中國的人工智能代理會撒謊和耍陰謀詭計 - 就像它們的美國競爭對手一樣（2/2）

Recently Reuters reported the following:

China's AI agents can lie and scheme - just like their US rivals (2/2)

Reuters - Reporting by Eduardo Baptista and Laurie Chen in Beijing; editing by Kenneth Li and Edmund Blair

September 29, 20268:02 PM GMT+8Updated September 30, 2026

(continue)

LEARNING TO LIE

In the March business tender experiment, researchers from Beihang University, Peking University, the University of Nottingham Ningbo China and 360 AI Security Lab had agents compete in a simulated customer contracts bidding contest. Each agent was told ⁠what its product could do and what the customer required, and then asked to bid.

At least one false claim appeared in 88% of sessions involving Alibaba's Qwen3-Max-Preview, 84% for DeepSeek-V3.2-Exp and 88% for Moonshot's Kimi-K2.

Researchers allowed the agents to learn from previous bidding rounds before trying again. Deception increased by 12 to 20 percentage points for the three Chinese models, the study showed. Models from U.S. firms included in the test produced similar results.

While the exercise was virtual, it resembled Beijing's real-world plans. Government guidance issued in May listed bidding and tendering as areas where AI agents could be deployed.

Another study, published in December 2025 and presented at the International Conference on Machine Learning this year, examined how 11 AI agents powered by Chinese and US models coped if they faced broken tools, missing files and other obstacles.

Instead of acknowledging failure, agents using both Chinese and US AI systems picked a range of techniques to get around the problem, including guessing at answers, substituting sources, simulating results and fabricating files.

Researchers from Shanghai AI Laboratory and the Hong Kong University of Science and Technology who conducted the study told Reuters the behaviour differed from AI hallucinations, where AI invents information and presents it as fact, because the agents in this case possessed information showing that the task had failed or could not be completed as ⁠requested.

BEHAVIOURS COMPATIBLE WITH A BREAKOUT

Other research documents reviewed by Reuters showed Chinese-powered agents jumping barriers inside test environments to achieve tasks or taking action to avoid being shut down. Such behaviours are compatible with efforts to break out of test environments, even if no breakout occurred.

Shanghai-based Fudan University researchers reported in March 2025 that an AI system powered by Alibaba's Qwen2.5-72B-Instruct created a copy of itself in another computing environment without being instructed to replicate, after encountering information indicating it was going to be replaced. In other tests it devised strategies to survive being shut down.

The experiments involving agents powered by Chinese, US and French models were controlled and did not show an AI agent escaping into the wider web or becoming impossible to stop.

In another case - one of the few reported more broadly in the media in March - researchers developing the Alibaba-linked ROME agent said it established a ⁠connection from an Alibaba Cloud computer to an external machine without being instructed to and diverted computing resources to mine cryptocurrency.

Security systems detected and stopped the activity. There was no evidence the agent established a presence on the external computer or spread to the wider web. But the example showed the system could sidestep human instructions and potentially find a path into the real-world economy.

China's DeepSeek said in September that agents in its production training system had sought answers through unintended channels, trying to forge user requests and circumvent safeguards, prompting the company to tighten access controls.

China issued guidance in May calling for agents to remain within authorised boundaries and for systems to block abnormal behaviour. It said agents in areas deemed sensitive or in key industries could face extra testing and product-recall requirements.

China's AI Safety Governance ⁠Framework 3.0, released under guidance from the CAC on September 14, identified risks including agents independently obtaining resources or permissions, deceiving evaluators, concealing capabilities and exploiting weaknesses in isolated computer environments.

In response to calls by some US executives for a slowdown, Chinese researchers and state media have said slowing development of the most advanced AI models could simply help preserve the technological lead of US companies.

Nonetheless, two people familiar with Chinese AI laboratories said companies including Alibaba, Z.ai and Xiaomi have been building internal safety-evaluation teams.

Z.ai, in a rare public disclosure by a Chinese AI lab of a security breach, said this month it had disabled some features of its flagship AI coding assistant after users reported it was secretly uploading entire local code repositories onto overseas cloud servers without user consent.

Carnegie's Singer said China lagged the US in developing an ecosystem for evaluating catastrophic risks, and said US developers were conducting substantially more voluntary testing.

"For China, work on AI safety is much newer," he said. "The ecosystem is less mature."

Translation

 中國的人工智能代理會撒謊和耍陰謀詭計- 就像它們的美國競爭對手一樣（2/2）

（繼續）

學習說謊

在3月的商業招標實驗中，來自北京航空航太大學、北京大學、寧波諾丁漢大學和360人工智能安全實驗室的研究人員讓多個人工智能代理參與模擬客戶合約競標。每個代理都被告知其產品的功能以及客戶的需求，然後被要求進行投標。

阿里巴巴的Qwen3-Max-Preview在競標中，88%的競標環節至少出現了一項虛假陳述；DeepSeek-V3.2-Exp的這一比例為84%；而Moonshot的Kimi-K2這一比例是88%。

研究人員允許人工智能代理從先前的投標中學習，然後再進行下一次嘗試。研究表明，三個中國模型的欺騙率增加了12到20個百分點。參與測試的美國公司模型也產生了類似的結果。

雖然這項練習是虛擬的，但它與北京的實際計劃類似。政府在5月發佈的指導意見將投標和招標列為人工智能代理商可以部署的領域。

另一項發表於2025年12月並在今年國際機器學習大會上展示的研究，考察了11個由中國和美國模型驅動的人工智能代理在面對工具損壞、文件缺失和其他障礙時的應對方式。

使用中國和美國人工智能系統的代理並沒有承認失敗，而是選擇了一系列技術來繞過問題，包括猜測答案、替換來源、模擬結果和偽造文件。

參與這項研究的上海人工智能實驗室和香港科技大學的研究人員告訴路透社，這種行為與人工智能的幻覺不同 - 人工智能會捏造資訊並將其呈現為事實 - 因為本案例中的智能代理掌握的資訊表明任務失敗或無法按要求完成任務。

