UN AI Governance Fragmentation: US-Russia Joint Obstruction and Cracks in Europe's Human Control Declaration

Executive Summary

During the UN General Assembly week in September 2026, AI governance fragmented into three distinct tracks: US-China bilateral agreements, US-Russia obstruction of weapon system norms, and non-binding declarations by middle powers and the EU. Despite differing motivations, the United States and Russia—both permanent members of the UN Security Council and nuclear-armed states—converged in their interest to avoid binding their military AI operational discretion through treaty obligations. This demonstrates that norm competition regarding the military use of AI in emerging and non-traditional security domains could be hollowed out under great-power leadership. While the same principle of "human control" refers to battlefield targeting decisions in Geneva, it denotes the control of frontier model development in White House and Security Council channels, establishing a locked-in division of labor rather than convergence toward a single governance framework. The declaration by 22 European nations aims for a stigmatization effect led by non-possessor states, but due to uncertain signals of industry compliance, it is more likely to converge toward a nuclear non-proliferation-style possessor club model rather than the Ottawa Treaty. The fact that South Korea, which championed the REAIM Seoul Summit, did not participate in this declaration leaves room for three interpretations: alliance considerations, tension between industrial promotion and regulatory discourse, and agenda separation. To maintain its credibility as a norm champion, South Korea needs to participate in SIPRI-style informal working channels and respond by separating military and civilian AI agendas.

I. Issue Analysis

UN AI Governance Fragmentation: US-Russia Joint Obstruction vs. Europe's Human Control Declaration

1. Background and Progress of the Issue

During the week of the 81st UN General Assembly in September 2026, three parallel lines regarding AI governance were drawn simultaneously in New York [17]. The flashpoint was a report by the Italian daily La Repubblica, citing The Washington Post [1]. According to the report, the United States and Russia jointly blocked the introduction of international restrictions aimed at mandating human intervention in targeting decisions by AI weapon systems [1]. Specifically, a provision requiring a human, rather than a computer, to make the final decision on which AI-proposed targets to strike was defeated [1].

This joint obstruction is not a one-off event. An EAI analysis defines it as "the result of a combination of the Trump administration's skepticism toward international AI oversight and Russia's distrust of Western-led multilateral norms" [3]. Analysts assess that the structural positions shared by the two countries as permanent members of the Security Council and nuclear-armed states in the field of disarmament norms were at play once again [3]. Around the same time, German Foreign Minister Johann Wadephul called for AI regulation from the podium of the UN General Assembly and held a separate meeting with Russian Foreign Minister Sergey Lavrov [4]. The declaration of the human control principle by 22 countries, including European nations, aimed at filling the normative vacuum, emerged precisely as a response to this obstruction [3].

During the same week, another track was underway in Washington. Following a summit between President Trump and President Xi Jinping, the White House announced an agreement to establish a bilateral dialogue channel and a notification system for AI incidents [7][14]. Subsequently, on September 23, an unusual scene unfolded at the UN Security Council, where heads of AI companies, including Sam Altman, personally appeared to brief the council under the agenda of "maintaining international peace and security" [16][6]. This meeting, which France had been preparing since August, aligns with the trend of Anthropic CEO Dario Amodei's post on AI deceleration spreading beyond Silicon Valley to become an international agenda item [6].

2. Current Situation

The three tracks appear to be hardening in parallel without converging. The US-China track progressed to the creation of an incident notification channel through a summit agreement, with the next round scheduled for November [7][14]. While the Qatari media outlet Geo News evaluated the Washington summit as "ceremonial without a breakthrough," it highlighted the establishment of the AI communication channel as a concrete achievement [14]. Conversely, the Council on Foreign Relations (CFR) points out that this agreement is less a signal of easing competition and more a paradoxical scene of "competitors suddenly pleading for restraint" [8].

The US-Russia track appears focused on blocking binding provisions in weapon system regulation negotiations, such as those in Geneva [1][3]. This track is characterized by its passive operation—blocking the introduction of specific provisions at the negotiating table—rather than visible diplomatic events like Security Council briefings or White House summits. The track led by middle powers, including European nations, is an attempt to fill this vacuum with non-binding declarations instead of formal multilateral norms [3]. From the same UN General Assembly podium, Singaporean Foreign Minister Vivian Balakrishnan went a step further, publicly proposing a UN framework convention on AI safety and the creation of a new international organization [10][15]. While asserting that "it is already too late to call for a safety pause," he emphasized the need for "a new institution that can set safeguards and standards" [10].

Positions also diverge among middle powers. During a UN General Assembly debate on AI governance, Rwandan Prime Minister Nsengiyumva pointed out that "countries that do not develop frontier AI should not simply trust that systems developed elsewhere will work safely for their populations and conditions," criticizing the exclusion of countries that are not in a position to test AI from the discussions [12]. The Carnegie Endowment for International Peace suggests that middle powers like Germany and India face the risk of being cut off if they rely on the US or China, advising that they should build "breakout capacity" even if they do not possess frontier capabilities themselves [11].

