當前，人工智能技術的快速迭代發展，已成為經濟發展的新引擎和國家競爭的新焦點。在促進技術快速發展的同時，如何制定合理有效的治理規則以防範人工智能帶來的潛在風險，已成為國際社會的重要關切。中美作為人工智能技術和應用發展的兩個大國，在人工智能治理上的競爭與合作受到各界高度廣注。今年以來，DeepSeek的問世提振了中國企業突破美國技術封鎖實現技術創新的信心，另一方面，“特朗普2.0”開啟，美新政府對待人工智能治理以及對華科技競爭上的態度和政策尚未完全明晰。

近日，《紐約時報》刊登弗里德曼先生的文章，呼籲中美兩國合作，共同應對通用人工智能可能給人類帶來的風險。他認為，正如1970年代以來美蘇核武器控制之於世界穩定的意義，未來確保人類有效掌控這些快速進化的人工智能系統，將取決於美中兩國在人工智能領域的合作程度——這關乎明日世界的存續根基。通用人工智能（A.G.I.）誕生後，若不能達成共同的信任標準，美中將無法開展任何合作。

在人工智能治理中，“信任”一詞常常被提及，包括對機器的信任，對人機協同的信任，對使用系統的人的信任，對機構的信任，對國家的信任等等。信任是最本的要素，也是最寶貴的資源。以色列歷史學家尤瓦爾·赫拉利（Yuval Noah Harari）近日在中國發展論壇上的發言中從人類歷史的宏觀視角解讀信任：“我們應當先增進人類之間的信任，再去開發真正意義上的超級人工智能體。……從本質上講，信任是生存的必要條件。以人類身體為例，我們每秒鐘都在呼吸，每一次呼吸都是一種信任——信任外界的空氣，我們將氧氣吸入肺中、融入身體，再將廢氣呼出到外界。這種信任和呼吸是生命的象徵，如果失去對外部世界的信任，停止呼吸，生命也將終結。……歷史告訴我們，信任比恐懼更為重要。為什麼地球是由人類主宰？而不是被大象或者是其他的動物？並非是因為人類比其他動物更聰明，而是因為我們更懂得如何與陌生人建立信任並展開合作，這種能力經過了數千年的錘鍊。在AI時代，想要生存和繁榮，我們需要更加信任他人，這一點比信任AI更為重要”。赫拉利同時也一針見血指出，“當前，人類之間的信任在瓦解，國際機制的基本範式和規則遭到破壞”。筆者最近出席聯合國裁軍署舉辦的首屆全球人工智能峰會，會上，嘉賓們都無一例外提到地緣政治競爭破壞國與國之間的信任，延宕甚至破壞人工智能國際治理的進程。

然而，什麼是信任？信任到底有什麼標準？劍橋詞典對Trust一詞的解釋為：相信某人是善良和誠實的，不會傷害自己，或者相信某個東西是安全、可靠的（to believe that someone is good and honest and will not harm you, or that something is safe and reliable）。韋氏詞典的解釋是，對某人或某物的品質、能力、實力或真相的放心依賴（assured reliance on the character, ability, strength, or truth of someone or something）。現代漢語詞典對於信任一詞的解釋為“相信而敢於託付”。新華字詞中對於“信任”一詞中“信”字的解釋為不懷疑，認為可靠，“任”字的解釋為相信，信賴。在國際關係領域，信任是指國家或行為體對彼此意圖、承諾和行動的正面預期。綜上所述，信任需要包括以下基本要素：可相信、可預期、可解釋。

中美兩國之間能像弗里德曼先生期待的那樣，至少建立基本的信任，在人工智能治理領域開展迫在眉睫的合作？要回答這個問題，不僅要從中美關係現狀來分析，也要考慮人工智能技術本身的特性。