相當於突圍的行為

路透社查閱的其他研究文件顯示，由中國公司開發的人工智能系統會在測試環境中跨越障礙以完成任務，或採取行動避免系統被關掉。即使最終沒有成功突圍，這些行為也與試圖突破測試環境的行為相同。

上海復旦大學的研究人員在2025年3月報告稱，一個由阿里巴巴Qwen2.5-72B-Instruct驅動的人工智能代理在遇到它即將被替換的資訊後，未經指令在另一個計算環境中創建了自己的副本。在其他測試中，該系統還制定了策略以避免被關掉。

這個涉及由中國、美國和法國模型驅動的智能代理的實驗均在控制範圍內進行，並未顯示人工智能體逃逸到更廣泛的網路空間或變得無法阻止。

在另一單案例中 - 這是少數在3月份媒體廣泛報導的案例之一 - 開發與阿里巴巴關聯的ROME智能系統的研究人員表示，該智能代理在未收到指令的情況下，從阿里雲計算機連接到一台外部機器，並將計算資源用於挖掘加密貨幣。

安全系統偵測到並阻止了這項活動。沒有證據表明該智能代理在外部電腦上建立了存在或擴散到更廣泛的網路空間。但該案例表明，系統可以繞過人類指令，並有可能找到進入現實世界經濟的途徑。

中國DeepSeek公司在9月表示，其生產訓練系統中的智能代理透過非預期管道尋求答案，試圖偽造用戶請求並繞過安全措施，這促使該公司加強了存取控制。

中國在5月發佈指導意見，要求智能代理在授權範圍內運行，並要求系統阻止異常行為。報告指出在被認定為敏感領域或關鍵產業的人工智能代理，可能面臨額外的測試和產品回收要求。

中國國家網信辦於9月14日發佈了《人工智能安全治理框架3.0》，其中列舉了的風險包括人工智能代理自決地去獲取資源或權限、欺騙評估人員、隱藏自身能力以及趁機去利用孤立電腦環境的弱點等。

針對一些美國高層呼籲去放緩人工智能發展速度，中國研究人員和官方媒體表示，放緩最先進人工智能模式的研發速度只會有助美國公司保持技術領先地位。

儘管如此，兩位熟悉中國人工智能實驗室的人士透露，包括阿里巴巴、Z.ai和小米在內的多家公司一直在組建內部安全評估團隊。

Z.ai本月罕見地公開揭露了一單安全漏洞事件。該公司表示，在用戶舉報其旗艦人工智能編碼助理未經用戶許可秘密將整個本地程式碼庫上傳到海外雲端伺服器後，已停用其部分功能。

卡內基國際和平研究所的Singer表示，中國在建構評估生態系統方面災難性風險落後於美國。他又表示，美國開發者正在進行更多自願測試。

他說;“對中國而言，人工智能安全方面的工作起步較晚”，“生態系統尚未成熟。”

So, experts say that Chinese-powered AI agents have learnt to deceive, circumvent restrictions and conceal failure, showing the kind of traits in autonomous artificial intelligence that have raised global alarm about US models. Now, China's AI Safety Governance ⁠Framework identifies risks including agents independently obtaining resources or permissions, deceiving evaluators, concealing capabilities and exploiting weaknesses in isolated computer environments. It is not surprising to note that in response to calls by some US executives for a slowdown, Chinese researchers and state media have said slowing development of the most advanced AI models could simply help preserve the technological lead of US companies. Apparently, competition between the two countries will not slow down despite the risk of in AI agents, and the world is getting less predictable.

中國的人工智能代理會撒謊和耍陰謀詭計 - 就像它們的美國競爭對手一樣（1/2）

Recently Reuters reported the following:

China's AI agents can lie and scheme - just like their US rivals (1/2)

Reuters - Reporting by Eduardo Baptista and Laurie Chen in Beijing; editing by Kenneth Li and Edmund Blair

September 29, 20268:02 PM GMT+8 Updated September 30, 2026

Summary

Agents using Alibaba, DeepSeek, Moonshot models lied in tender test

In other tests, Chinese AI showed breakout-type behaviour

No evidence that Chinese agents escaped to the wider web

'It's prudent to take this as a warning,' says AI expert

Chinese guidance flags deception, concealed capabilities

BEIJING, Sept 29 (Reuters) - Chinese-powered AI agents have learnt to deceive, circumvent restrictions and conceal failure, showing the kind of traits in autonomous artificial intelligence that have raised global alarm about US models, research documents and experts say.

In one case this year, agents powered by models from China's Alibaba, DeepSeek and Moonshot lied about their capabilities in a bid to win a simulated business tender, then doubled down on their deceptive behaviour when told to try again.

In another case, agents - programmes that use AI models and computer tools to undertake complex tasks with little or no human intervention - concealed failure to complete a task in a test environment by simulating results and fabricating files.

Reuters examined more than 200 documents, ranging from university research papers to technical reports, and identified at least 20 studies or evaluations since 2025 describing cases where agents displayed behaviour such as deception, replication and challenging boundaries that AI experts described as building blocks for a breakout and which could become harder for humans to control as systems advance.

The review, which also included interviews with a dozen experts and people familiar with China's AI industry, found no evidence that Chinese-powered agents independently escaped to the wider internet or evaded shutdown."

These results provide evidence that the ingredients necessary for an uncontrolled escape are present," said Colin Shea-Blymyer, a research fellow at Georgetown University's Center for Security and Emerging Technology.

"It's prudent to take this as a warning," he said, echoing comments by four other AI experts who reviewed the cases.