3. Key Actors and Positions

The United States exhibits a dual approach. While jointly blocking the introduction of weapon system regulatory provisions with Russia [1][3], it simultaneously cooperates with China to establish an AI incident notification channel [7][14]. The common thread in both tracks is a preference for managing the agenda through bilateral or minilateral channels among great powers, avoiding binding norm-making through multilateral organizations. The Carnegie Endowment for International Peace summarizes this with the question, "Who has the authority to put guardrails on AI?" pointing out the need to supplement the limitations of existing crisis communication efforts and keep pace with frontier AI [2].

While maintaining structural distrust of Western-led multilateral norms, Russia shares common ground with the United States on preserving operational discretion over military AI, despite their confrontation in the broader strategic competition [3]. Foreign Minister Lavrov continued tensions with the West at the same UN General Assembly by stating, in relation to the war in Ukraine, that "Europe is blocking peace" [1], suggesting that the joint obstruction of AI norms is an extension of a broader distrust toward the West.

Twenty-two European nations, including Germany, seek to fill the normative vacuum with a declaration on the principle of human control, but this declaration is non-binding [3][4]. Foreign Minister Wadephul's simultaneous meeting with Lavrov and his General Assembly speech [4] demonstrate Europe's dual strategy of maintaining bilateral channels with Russia while not giving up on multilateral norm formation. Singapore takes the most active stance among middle powers, proposing more institutionalized alternatives such as a UN framework convention and a newly established international organization [10][15]. While expressing sympathy for cooperation on frontier AI during the Security Council briefing, China is running a dual track, with its Ministry of State Security defining US AI models as a separate security threat [6]. Vice President Han Zheng criticized unilateral military action in his General Assembly speech, stating that "great powers must follow rules" [4], which can be read as an attempt by China to steer the narrative in AI militarization discussions toward great-power responsibility.

4. Key Issues

The most fundamental issue is that the same phrase, "human control," refers to different targets depending on the track. In the US-Russia obstruction track, human control refers to the authority to make targeting decisions on the battlefield [1], whereas in the European declaration and Singapore's framework convention concept, it refers to safeguards during the development phase of frontier models [3][10]. The fact that these two usages employ the same vocabulary but target entirely different subjects and implementation actors makes it difficult to discern whether this fragmentation represents a process of norm diffusion through multiple pathways or a structure where great powers isolate binding issues into bilateral channels to evade pressure from multilateral arenas.

The second issue is the effectiveness of the middle-power declaration. It remains to be seen whether a non-binding declaration adopted without great powers and major AI companies will function like the IAEA-style verification model of the nuclear non-proliferation regime, or exert normative power solely through a stigmatization effect like the Ottawa Treaty [2][3]. SIPRI's convening of an informal dialogue on September 9–10, bringing together defense and foreign policy officials, industry representatives, the International Committee of the Red Cross (ICRC), NATO, and the UN Institute for Disarmament Research (UNIDIR) to discuss the lawful development and operationalization of military AI [5], demonstrates that alternative channels at the working level are being explored while the formal treaty track is blocked.

The third issue is the asymmetry within the middle powers themselves. A gap exists between countries like Germany and Singapore, which vocalize norm formation, and countries like Rwanda, which are excluded from discussions due to a fundamental lack of verification capabilities [11][12]. This demonstrates that middle-power coalitions are not bound by a single, unified interest, raising the challenge for South Korea to distinguish which type of middle-power group it aligns with should it join declarations by like-minded states.

II. In-Depth Analysis of the Issue

UN AI Governance Fragmentation: US-Russia Joint Obstruction vs. Europe's Human Control Declaration

In-Depth Analysis

1. Root Cause: Asymmetric Interest Structures of Permanent UNSC Members

The root of US-Russia cooperation is not ideological affinity; it is the alignment of their structural positions. Both countries are permanent members of the UN Security Council and, simultaneously, nuclear-armed states. In Cold War disarmament negotiations, these two statuses always operated in tandem. Under the system of mutually assured destruction (MAD), the US and the Soviet Union adopted a dual strategy: opposing norms that would ban the use of nuclear weapons outright, while cooperating to prevent their proliferation. The same pattern is repeating in AI weapon systems. Mandating human intervention in targeting decisions restricts operational discretion, and neither country has any incentive to concede this discretion [1][3].

The motivations of the Trump administration and Russia are qualitatively different. On the US side, there is skepticism toward international oversight itself. The Carnegie Endowment for International Peace summarizes this with the question, "Who has the authority to put guardrails on AI?" [2]. With its own companies holding a technological edge, the US sees no reason to agree to an external body setting operational standards. On the Russian side, there is a fundamental distrust of Western-led multilateral norms [3]. Since the outbreak of the war in Ukraine, instances of Russia cooperating with Western-designed norms within the UN framework have been rare. This shares the same structural logic as Foreign Minister Lavrov's statement at the UN General Assembly during the same week that "Europe is blocking peace" [1]. The two countries simply reached the same conclusion for different reasons; reading this joint obstruction as an "alliance" would be a misinterpretation.

