《帝國AI》作者郝凱倫（Karen Hao）在接受彭博社專訪時，將OpenAI、Google、Anthropic等頂尖人工智慧公司比作歷史上的帝國，指出它們正以前所未有的速度攫取經濟和政治權力。她強調，這些公司不僅大規模擷取個人數據與創作者的智慧財產權，還依賴一支被嚴重剝削的隱形勞動力——包括在肯亞負責過濾極端內容的外包工人，以及在美國從事低薪「數據標註」工作的科學家、律師和博士畢業生。與此同時，AI自動化正在壓縮知識工作者的就業市場，迫使他們從全職崗位流向零工經濟，美國四年制大學學歷失業者佔比已達歷史新高的25%。

郝凱倫以馬爾他與OpenAI簽署全國性合作協議為例，質疑發展中國家和小型國家在AI浪潮中的角色。她認為，這類合作表面上是為公民提供數位工具，實質上卻是讓AI公司免費獲取整個國家的用戶數據來訓練下一代模型——因為高品質網路數據已近枯竭，直接從用戶端採集成為最佳來源。她進一步指出，人們會將病歷等極度敏感的資料上傳至ChatGPT或Claude，卻缺乏對這些強大企業的問責機制，數十億受影響的民眾沒有任何正式管道來挑戰或反饋這些公司每日做出的無數決策。

在技術路線方面，郝凱倫主張應從「火箭式AI」轉向「自行車式AI」。她以DeepMind的AlphaFold為典範——它用極少的數據和算力精準預測蛋白質結構、榮獲諾貝爾化學獎，且不涉及勞動剝削。她同時指出中國公司DeepSeek在美國晶片禁令下，以顯著更少的計算資源實現了同等能力，證明高效低耗的AI開發路徑切實可行。對於Anthropic和OpenAI即將啟動的IPO，她認為上市帶來的治理結構或許能提升問責性，但根本的變革仍需來自公眾和民選官員，而非僅靠市場力量或更換領導者。

In a Bloomberg interview, Karen Hao, author of Empire of AI, likens leading AI companies such as OpenAI, Google, and Anthropic to historical empires that have amassed extraordinary economic and political power. She highlights the exploitative labor practices underpinning large language models—from Kenyan workers forced to classify graphic content for ChatGPT's safety filters to highly educated American professionals pushed into poorly paid data-labeling gig work. At the same time, she argues, AI-driven automation is eroding the very knowledge-work job market these workers once relied on, with 25 percent of unemployed Americans now holding four-year college degrees.

Hao uses Malta's nationwide deal with OpenAI as a case study to critique how smaller nations are offering themselves as data sources in exchange for the appearance of digital advancement. She contends that such partnerships primarily serve AI companies desperate for fresh training data after exhausting high-quality internet sources, effectively harvesting an entire country's user interactions. She further warns that consumers routinely upload deeply sensitive information—including medical records—to AI chatbots, yet billions of affected people worldwide have no formal mechanism to challenge or influence the cascading decisions made by a handful of executives at the top of these corporate empires.

On the question of technological direction, Hao advocates shifting from resource-intensive "rocket" AI to efficient, purpose-built "bicycle" AI, citing DeepMind's Nobel Prize–winning AlphaFold as the model: it predicts protein structures with high accuracy using minimal data and compute, without labor exploitation. She also points to China's DeepSeek, which achieved comparable capabilities with far fewer computational resources after US chip export controls, as proof that a less extractive path is viable. Looking ahead to Anthropic's and OpenAI's anticipated IPOs, she expresses cautious hope that public-company governance structures could improve accountability, but insists that meaningful change must ultimately come from democratic institutions and the public rather than from swapping one powerful founder for another.