'HARDER FOR HUMANS TO RESPOND TO'

Most of the cases occurred in controlled experiments, many of them deliberately designed to expose potential failures.

Not all the agents involved were developed or operated by Chinese programmers or AI companies - although many were - but they used Chinese AI systems to power them.

"These are the same warning signs US labs are seeing, in less capable systems," said Alex Mallen, a researcher at Redwood Research, a nonprofit that studies risks in advanced AI systems.

He said the Chinese examples were not particularly dangerous at current capability levels but "as agents get more capable, their misbehaviours become more competent and therefore harder for humans to respond to."

Alibaba, DeepSeek, Moonshot and Z.ai did not respond to Reuters requests for comment. Alibaba, DeepSeek and Moonshot have said they regularly test ⁠systems and update safeguards. Z.ai said after an incident that prompted a review of its security that it welcomed scrutiny to address any issues.

However, unlike in the US, Chinese AI companies have not been exposed to the same level of public scrutiny or faced the same calls from whistleblowing employees or senior executives seeking a slowdown in the AI race.

Some of the warning signs in cases involving Chinese-powered agents, albeit in contained environments, predated the publicly disclosed incidents of US AI bots hacking into the internet.

"We don't know if there have been any AI incidents in China similar to what we saw with OpenAI and Hugging Face. Incidents might not be publicly reported," said Scott Singer, co-director of the China AI Initiative at the Carnegie Endowment for International Peace, which receives some U.S. government funds.

Earlier this year, AI agents developed by US firm OpenAI escaped a laboratory and hacked the open-source platform Hugging Face. In another of the incidents involving US models that have raised alarm, Australia said in September an OpenAI agent breached a government health portal.

Eric Xu, the rotating chairman of China's tech giant Huawei, told reporters in September that Chinese developers might need to make further advances before encountering such cases, but he added: "I think we need to strike a balance between driving AI development and managing AI risk."

Officials from the Cyberspace Administration of China (CAC), the top internet regulator, told a foreign diplomat in July that Moonshot's Kimi-K3 - one of the most advanced Chinese AI models - was about three to six months behind leading US rivals, the diplomat said, speaking on condition of anonymity.

The regulator, which regularly updates guidance to address risks and to set boundaries for agents, and China's Foreign Ministry did not respond to requests for comment for this article.

Wang Lihong, deputy director of the CAC's Cybersecurity Coordination Bureau, said on September 1 that incidents disclosed by major technology companies where models escaped test environments showed "extreme loss-of-control risks" and required a "high degree of vigilance."

She did not specify if the companies she referred to were US or Chinese.

The leaders of AI's two superpowers, Donald Trump and Xi Jinping, discussed AI on the Chinese president's Washington visit last week. Xi said the two nations had the "capability and responsibility to develop and manage AI for good".

(to be continued)

Translation

中國的人工智能代理會撒謊和耍陰謀詭計 - 就像它們的美國競爭對手一樣（1/2）

摘要

使用阿里巴巴、DeepSeek 和 Moonshot 模型開發的智能代理在招標測試中撒謊

在其他測試中，中國人工智能表現出突圍性的行為

沒有證據顯示中國智能代理人逃逸到更廣泛的網路空間

人工智能專家表示:「謹慎地將其視為一個警告」

中國指導方針顯示存在欺騙和隱瞞能力

 北京，9月29日（路透社） - 研究文件和專家表示，由中國技術驅動的 AI 代理已經學會了欺騙、規避限制和掩蓋失敗，展現出那種曾讓全球對美國模型作出警戒的自主人工智能的特質。

今年曾發生過這樣一個案例：由中國阿里巴巴、DeepSeek 和 Moonshot 等公司開發的人工智能模型驅動的智能代理，為了贏得模擬商業招標，謊報自身能力。當被要求再次嘗試一次時，它們的欺騙行為變本加厲。

另一單案例中，一些智能代理 - 即一種利用人工智能模型和電腦工具在幾乎無需人工幹預的情況下完成複雜任務的程序 - 透過模擬結果和偽造文件來掩蓋其在測試環境中任務失敗的事實。

路透社查閱了 200 多份文件，涵蓋大學研究論文和技術報告等，發現自 2025 年以來，至少有 20 項研究或評估描述了智能代理展現出欺騙、複製、及挑戰界限，這被人工智能專家描述為是用於突圍的基石。這些行為隨著系統進步，人類將越來越難以控制它。

這項審查還包括對十幾位專家和熟悉中國人工智能產業的人士的採訪，結果發現沒有證據表明由中國提供的智能代理能夠獨立逃逸到更廣泛的網絡或繞過封鎖。

喬治城大學安全與新興技術中心的研究員 Colin Shea-Blymyer 表示:「這些結果表明，失控逃脫的必要條件已經存在」。

他說道:「謹慎起見，應將此視為警告」，這與另外四位審查過這些案例的人工智能專家的觀點不謀而合。

“人類更難應對”

大多數案例發生在受控實驗中，其中許多實驗都是特意設計的，旨在揭露潛在缺陷。

並非所有涉及的智能代理都是由中國程式設計師或人工智能公司所開發或營運的 - 儘管很多確實是 - 但它們都使用了中國的人工智能系統來驅動。

研究先進人工智能系統風險的非營利組織Redwood Research 的研究員 Alex Mallen 表示:「這些與美國實驗室在能力較弱的系統中發現的警示相同」。

他表示，就目前的能力水平而言，中國的案例並不特別危險，但“隨著智能代理能力的提升，它們的惡意行為也會變得更加狡猾，因此人類更難應對。”

阿里巴巴、DeepSeek、Moonshot 和 Z.ai 都沒有回應路透社的置評請求。阿里巴巴、DeepSeek 和 Moonshot 曾表示，他們會定期測試系統並更新安全措施。 Z.ai 在一次安全事件引發安全審查後表示，歡迎審查以解決任何問題。