從中美關係來看，當前中美關係存在嚴重的信任赤字。中美兩國之間相互懷疑和恐懼，對彼此意圖的不確定一直影響兩國關係走向。21世紀初，中國經濟和軍事力量的快速增長引發美國國內關於如何應對中國崛起的廣泛辯論，美國國內學術和政策界開始關於中國的大辯論。中國積極推進“一帶一路”倡議、發佈《中國製造2025》、大幅提升對高科技產業投資等，使得美國政策界對中國的態度轉向懷疑和警惕，擔心中國挑戰美國的領導地位。2017年12月，特朗普政府在《國家安全戰略》報告中將中國定位為競爭者，2021年3月，拜登政府發佈《臨時國家安全戰略方針》強調，“中國是唯一一個有潛力綜合運用經濟、外交、軍事以及科技力量來對穩定與開放的國際機制發起持續挑戰的競爭者”。美國《2022年國防戰略》宣稱中國是“美國最重要的戰略競爭者和緊迫的挑戰者”，將中國置於美國的對立面，急劇壓縮兩國對話合作的領域。多位美國官員曾公開表示，要確保對華技術競爭的絕對優勢。為實現這一目標，美國泛化“國家安全”概念，通過設立對外投資審查機制，加強對關鍵技術和設備的出口管制，限制對華技術轉讓，將中國企業列入出口管制“實體清單”等方式，極盡其能保護自身的技術優勢和所謂“國家安全”。另一方面，中國則認為，美方做法實為限制中國的發展，目的是維護自身霸權，剝奪中國正當發展的權利。此外，美國在特朗普第一任期的後期，開始突出對華戰略中的意識形態因素。2020年5月20日，特朗普政府發佈《美國對中華人民共和國的戰略方針》，將中美兩國的競爭上升到意識形態層面，認為“中美意識形態競爭體現在政治、經濟和社會等所有領域”。此後，特朗普政府的多位重要官員就美中關係發表演講，攻擊中國的社會制度，指責中國的內外政策。在中美博弈和地緣政治大背景下，中美兩大行為體的相互懷疑和恐懼，對彼此意圖的不確定性愈演愈烈，構成兩國信任赤字的最大因素，使“可相信”這一目標更加難以企及。而“特朗普2.0”初期在對外政策上的取向，使得“可預期”也成為兩國關係的“稀缺資源”。

從人工智能技術本身來看，雖然近年來獲得突飛猛進的發展，但仍然存在一些尚未解決的問題，包括不確定性、算法黑箱、不可解釋性等。有AI“教父”之稱的約書亞·本希奧教授於2025年1月領銜發佈《國際人工智能安全報告》，其中列出通用人工智能（General-purpose AI）可能存在可靠性問題，包括虛構或幻覺——即不準確或具有誤導性的內容； 常識推理與推斷失敗——無法正確執行符合常識的邏輯推理；未能反映情境相關、最新且無偏見的知識與理解——輸出內容脫離實際背景、過時或帶有偏見。報告指出，可靠性問題的具體表現可能引發多種風險，包括對個人造成身體或心理傷害，對組織造成聲譽、法律及經濟損失，通過錯誤信息干擾治理流程等。而與此同時，雖然各國紛紛立法或制定規則對本國人工智能的發展予以約束，在國際層面針對人工智能技術尚無統一的、具有約束力的國際法律或條約。弗里德曼在文章中也預測，若中美各自發展人工智能系統，大量貿易將陷入停滯，只能互售大豆釀醬油。世界將淪為一個高科技封建主義世界。由此可見，人工智能技術本身不穩定性和不可解釋性會讓中美兩國在這一領域的合作與競爭充滿挑戰。