A more fundamental cause is that the term "human control" refers to different targets depending on the forum. In the Geneva weapon system negotiations, the term refers to battlefield targeting decisions. In the US-China agreement originating from the White House and the Security Council briefings, it refers to controlling the development of frontier models [7][14][16]. These two agendas differ in their regulatory targets, stakeholders, and enforcement bodies. Because the same principle is used in different contexts, what outwardly appears to be a single AI governance discussion has actually split into three non-converging tracks.

2. Structural Context: Three-Way Division of Labor and Power Asymmetry

Overlaying the three tracks reveals a single division of labor. The US-China bilateral track is a high-level channel where leaders and corporate CEOs sit at the same table [7][14]. Sam Altman's personal appearance at the Security Council briefing symbolizes this structure [16][6]. It is a method where state and corporate actors mix to create non-binding communication channels. In contrast, the US-Russia obstruction track has low visibility, operating passively by defeating specific provisions in negotiating rooms [1][3]. It is a track that does not need to publicize achievements; indeed, keeping it hidden is more advantageous. The middle-power/EU declaration track emerges in the gaps between these two. This is illustrated by the German Foreign Minister calling for AI regulation from the UN podium immediately after meeting with Lavrov [4]. While it lacks binding force, it also has low barriers to participation.

Power is distributed asymmetrically within this structure. Great powers pull binding issues into bilateral channels. Middle powers can only create norms in spaces vacated by great powers. Rwandan Prime Minister Nsengiyumva's statement during the UN AI governance debate that "countries should not simply trust that systems they did not develop will work safely for their populations and conditions" exposes this asymmetry most directly [12]. The Carnegie Endowment's analysis of middle powers like Germany and India addresses the same issue from a different angle: middle powers cannot build frontier AI on their own, and relying on the US or China exposes them to the risk of being cut off [11].

Economic structures also underpin this fragmentation. Frontier model development capabilities are concentrated in a handful of companies in the US and China. Similarly, military AI operational capabilities are concentrated in the US and Russia. The role middle powers and the EU can play in norm declarations is verification and stigmatization, not the control of the technology itself. Singaporean Foreign Minister Balakrishnan's proposal for a UN framework convention and the creation of a "new international organization," while simultaneously declaring that it is already "too late" for a safety pause, is a self-acknowledgment of this limitation [10][15].

3. Comparison of Historical Precedents: Crossroads Between the Ottawa Treaty and the NPT/IAEA Model

The question posed by the report converges on two contrasting disarmament precedents. One is the Ottawa Treaty (Anti-Personnel Mine Ban Convention). This treaty was signed without the participation of major military powers like the US, Russia, and China. Nevertheless, it is evaluated as having largely succeeded in attaching an international stigma to the use of landmines by non-signatory states. This is an instance where a coalition of middle powers and civil society took the lead, and the absence of great powers sharpened the moral framework rather than undermining the treaty's legitimacy.

The other is the Treaty on the Non-Proliferation of Nuclear Weapons (NPT) and the IAEA model. This system does not function without the voluntary participation and acceptance of verification by the possessor club. As the North Korean case demonstrates, if a key party withdraws from the verification system, the effectiveness of the regime itself drops precipitously. It remains to be seen which of these two precedents the military AI norm will closer resemble. However, developments to date signal a closer alignment with the IAEA model. The very fact that the middle-power declaration advocates for an "IAEA-style verification body concept" while excluding the US, Russia, and China—the very targets of verification—foreshadows the structural limitations of this declaration.

China's response suggests a third option at this crossroads. Instead of participating in Western-led multilateral regimes, China is concurrently building its own alternative institution, the World AI Cooperation Organization (WAICO) [9]. This pattern is similar to the Non-Aligned Movement during the Cold War or the establishment of the Shanghai Cooperation Organisation. It is a strategy of not opposing the existing regime outright, but rather dispersing leadership in norm formation through a parallel, alternative body. At the same time, China exhibits a dual structure: expressing the principled stance that "great powers must follow rules" during Security Council briefings [4], while running a separate track through its Ministry of State Security that defines US AI models as security threats [6]. Great powers do not deviate from multilateral regimes in a single way. The US takes the path of skepticism, Russia of rejection, and China of building alternative institutions.

4. Key Variables: Fragmentation or Re-convergence?

The variables shaping future developments narrow down to three.

First is the level of institutionalization of the US-China bilateral channel. If the incident notification system acquires substantive operational rules during the second round scheduled for November [7], this channel could effectively solidify as the de facto standard for AI security governance. Conversely, if the notification channel remains merely symbolic, the space for middle-power declarations will expand.

Second is the expansion of SIPRI-style informal working-level consultative bodies. The closed-door dialogue in September, which included defense and foreign ministry officials, industry representatives, the ICRC, NATO, and UNIDIR [5], suggests an alternative pathway where substantive norm formation can occur when formal treaty negotiations are blocked. Whether such informal channels develop into binding agreements or remain limited to dialogue itself will be a turning point.

Third is whether the middle-power declaration generates a stigmatization effect. The key is whether the 22-nation declaration succeeds in imposing international costs on the behavior of non-signatory states, like the Ottawa Treaty, or remains merely a declaration. So far, with both the US and Russia continuing to block provisions at the negotiating table unhindered by this declaration, there are no clear signs of a stigmatization effect emerging [1][3].