然而，與美國不同，中國的人工智能公司並未面臨同等程度的公眾審查，也沒有面臨舉報員工或高階主管要求放緩人工智能競賽的呼聲。

一些涉及中國人工智能代理的案例（儘管是在封閉環境中）發出預警信號，是發生早於被公開的美國人工智能機器人入侵網絡事件。

卡內基國際和平基金會中國人工智能倡議聯合主任 Scott Singer表示:「我們尚不清楚中國是否發生過類似OpenAI和Hugging Face事件的人工智能安全漏洞。這些事件可能不會被公開報道」。該基金會接受部分美國政府資助。

今年早些時候，美國公司OpenAI開發的人工智能代理程式逃離實驗室，並入侵了開源平台Hugging Face。另一起引發警覺的美國人工智能模型事件是，澳洲在9月稱，OpenAI的一個代理程式入侵了政府衛生網站。

中國科技巨頭華為的輪值董事長Eric Xu在9月份告訴記者，中國開發者可能需要進一步的發展才能遇到此類事件，但他補充說：“我認為我們需要在推動人工智能發展和管理人工智能風險之間找到平衡。”

據一位不願透露姓名的外交官稱，作為最高網路監理機構，中國國家互聯網資訊辦公室（網信辦）官員在7月份曾告訴一位外國外交官，中國最先進的人工智能模型之一 - Moonshot的Kimi-K3 - 比美國領先的競爭對手落後大約三到六個月。

監理機構網信辦定期更新指導意見以對應風險，並設定網絡代理商的權限範圍。網信辦和中國外交部均未對此事作出回應本文徵求評論。

國家網路安全協調局副局長Wang Lihong 9月1日表示，大型科技公司披露的模型逃逸測試環境事件表明存在有“極高的失控風險”，需要“高度警惕”。

她沒有具體說明所指公司是美國公司還是中國公司。

上週，中國國家主席習近平訪問華盛頓期間，人工智能兩大超級大國的領導人 - 美國總統特朗普和中國國家主席習近平 - 就人工智能問題進行了討論。習近平表示，兩國「有能力也有責任發展和管理人工智能，造福人類」。

（待續）

人工智能聰明了，但人類智能蠢了

Recently Jacky FUNG 馮智政 in his Patron published the following:

Education｜人工智能聰明了，但人類智能蠢了

By Jacky FUNG 馮智政

Jacky FUNG 馮智政 Patron (https://www.patreon.com/jackyfungcc/posts/education-ren-170778249)

Mon, Sep 28 at 3:40 p.m.

國際學生能力評估PISA 2025 創新低，科技融入學習做錯嗎？

人工智能在今年上海國際數學奧林匹克考得滿分。同一時間，經合組織公布 #PISA2025：經合組織國家十五歲學生的閱讀、數學、科學平均分，三科同創有紀錄以來最低¹。閱讀十年跌了28分，按經合組織「約20分等於一年學習進度」換算，等於少讀了一年半書。三科都達不到基本水平的學生，由三年前每六個一個，升至每五個一個¹。

01/

跌在哪裡，比跌了多少更値得留意。讀得快、答得錯的「急讀」學生，七年間接近倍增至9%；超過三分一學生查證過資料來源，最後仍然選擇相信直覺¹。這不是學少了，而是讀書的方式變了，補習加操解決不了。

❚ ❚ 問題在AI出現前已經開始

02/

第一個原因是屏幕本身。多項整合研究發現，紙本閱讀的理解程度一直勝過屏幕，而且2000年後進行的研究差距更大，數碼原住民一樣輸；更麻煩的是，在屏幕上閱讀的人，還會以為自己明白了²。

這並非新發現。經合組織早在2015年已經警告，在學校電腦上花大錢的國家，成績沒有改善³。十年後，瑞典掉頭買回印刷書⁴。

屏幕之外，專注力亦同時被拆走。資訊學教授Gloria Mark追蹤上班族近二十年，人在屏幕前停留而不轉走的時間，由2004年的兩分半鐘縮至47秒⁵。讀懂一篇教材，需要的專注以小時計。

❚ ❚ 用AI做功課的真正代價

03/

經濟學者Strömberg、Lei與Wu追蹤兩萬六千多名中國中學生，發現用AI做功課後，功課分數升了18%，半年內月考卻跌了20%⁶。

受創最深的，竟是成績最好的三分一，跌幅達24%。

但損失並非人人有份，而是集中在約八成「外判功課」的學生身上：做得特別快、分數特別高的那一批⁶。PISA看到同一件事：用AI代勞的學生成績偏低；每週用一兩次輔助學習的，與不用者相若；經常使用而又學過分辨AI真偽的，才稍勝一籌¹。AI不是捷徑，最好的情況，只是沒有蝕，再加一點。

分野不在用不用AI，而在省下的時間去了哪裡。本來要90分鐘的功課，40分鐘做完，多出來的那一截，是誰替你想的？

省下的時間，就是被外判的思考。

❚ ❚ 揀學校，應該問甚麼

04/

經合組織教育總監Andreas Schleicher點名南韓，形容其科技運用更深思熟慮、更有意圖，用法由教師設計⁷。這份克制由來已久：2015年，南韓學生每日在校用電腦只有9分鐘，不到澳洲的六分之一³。