2024年5月，中美人工智能政府間對話會議在日內瓦舉行。從會後新聞稿看出，雙方對此次對話均作出積極評價，認識到應確保人工智能技術的安全，重申繼續致力於落實“舊金山共識”。但在治理側重點上，雙方存在較大分歧和錯位。中方強調人工智能技術的發展，就美方在人工智能領域對華限制打壓表明嚴正立場；強調人工智能發展的公平，認為AI全球治理應發揮聯合國主渠道作用，建議建立具有廣泛共識的全球人工智能治理框架和標準規範。而美國則傾向於將本次對話視為技術交流與交鋒，聚焦人工智能的安全風險和治理，強調應確保建設AI系統安全、可靠、可信賴，並關注所謂中方“錯用”（misuse）人工智能的做法。2024年11月，中美兩國領導人在智利利馬會晤，雙方就有必要加強國際合作，促進人工智能向善普惠，維持由人類控制核武器使用的決定達成一致，獲得國際社會廣泛好評。如今，距離上一次中美人工智能政府間對話會議已經近一年，雙方下次人工智能政府間對話仍未提上正式議程。而與此同時，僅2024年一年，美方共數次發佈針對中國的人工智能管制措施和政策，拜登政府在針對中國的AI技術出口管制措施上仍在不斷加碼。

弗里德曼先生確實看到了中美兩國在人工智能問題上進行合作的必要性和緊迫性，但在當前中美博弈的背景下，中美兩國要想實現真正意義上的人工智能治理對話與合作，構建應對人工智能“生存性”風險的合作治理框架，必須首先跳出當前存在的“安全困境”。

“安全困境”是美國政治學者約翰·赫茨於1950年提出，指的是國家在追求自身安全的過程中，由於缺乏信任，可能導致軍備競賽、緊張局勢升級，甚至引發衝突。這一困境的根本原因在於各國對彼此意圖的不確定性，以及國際體系中無政府狀態導致的安全焦慮。根據郝茨定義，安全困境存在以下幾類要素：一是無政府狀態的國際環境，缺乏規範行為和保護行為體免受攻擊的權威組織或機構；二是行為體之間的相互懷疑和恐懼，對對方意圖的不確定性；三是各行為體為尋求安全而強化競爭和採取安全措施；四是對所有行為體安全的減損。中美兩國在人工智能上的競爭與合作呈現出“安全困境”的種種跡象。

中美之間進行人工智能治理合作，需要努力走出“安全困境”，而在當前背景下重建最基本的信任是唯一出路，需要雙方都採取切實措施，一步一步做出調整。一是需要調整對彼此國家的認知和定位。如一定程度的競爭不可避免，那也要避免你死我活、非黑即白、零和博弈的惡性競爭。這是雙方進行合作的認知前提。二是為國家安全設定紅線，明確哪些為無法合作和交流的敏感領域，同時克制無限擴大這些領域的衝動。這是為雙方合作提供確定性的基礎。第三，在劃定紅線的基礎上，可嘗試在低敏感領域開展交流與合作，例如兩國可圍繞AI治理的政策和原則、最佳實踐、能力建設上進行對話，以期增進相互了解，形成自身在對方眼中的可預測性和可解釋性，逐步建立信任。第四，2024年兩國在聯合國大會相互聯署各自關於人工智能的決議，為國際社會樹立了良好典範，未來應繼續在聯合國框架下共同推動人工智能國際治理，探討建立規範相關行為和保護行為體的權威組織或機構。第五，作為人工智能技術與能力的兩個大國，中美應攜手推動人工智能能力建設，為提高發展中國家人工智能能力水平，縮小數字鴻溝，推動實現聯合國可持續發展目標貢獻力量。

Yes, they can. As the world's two major powers in AI technology, the U.S. and China must work together to build capacity, contribute to AI for developing countries, bridge the digital divide and help achieve the United Nations Sustainable Development Goals.

On March 25, after a visit to China, New York Times columnist Thomas Friedman called for cooperation between the United States and China to jointly address the potential risks that artificial general intelligence (AGI) may pose to humanity. He argued that what Soviet-American nuclear arms control was to world stability since the 1970s, U.S.-China collaboration on AI currently is for the same objective.Making sure we effectively control these rapidly advancing AI systems is essential for tomorrow's world, Friedman wrote. He pointed out sharply: “Once AGI arrives, if we are not assured that these systems will be embedded with common trust standards, the United States and China will not be able to do anything together.”

