groups with full AI access versus AI cut off mid-task. The result: accuracy for the AI-assisted group dropped sharply once help disappeared at problem 12, while the control group that practiced alone held steady throughout.

How were the three experiments structured?

  • Group 1: Full ChatGPT access throughout the problem set
  • Group 2: AI access removed after a set point, forced to finish solo
  • Group 3 (control): No AI access at all, solving everything independently

Across multiple randomized controlled trials, the pattern held: the group that lost AI assistance showed a clear drop in both willingness to keep trying and accuracy — matching the paper's title, "AI Assistance Reduces Persistence and Hurts Independent Performance."

Why does AI dependency hurt problem-solving?

The researchers' interpretation isn't that AI makes people dumber — it's that when an answer is always one prompt away, the brain defaults to "ask AI" the moment it hits friction, lowering tolerance for struggle over time. It's similar to how calculators or GPS navigation can dull specific skills, except the effect here showed up startlingly fast — within a single 10-minute session.

How to avoid AI dependency in daily use

  • Attempt the problem yourself first; only bring in AI after a reasonable struggle period
  • Use AI output to verify your work, not to replace doing it
  • Schedule AI-free practice sessions for skills you actually want to keep
  • When rolling out AI tools on a team, weigh long-term skill erosion alongside short-term efficiency gains

None of this means ditching AI tools — it means staying aware of the cost of full automatic reliance, especially for students and junior engineers still building core skills. Persistence is trained, not borrowed from a model.

好不好用,試了才知道。

Sources / 資料來源

常見問題 FAQ

AI依賴症是什麼?

指長期靠AI輔助解決問題,導致大腦遇到困難時習慣直接問AI,獨立解題的持續力與耐受度隨之下降的現象。

柏克萊這份研究可信嗎?

是的,研究由柏克萊人機共存中心主導,1,222人隨機對照實驗,已在COLM會議正式發表peer-reviewed論文。

是不是完全不能用AI輔助?

不是,研究重點是提醒全自動依賴的代價,建議先自己嘗試再用AI核對答案,而不是一卡住就直接問AI代寫。

AI依賴多快會出現效果?

研究顯示僅需10分鐘的AI輔助使用,一旦AI被移除,解題持續力與正確率就會明顯下滑。

延伸閱讀 / Related Articles


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

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code