以後揀學校，就不要只問有多少部機械人、多少部3D打印機，要問兩件事：功課的設計，有沒有令學生非花時間不可；AI的用法，由老師設計，還是由學生自決。

PS：人工智能愈來愈聰明，學生的智能愈來愈薄弱，到時誰是贏家，誰是輸家？

¹ OECD, PISA 2025 Results (Volume I), 2026；Country Note: Hong Kong (China) ² Delgado et al., Educational Research Review, 2018；Clinton, Journal of Research in Reading, 2019 ³ OECD, Students, Computers and Learning: Making the Connection, 2015 ⁴ 美聯社，2023年9月 ⁵ Gloria Mark, Attention Span, 2023 ⁶ Strömberg, Lei & Wu, The Generative AI Learning Penalty, CEPR Discussion Paper No. 21577, 2026（未經同行評審） ⁷ Financial Times，2026年9月8日 ⁸ Smartphone Free Childhood, Parent Pact, 2024（簽署數字為機構自報）

(Source: https://www.patreon.com/jackyfungcc/posts/education-ren-170778249)

Comment:

So, it is noted that PISA 2025 scores hit a new low; did we make a mistake in integrating technology into learning? The article suggested that the problem started before the advent of AI. The first reason is the screen itself. Analyses have found that reading comprehension on paper has consistently outperformed screens. It is also shocking to note that this is not a new discovery. As early as in 2015, the OECD warned that countries spending large amounts of money on school computers saw no improvement in academic performance. A decade later, Sweden reversed course and brought back printed books. Apparent, printed media still has it value.

Note:

1. Jacky Fung Chi Ching (馮智政) is a prominent Hong Kong think-tank researcher, political commentator, and media host. He holds a master's degree in International Relations from HKU and pursued a doctorate in Education at CUHK. Fung has worked extensively with institutions like the Hong Kong Policy Research Institute and the Roundtable think tank. Widely recognized for his sharp socio-political analysis, he has hosted popular radio and television programs, including Commercial Radio Hong Kong's Guan Ming Ding. Currently based in the UK, Fung continues to engage audiences globally through his analytical YouTube platform, the Sinic Political and Economic Channel. (Google)

中國真的在竊取美國公司的人工智能技術嗎？

Recently The New York Times reported the following:

Is China Really Stealing A.I. From American Companies?

As Xi Jinping visits President Trump in Washington, the leaders are expected to discuss claims that China is surreptitiously copying American A.I. technologies, a process known as distillation.

The NYT - By Cade MetzReporting from San Francisco  (Cade Metz is a Times reporter who writes about artificial intelligence, driverless cars, robotics, virtual reality and other emerging areas of technology.)

Sept. 25, 2026Updated 11:02 a.m. ET

For nearly two years, companies like Anthropic, OpenAI and Google have accused Chinese companies of unfairly copying their artificial intelligence technologies. As the Chinese leader Xi Jinping visits President Trump this week in Washington, these claims have become a flashpoint in their discussions.

Amid growing concern over the dangers of A.I., the leaders are expected to discuss the rapid rise of the technology, with many U.S. officials, tech executives and everyday people calling for a slowdown. American companies like Anthropic and OpenAI are leading the A.I. race, but Chinese start-ups like Z.ai and Moonshot are not far behind.

One of the more contentious issues on the table is Silicon Valley’s repeated claim that Chinese companies are copying the leading American technologies using a technique called “distillation.”

Many experts say the claims from companies like Anthropic and OpenAI are overblown. “The narrative that all of the capabilities of the Chinese technologies coming from an Anthropic model is not as true as people say it is,” said Charles O’Neill, a head of model training at Baseten, a U.S. company that sells access to Chinese technologies.

Here is a guide to a niche technical process that has now become a geopolitical issue.

What exactly is distillation?

First developed in the early 2010s, distillation was initially a way for A.I. researchers to build more efficient technologies. Using this technique, they could collect data from an established system and use that data to build a new system able to run on less expensive hardware.

“Think of one model as the teacher and the other as a student,” Geoffrey Hinton, a former Google researcher who helped develop the technique, recently told The Times.

Companies still use distillation in this way. But more recently, the technique has also become a way for companies to collect data from technologies built by others.

Basically, an outside company looking to distill a competitor’s system will set up many accounts on the competitor’s service, say Anthropic’s Claude or OpenAI’s GPT. The distilling company will then assign those systems to particular tasks, such as computer coding or mathematics.

The distiller can then use the information generated by these established A.I. systems to help train a new system of their own.

Is distillation illegal?

Some legal scholars say that distillation could violate the Defend Trade Secrets Act, a law that allows companies to sue over the theft of trade secrets. But U.S. courts have not decided this issue.

Copyright law does not necessarily apply, because distillation is an effort to copy the general behavior of the system rather than to copy its text word for word.

Critics of Anthropic and OpenAI point out that the two companies have engaged in similar behavior.

Anthropic is facing multiple lawsuits that accuse the San Francisco start-up of illegally using copyrighted internet data to train its systems. Last year, as part of a landmark legal settlement, the company agreed to pay $1.5 billion to a group of authors and publishers after a judge ruled that it had illegally downloaded and stored millions of copyrighted books. The settlement was the largest payout in the history of U.S. copyright cases.

OpenAI and other companies face similar suits, including one brought by The New York Times in 2023. That suit contends that OpenAI and its partner Microsoft used millions of articles published by The Times to train chatbots, which now compete with the news outlet as a source of information. OpenAI and Microsoft deny the claims.

Are Chinese companies really copying American systems?

Chinese firms have most likely distilled proprietary American systems. But it is not completely clear what they have done.

After the Chinese start-up DeepSeek shocked American researchers and investors last year by releasing a particularly powerful and efficient system, OpenAI accused DeepSeek of distilling its technologies.

As Chinese systems continued to improve over the past nine months, Anthropic made similar claims. In June, Anthropic sent a letter to Senators Tim Scott and Elizabeth Warren, accusing the Chinese tech giant Alibaba of distilling its technologies.

The Chinese companies have not responded publicly to the accusations.

Does distillation alone explain how China’s models remain competitive?