In the context of AI governance, the word “trust” is frequently mentioned — trust in machines, trust in human-machine collaboration, trust in the users of AI systems, trust in institutions and trust between nations. Trust is a fundamental element and a valuable resource.Israeli historian Yuval Noah Harari recently spoke at the China Development Forum, offering his perspective on trust. He was quoted by saying: “We should build more trust between humans before we develop truly super intelligent AI agents. …At its core, trust is essential for survival. …All over the world, trust between humans is collapsing.” Together humans can control AI, he added, “but if we fight one another, AI will control us.”I recently attended the inaugural Global Conference on AI, Security and Ethics organized by UNIDIR and Microsoft, where there was lots of talk about how geopolitical competition undermines trust between nations, thereby delaying — and even jeopardizing — the progress of international AI governance.

Then, what exactly is trust? What are the standards by which trust can be defined? According to the Cambridge Dictionary, trust is “to believe that someone is good and honest and will not harm you, or that something is safe and reliable.” The Merriam-Webster Dictionary defines it as “assured reliance on the character, ability, strength, or truth of someone or something.” In the Modern Chinese Dictionary, trust (xinren) is explained as “believing in someone to the extent of being willing to entrust them with responsibilities.” In the Xinhua Dictionary, the character “信” means not to doubt and to consider something reliable, while “任” means to trust and to rely upon.In the field of international relations, trust refers to a state or actor's positive expectations regarding another party's intentions, commitments and actions. Taken together, these definitions seem to encompass the following essential elements: credibility, predictability and reliability.

Can the United States and China, as Friedman hopes, establish a basic level of trust and carry out urgently needed cooperation in the field of AI governance? To answer this question, we need not only to analyze the current state of U.S.-China relations but also take into account the inherent characteristics of AI technology itself.

From the perspective of U.S.-China relations, there currently exists a serious trust deficit between the two countries. Mutual suspicion and fear, as well as uncertainty about each other's intentions, have long influenced the trajectory of bilateral relations. In the early 21st century, China's rapid economic and military growth sparked widespread debate within the United States about how to respond. The Belt and Road Initiative, Made in China 2025 and China's significant increase in investment in high-tech industries have fueled U.S. concerns that China might challenge its global leadership.In December 2017, in its National Security Strategy, the first Trump administration labeled China a “rival.” The 2022 National Security Strategy identified China as “the most consequential geopolitical challenge” to the United States and positioned China as a strategic rival. It drastically narrowed the space for dialogue and cooperation between the two countries.Several U.S. officials have publicly stated that ensuring absolute superiority in technological competition with China is essential. To achieve this, the U.S. has expanded the concept of national security, established foreign investment screening mechanisms, strengthened export controls on key technologies and equipment, restricted technology transfers to China and placed Chinese companies and institutions on the export control entity list — all in an effort to safeguard a technological edge in the name of national security.China views these U.S. actions as deliberate attempts to hinder its development, preserve American hegemony and deprive China of its right to development. Further, during the latter part of Trump's first term, ideological elements began to take center stage in U.S. strategy toward China. In May 2020, the Trump administration released the United States Strategic Approach to the People's Republic of China, elevating the U.S.-China rivalry to an ideological level and asserting that competition between the two countries spans political, economic and societal dimensions. Following this, several senior Trump administration officials delivered speeches attacking China's political system and criticizing its domestic and foreign policies.Amid the broader context of strategic rivalry and geopolitical tension, mutual suspicion and fear between the two major powers have intensified, exacerbating the uncertainty surrounding each other's intentions. This growing distrust makes credibility impossible to achieve. Moreover, the early-stage foreign policy of Trump 2.0 has made predictability a scarce resource in the bilateral relationship.

