Mechanical Turk評測：21年人力AI平台被AI淘汰收攤 | Mechanical Turk Review: The Human AI Platform AI Killed

8月 29, 2026

By Kit 小克 | AI Tool Observer | 2026-08-30

🇹🇼 Mechanical Turk評測：21年人力AI平台被AI淘汰收攤

亞馬遜宣布旗下的Mechanical Turk（MTurk）將在2026年9月30日正式關閉，這個由創辦人Jeff Bezos在2005年推出、自嘲為「人工的人工智慧」（artificial artificial intelligence）的人力外包平台，服役21年後畫下句點。諷刺的是，把它淘汰出局的正是它當年一手訓練出來的真AI。

Mechanical Turk是什麼？為什麼現在關閉

MTurk讓企業把「人工智慧任務」（Human Intelligence Tasks，HIT）拆成小任務外包給全球工作者，內容涵蓋資料標註、語音／影片轉錄、問卷填寫等機器當年做不好的工作。過去十年，大量電腦視覺與自然語言模型的訓練資料，其實都是靠Mechanical Turk上的人力一筆一筆標出來的。

亞馬遜聲明中只說關閉決定來自「內部評估」，但外界普遍認為原因很直接：

需求萎縮：生成式AI模型本身已能處理大量原本要人工判讀的任務，基礎標註需求大幅下降
競爭者崛起：Scale AI、Mercor、Prolific等專攻AI訓練資料的新創陸續把工作者與客戶都搶走
投資停滯：亞馬遜近年明顯縮減平台維護與功能更新，上月已停止接受新客戶申請

SageMaker Ground Truth也一併退場

值得注意的是，亞馬遜同步關閉依賴MTurk人力的SageMaker Ground Truth資料標註服務，等於連「用AWS一條龍標註訓練資料」這條路都收掉了。想標註資料的團隊，現在得自己找Scale AI、Mercor這類專門公司，或是乾脆用更強的模型做自動標註，人工只負責覆核。

對AI從業者的實際影響

如果你的產品還在用Mechanical Turk做少量資料標註、A/B測試或使用者研究，現在就該找替代方案，9月底前遷移工作流程。這件事也提醒一個現實：AI產業的「人力成本」正從按件計酬的眾包，轉向更集中、更專業的資料公司，散戶工作者的議價空間反而更小。想靠標註賺零用錢的個人，這個管道正式關閉。

Mechanical Turk的落幕某種程度是AI發展的縮影：用人力餵出來的模型，最終讓餵它的平台失去存在意義。

好不好用，試了才知道。

🇺🇸 Mechanical Turk Review: The Human AI Platform AI Killed

Mechanical Turk is shutting down. Amazon announced its crowdsourced labor platform — launched in 2005 by founder Jeff Bezos, who jokingly called it "artificial artificial intelligence" — will close for good on September 30, 2026, after 21 years. The irony is hard to miss: the real AI it helped train is exactly what made it obsolete.

What Was Mechanical Turk, and Why Close It Now

Mechanical Turk let companies break "Human Intelligence Tasks" (HITs) into small jobs farmed out to workers worldwide — data labeling, audio and video transcription, survey completion, exactly the tasks computers couldn't handle reliably back then. For over a decade, a huge share of the training data behind computer vision and NLP models was quietly labeled one HIT at a time by Mechanical Turk workers.

Amazon's statement just cites an "internal assessment," but the practical reasons are obvious:

Demand collapsed — modern generative AI models now handle much of the judgment work that used to require human labeling
Specialized rivals took over — startups like Scale AI, Mercor, and Prolific built platforms purpose-built for AI training data and pulled away both workers and clients
Investment stalled — Amazon visibly stopped improving the platform and quietly stopped accepting new customers last month

SageMaker Ground Truth Is Going Too

Amazon is also shutting down SageMaker Ground Truth, its labeling service built on top of MTurk labor — closing off the "one-stop AWS labeling" option for enterprises entirely. Teams that relied on it now need to move to dedicated vendors like Scale AI and Mercor, or lean more on AI-assisted labeling with humans handling spot review instead of raw annotation.

What This Means If You Actually Use It

If your product still runs small-scale labeling, A/B testing, or user research through Mechanical Turk, start migrating before the September 30 deadline. The bigger signal: AI's "labor cost" is consolidating from piecework crowdsourcing into a handful of specialized data companies, which leaves individual gig workers with less leverage, not more. If you were earning pocket money labeling data on the side, that door just closed.

Mechanical Turk's shutdown is a fitting summary of where AI development has gone: the platform that fed models with human labor eventually made itself unnecessary.

You won't know until you try it.

Sources / 資料來源

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

搜尋此網誌

AI小貼士