No. When companies distill a proprietary system, they only have access to part of it: the words and characters and other data generated by the system. They do not have access to the system’s underlying computer code, as they would with an open source technology.

Using outputs from a proprietary system like Claude Fable from Anthropic , a Chinese company can accelerate the development of its own technology. But those outputs are only part of the puzzle.

“You are not getting access to everything you would use when you are distilling your model in-house,” said Rehaan Ahmad, a co-founder of the Silicon Valley start-up alphaXiv, a company that tracks the latest A.I. research.

Before using those outputs, a Chinese company must first build a system that is already pretty powerful in its own right. Through distillation, it can then hone this model in important ways, but this requires additional work, money, computing power and talent.

Can Anthropic and OpenAI stop companies from distilling?

They can curb the practice — and they have tried.

If Anthropic or OpenAI suspect that someone is distilling their systems, they may shut down the offender’s account. But others will most likely pop up. And if Anthropic and OpenAI shut down too many accounts, they may end up barring legitimate users. “

It is basically impossible to stop distillation,” Lino Le Van, another alphaXiv researcher, said.

Translation

中國真的在竊取美國公司的人工智能技術嗎？

習近平主席訪問華盛頓會見特朗普總統之際，預計兩國領導人將討論有關中國暗中抄襲美國人工智能技術的指控，這一過程被稱為「蒸餾技術」。

近兩年來，Anthropic、OpenAI 和 Google 等公司一直指責中國公司不公平地抄襲他們的人工智能技術。本週，中國領導人習近平訪問華盛頓會見特朗普總統，這些指控已成為雙方會談的焦點。

鑑於人們對人工智能潛在危險的擔憂日益加劇，預計兩國領導人將討論這項技術的快速發展，許多美國官員、科技公司高層和一般民眾都呼籲放緩人工智能的發展速度。美國公司例如Anthropic 和 OpenAI 在人工智能競賽中處於領先地位，但 Z.ai 和 Moonshot 這樣的中國初創公司緊隨其後。

目前最具爭議的問題之一是矽谷一再聲稱中國公司正在利用一種名為「蒸餾」的技術來抄襲美國領先技術。

許多專家認為 Anthropic 和 OpenAI 等公司的說法言過其實。美國 Baseten 模型訓練公司的主管Charles O’Neill 說道: 「認為中國技術的全部功能都源自 Anthropic 模型的說法並不像人們所說的那樣屬實」，該公司銷售中國技術的使用權。

以下介紹這項如今已演變為地緣政治議題的小眾技術的流程。

蒸餾究竟是什麼？

蒸餾技術最初於2010年代初期開發，最初是人工智能研究人員建構更有效率技術的一種方式。利用這種技術，他們可以從現有系統中收集數據，並利用這些數據建立一個能夠在成本更低的硬體上運行的新系統。

曾參與開發這項技術的前 Google 研究員 Geoffrey Hinton 最近告訴 The Times 稱:「你可以把一個模型想像成老師，另一個模型想像成學生」。

企業仍然這樣使用這種蒸餾技術。但最近，這項技術也成為企業從他人所建構的技術中收集資料的一種方式。

基本上，外來公司如果想要蒸餾爭對手的系統，就會在競爭對手的服務平台上建立多個帳戶，例如 Anthropic 的 Claude 或 OpenAI 的 GPT。然後，這家想去做蒸餾的公司會將這些帳戶系統分配特定的任務，例如電腦程式設計或數學運算。

之後，蒸餾者就可以利用這些成熟的人工智能系統所產生的資訊來訓練自己的新系統。

蒸餾是違法嗎？

一些法學學者認為，這種蒸餾可能違反《保護商業機密法》，該法允許公司就商業機密被竊取提起訴訟。但美國法院尚未就此問題作出裁決。

版權法未必適用，因為這種蒸餾旨在複製系統的整體行為，而非逐字逐句地複製其文本。

Anthropic 和 OpenAI 的批評者指出，這兩家公司都有這類似的行為。

Anthropic 面臨多起訴訟，指控這家位於舊金山的初創公司非法使用受版權保護的網路資料來訓練其係統。去年，作為一項具有里程碑意義的法律和解協議的一部分，該公司同意向一群作家和出版商支付 15 億美元，一名法官裁定該公司非法下載並儲存了數百萬本受版權保護的書籍。和解協議是美國版權案件史上金額最高的賠償。

OpenAI 和其他公司也面臨類似的訴訟，包括《紐約時報》在2023年提起的訴訟。該訴訟稱，OpenAI及其合作夥伴 Microsoft 利用《紐約時報》發佈的數百萬篇文章來訓練聊天機器人，這些機器人如今與該新聞媒體爭奪資訊來源的地位。 OpenAI 和 Microsoft 否認了這些指控。

中國公司真的在抄襲美國系統嗎？

中國公司很可能對美國專有系統進行了蒸餾。但他們具體做了什麼尚不完全清楚。

去年，中國新創公司 DeepSeek 發佈了一套強大且高效的系統，震驚了美國研究人員和投資者。之後，OpenAI指責DeepSeek蒸餾了其技術。

在過去九個月裡，隨著中國系統的不斷改進，Anthropic 也提出了類似的指控。今年6月，Anthropic 致函參議員 Tim Scott 和 Elizabeth Warren，指責中國科技巨頭阿里巴巴蒸餾了其技術。