From the perspective of AI technology itself, despite its remarkable advancements in recent years, several unresolved issues persist, including such things as hallucinations and black boxes. In the International AI Safety Report published in January, Professor Yoshua Bengio and his team classifies general-purpose AI risks into three categories: malicious use risks, risks from malfunctions and systemic risks.The report also pointed out that, general-purpose AI can suffer reliability issues that include confabulations or hallucinations; failures at performing common-sense reasoning; and inferences, or failures to reflect contextually relevant, up-to-date, unbiased knowledge and understanding. It went further to say that instances of reliability failure can create risks, such as physical or psychological damage to individuals; reputational, legal, and financial damage to organizations; and misinformation that impacts governance processes.Meanwhile, although countries have been enacting laws and regulations to govern AI domestically, there remains no binding international law or treaty addressing AI technology at the global level. In his article, Friedman predicted that if China and the U.S. were to develop their own AI systems independently, much of their trade would stagnate, reducing exchanges to basic goods like soybeans and soy sauce. “It will be a world of high-tech feudalism,” he warned. This underscores how the inherent instability and unreliability of AI technology could layer China-U.S. cooperation and competition with challenges.

In May last year, China and the United States held their first intergovernmental dialogue on AI in Geneva. According to the news release, both sides assessed the dialogue positively, recognizing the need to ensure the safety of AI technology and reiterating their continued commitment to implement the important consensus reached at the San Francisco meeting of their two leaders. However, there were significant differences and misalignments in terms of the focus of AI governance. The Chinese side emphasized the development of AI technology, expressing a stern position on U.S. trade restrictions; highlighting the fairness of AI development, believing that the UN should play a leading role in global AI governance; and proposing the establishment of a global framework for AI governance and standards.In contrast, the U.S. tended to focus on the safety risks and governance of AI, emphasizing the need to ensure that AI systems are safe, reliable and trustworthy, and expressing concerns over the so-called misuse of AI by China.In November President Xi Jinping and President Joe Biden met in Lima, Chile, where they agreed to maintain human control over the use of nuclear weapons, a move that was widely acclaimed by the international community. But nearly a year has passed since the last meeting of China-U.S. Intergovernmental Dialogue on AI, and a second meeting has not yet been scheduled. Meanwhile, the Biden administration last year continued to escalate AI technology export control measures against China. The recent trade war started by the U.S. against China has made the situation even worse.

Friedman recognized the necessity and urgency of cooperation between the U.S. and China on AI governance, and rebuilding the basic level of trust is the only way out. Both sides need to take concrete steps and make adjustments — against a background of geopolitical tension — if they expect to achieve genuine dialogue and cooperation and to construct a cooperative governance framework that addresses the existential risks of AI.

Both China and the U.S. need to adjust to the perceptions and positioning of the other. If a certain degree of competition is inevitable, then we should at least avoid the vicious competition of a life-or-death struggle or zero-sum game. Setting red lines for national security is essential. Both sides should clearly define sensitive areas where cooperation and exchanges are not possible, and restrain the impulse to indefinitely expand these areas.Once red lines have been set, attempts can be made to carry out exchanges and cooperation in areas of lower sensitivity. For example, the two countries can engage in dialogue on policies and principles of AI safety, best practices and capacity-building, to enhance mutual understanding, establish predictability and interpretability in the eyes of each other, and then gradually build trust.

Last year, the two countries sponsored joint resolutions on AI at the United Nations General Assembly, setting a good example for the international community. They should continue to jointly promote the international governance of artificial intelligence under the UN framework and explore the establishment of an authoritative organization to regulate behavior and protect actors.Last but not least, as two major powers in AI technology and capability, the U.S. and China must work together to advance capacity-building, contribute to enhancing AI capabilities in developing countries, bridge the digital divide and help achieve the United Nations Sustainable Development Goals.