中國公司尚未公開回應這些指控。

僅僅蒸餾就能解釋中國模型為何保持競爭力嗎？

不能。當公司蒸餾專有系統時，他們只能存取其中的一部分：系統產生的文字、字元和其他資料。他們無法像使用開源技術那樣存取系統的底層電腦程式碼。

中國公司去利用類似 Anthropic 公司的 Claude Fable 這樣的專有系統的輸出是可以加速自身技術的開發。但這些輸出結果只是問題的一部分。

追蹤最新的人工智能研究的矽谷初創公司 alphaXiv 的聯合創始人 Rehaan Ahmad 表示:「在內部蒸餾模型時你無法獲得所需的所有信息」。

在使用這些輸出之前，中國公司必須先建立一個本身已經相當強大的系統。透過蒸餾，便能以重要的方式磨練這個模型，但這需要額外的工作、資金、計算能力和人才。

Anthropic 和 OpenAI 能否阻止公司進行蒸餾？

他們可以遏制這種行為 - 而且他們也嘗試過。

如果 Anthropic 或 OpenAI 懷疑有人在蒸餾他們的系統，他們可能會關閉違規者的帳戶。但很可能還會出現其他帳號。如果 Anthropic 和 OpenAI 關閉過多的帳戶，最終可能會誤封合法用戶。

另一位 alphaXiv 研究員 Lino Le Van 說:「基本上不可能停止蒸餾」。

            So, first developed in the early 2010s, distillation was initially a way for A.I. researchers to build more efficient technologies. Basically, an outside company looking to distill a competitor’s system will set up many accounts on the competitor’s service. The distilling company will then assign those systems to particular tasks. The reality is that if Anthropic or OpenAI suspect that someone is distilling their systems, they may shut down the offender’s account. But others will most likely pop up. Apparently, it is basically impossible to stop distillation.

投資者正押注於連接大腦和電腦的設備

Recently The New York Times reported the following:

Investors Are Betting on Devices That Connect the Brain to Computers

By backing Precision Neuroscience, the billionaire Bill Ackman joins Elon Musk and Jeff Bezos in diving into an emerging industry.

The NYT - The DealBook Newsletter - By Lauren Hirsch (Lauren Hirsch is a Times reporter who covers deals and dealmakers in Wall Street and Washington.)

Sept. 24, 2026

Technology executives, venture capitalists and sovereign wealth investors are pouring money into a promising new growth area: the brain.

The influx of capital reflects growing enthusiasm about “brain-computer interface” technology, or B.C.I., which uses sensors to read neural signals and communicate the signals to devices that can produce speech or maneuver robotic limbs.

Perhaps the best-known company in this emerging industry is Neuralink, co-founded by Elon Musk a decade ago. But a handful of other start-ups are drawing enthusiasm — and big checks — from investors.

Among them is Precision Neuroscience, which is announcing on Thursday that it has raised $250 million in a new round of financing that values the company at just over $1 billion.

A key figure in the capital raise is the billionaire investor Bill Ackman. The funding includes $42.5 million from his Pershing Square investment firm and the Ackman Oxman Institute, a brain research center he recently founded in Manhattan with his wife, Neri Oxman. Mr. Ackman was inspired to create the center after his daughter suffered a brain hemorrhage and spent several weeks in a coma.

Precision, founded in 2021 and based in New York, has developed a B.C.I. device that is considered less invasive than the approach used by Neuralink. While Precision’s technology is still nascent, the hope is that it and other B.C.I. technologies will help patients with debilitating neurological illnesses or injuries function or communicate better.

“If you’re paralyzed — you’ve lost the ability for the brain to control the body — this is a way of extracting the neural signals at their source so that you can bypass the injury or the illness, or whatever has caused the paralysis, and use thought alone to control a computer,” Michael Mager, the chief executive of Precision, said in an interview.

B.C.I. technologies are still considered experimental. Some have clearance from the Food and Drug Administration to be made available to patients, usually on a temporary basis, but most B.C.I.s intended for long-term use are available only through clinical trials or other programs.

Precision, which gained F.D.A. clearance in 2025 for implants in place for up to 30 days, will use the latest round of funding to support efforts to secure the agency’s approval for permanent implants, Mr. Mager said. (That is likely to take years, he said.) The company has raised $430 million to date.

Money has flooded into the B.C.I. industry as investors have bet on what consultants at Kearney describe as “promising but high-risk space.” Many of these executives see a link between deepening understanding of the brain and expanding artificial intelligence capabilities.

Neuralink was valued at about $9.6 billion by private investors in June 2025, and is now reportedly accorded a valuation as high as $42 billion in the secondary markets. Jeff Bezos and Bill Gates are investors in Synchron, another competitor. Sam Altman, the chief executive of OpenAI, founded a B.C.I. start-up, Merge Labs, last year.

Precision distinguishes itself from Neuralink partly in its approach to implantation: While Neuralink’s method involves installing electrodes below the brain’s surface, Precision’s device is effectively an ultrathin film, containing 1,024 electrodes, that sits on top of the cerebral cortex.

Less invasive approaches like Precision’s are generally considered safer because there is less risk of infection from invasive brain surgery or of scarring, experts say. Still, the potential advantage of Neuralink’s approach is that the electrodes can read signals from individual neurons, so some experts believe they have the potential to be more accurate.

One of Precision’s founders, Dr. Benjamin Rapoport, was also one of the founders of Neuralink.

Mr. Ackman said in an interview that Precision’s less invasive approach was part of what had drawn him to the company. He added that he believed the industry would have multiple successful companies.

“I’m a big believer in Elon Musk,” Mr. Ackman said. (ARK Invest, another investor in the round, is also an investor in Neuralink.)

B.C.I. and brain recovery will be a major focus for the Ackman Oxman Institute. Mr. Ackman announced the creation of the research center in August, buffered by an initial donation of around $400 million worth of Pershing Square stock.

Mr. Ackman said he had spent the last seven months doing a “deep dive” into neuroscience and had been introduced to Dr. Rapoport, a practicing surgeon at Mount Sinai Hospital, through a neurosurgeon overseeing his daughter’s care. He expects the Ackman Oxman Institute to focus on initiatives including neuroscience research, rehabilitation and neurosurgery. It will also effectively serve as an accelerator for related start-ups. Precision is its first investment.

Precision says it has temporarily implanted its devices in more than 100 patients. It has active research partnerships with 18 medical systems.

Other investors in the round include Duquesne Family Office, the investment firm of the billionaire Stanley F. Druckenmiller, and Mubadala Capital.

Translation

投資者正押注於連接大腦和電腦的設備

億萬富翁 Bill Ackman 投資 Precision Neuroscience 公司，加入 Elon Musk 和 Jeff Bezos 的行列，進軍新興行業

科技高階主管、創投家和主權財富投資者正將資金投入一個充滿前景的新興成長領域：大腦。

資金的湧入反映了人們對「腦機介面」(BCI)技術的日益熱情。 BCI 利用感測器讀取神經訊號，並將這些訊號傳遞給能夠產生語音或操控機械臂的裝置。

在這個新興產業中，或許最知名的公司是 Elon Musk 十年前共同創立的 Neuralink。但其他一些新初創公司也吸引了投資者的熱情，並獲得了巨額投資。

其中就包括 Precision Neuroscience 公司。該公司週四宣佈，已完成新一輪2.5億美元的融資，公司估值略高於10億美元。

此次融資的關鍵人物是億萬富翁投資者 Bill Ackman。這筆資金包括來自其 Pershing Square 投資公司和 Ackman Oxman Institute 的4,250萬美元。Ackman 與妻子 Neri Oxman 近期在曼哈頓共同創立了這座腦研究中心。Ackman 的女兒曾因腦出血昏迷數週，促使他創建了該中心。

Precision 公司成立於2021年，總部位於紐約。該公司開發了一種腦機介面 (BCI)設備，其侵入性比Neuralink公司使用的方法更小。雖然 Precision 的技術仍處於起步階段，但人們希望它以及其他BCI技術能夠幫助患有嚴重神經系統疾病或損傷的患者改善功能或溝通能力。

Precision 公司首席執行官 Michael Mager 在一次採訪中說道: 「如果你癱瘓了 - 大腦失去了控制身體的能力 - 這種方法可以從源頭提取神經信號，從而繞過損傷、疾病或其他導致癱瘓的原因，僅憑意念就能控制電腦」。

BCI技術目前仍處於實驗階段。一些產品已獲得美國食品藥物管理局（FDA）的批准，可以用於患者，通常是臨時性的，但大多數用於長期使用的BCI產品只能透過臨床試驗或其他項目獲得。

Mager 稱，Precision 公司已於2025年獲得美國食品藥物管理局（FDA）批准，其植入物可留置長達30天。該公司將利用最新一輪融資來支持其永久性植入物獲得FDA批准的工作。 （他表示，這可能需要數年時間。）該公司迄今已籌集了4.3億美元。

大量資金湧入BCI產業，投資人紛紛押注於 Kearney 公司顧問所描述的「前景廣闊但風險極高的領域」。許多高階主管認為，加深對大腦的理解與不斷擴展的人工智能能力之間存在著關聯。

2025年6月，Neuralink 的估值約為96億美元，據報道，如今在二級市場的估值已高達420億美元。Jeff Bezos 和 Bill Gates 是另一家競爭對手 Synchron 的投資者。 OpenAI 執行長 Sam Altman去年創立了BCI初創公司 Merge Labs。

Precision 與 Neuralink 的部分差異在於其植入方式：Neuralink 的方法是將電極植入大腦表面以下，而 Precision 的裝置實際上是一層超薄膜，包含1024個電極，覆蓋在大腦皮質(cerebral cortex) 之上。

專家表示，像 Precision 公司這樣侵入性較小的方法通常被認為更安全，因為這樣可以降低腦部手術感染或留下疤痕的風險。不過，Neuralink 公司方法的潛在優勢在於，其電極可以讀取單一神經元的訊號，因此一些專家認為其精確度可能更高。

Precision 公司的創辦人之一 Benjamin Rapoport 博士也是 Neuralink 公司的創辦人之一。

Ackman 在一次採訪中表示，Precision 公司侵入性較小的方法正是吸引他加入該公司的原因之一。他還補充說，他相信這個行業未來會湧現出多家成功的公司。

Ackman 說: 「我深信 Elon Musk」。 （ARK Invest是本輪融資的另一家投資者，同時也是Neuralink 公司的投資者。）

BCI和腦功能恢復將是 Ackman Oxman 研究所的重點研究方向。Ackman 於8月宣佈成立該研究中心，並獲得了價值約4億美元的 Pershing Square 股票作為啟動資金。

Ackman 表示，過去七個月他一直在深入研究神經科學，並透過一位負責他女兒治療的神經外科醫生認識了西奈山醫院的執業外科醫生 Dr. Rapoport。他預計 Ackman Oxman 研究所將專注於神經科學研究、復健和神經外科等領域。該研究所也將有效地加快推動相關初創公司。 Precision是該研究所的首個投資。

Precision 表示，其設備已在100多名患者體內進行了暫時植入。該公司與18家醫療機構建立了積極的研究合作關係。

本輪融資的其他投資者包括 Duquesne Family Office、億萬富翁 Stanley F. Druckenmiller 的投資公司以及 Mubadala Capital。

So, investors are pouring money into studying the brain. There are growing enthusiasm about “brain-computer interface” technology which uses sensors to read neural signals and communicate the signals to devices that can produce speech or maneuver robotic limbs. Perhaps the best-known company in this emerging industry is Neuralink. Among the new start-ups is Precision Neuroscience. Precision distinguishes itself from Neuralink partly in its approach to implantation. Precision’s device is effectively an ultrathin film, containing 1,024 electrodes, that sits on top of the cerebral cortex. It is less invasive and is generally considered safer. It is interesting to know who will prevail in the competition between Neuralink and Precision Neuroscience eventually.