Why the Study of AI Geopolitics Needs Ethics

The field of AI geopolitics can be conceptualised through three clusters of research: how are a) developments in, b) governance and regulation of, and c) conflict deployment of AI technologies shaped by geopolitical processes?

The first cluster revolves around development. A Lazard (2023) Geopolitical Advisory report identified four key ‘building blocks’ to competitiveness in the global AI race – computing power (including semiconductors and data centre infrastructure), talent (researcher quality and quantity), data, and physical infrastructure (energy and water resources). These blocks are asymmetrically concentrated in a handful of economies and corporations – for example, NVIDIA, Taiwan Semiconductor Manufacturing Company (TSMC), Broadcom, and Advanced Semiconductor Materials International (ASML) (Mahindroo, Suryanarayanan, and de Jong Reference Mahindroo, Suryanarayanan and de Jong2025).

For instance, the US and China are both endowed with significant concentrations of high-skilled talents and natural resources. While countries such as Japan, Israel, and France produced far fewer top talents in absolute quantities, their highly integrated roles within a US-led AI ecosystem have enabled their corporations and research laboratories to access cutting-edge systems (NVIDIA 2024; Press Reference Press2024; École Polytechnique 2025). The flipside to such cross-national synergies consists of governmental restrictions to access. Such restrictions are often justified in terms of interest protection – for example, through preventing nefarious usage of their own citizens’ data by foreign companies.

Other literature has focused on restrictions put in place for developmental throttling – a chief case-in-point being attempts by governments to restrict the pace at which other countries are able to acquire and deploy cutting-edge semiconductors in GPUs, which play a critical role in AI applications. The bipartisan US CHIPS and Science Act, 136 Stat. 1366 (CHIPS Act 2022) was initially hailed in policy circles for curbing China’s access to advanced chips – with leading consumer-facing Chinese giants, such as Huawei, long depending upon fabricators such as Samsung and TSMC in providing crucial components (Shivakumar, Wessner and Howell Reference Shivakumar, Wessner and Howell2024). More recently, Shivakumar, Wessner, and Howell (Reference 89Shivakumar, Wessner and Howell2025) have argued that ‘the imposition of export controls has resulted in China doubling down on its [own] development efforts that could well produce breakthrough technologies’ – the announcement by Chinese start-up DeepSeek in January 2025 of an open-source research model, R1, which matched roughly leading models from Google and OpenAI, spurred speculation that Chinese developers were able to increase algorithmic efficiency and make use of insightful data compression and model optimisation strategies in order to make up for limitations in computational power and typically smaller datasets, as compared with their overseas counterparts.

The second cluster concerns the governance and regulation of AI. A prevalent approach among existing theorists (e.g. Bradford Reference Bradford2023) is to focus on key geopolitical clusters of AI powers – what she terms “digital empires” and Schmid et al. (Reference Schmid, Lambach, Diehl and Reuter2025) term “technopoles” – and ascribe to each region distinctive principles guiding their approach to regulation. Per this framing, the EU is portrayed as primarily ‘rights-centric’, leveraging its bureaucratic apparatus and effective legal dissemination methods in accomplishing the proverbial Brussels Effect – whereby European ordinances have an outsized impact on blocs of nation-states that possess extensive commercial connections to and legal ties with Europe (Wang et al. Reference Wang, Milne, Hu and Khan2024). The US can be characterised as “market-driven”, ostensibly because its government would “step aside to maximise the private sector’s unfettered innovative zeal” (Bradford Reference Bradford2023, 24). Finally, some would assert that the Chinese model is top-down and strives to advance both innovation and stability, coupled with effective enforcement mechanisms underpinned by strong state capacities and understanding of AI technologies (Reed Smith Reference 88Reed2024). More recently, Chinese state signalling suggests that it is endeavouring to position the country as a key proponent of openly available AI, building off the positive reception of DeepSeek; the 2024 Shanghai Declaration on Global AI Governance states, “[w]e advocate the spirit of openness and shared benefit and will promote exchanges and cooperation on global AI research resources.” (Ministry of Foreign Affairs, People’s Republic of China [MFA] 2024).

The reality is more complex than the tripartite framing of digital empires or technopoles suggests. As Zhang (Reference Zhang2024) observes, in China, the delicate interplay between the senior political leadership, provincial and local bureaucratic regulators, large platform economy companies, and individual workers tends to give rise to a volatile regulatory process generating fragile policy outcomes which may be open to interpretation and rearrangements based on market and leadership sentiments. Meanwhile, Fang and Hwang (Reference Fang and Hwang2023) argue that the US state is increasingly embracing its own form of “techno-nationalism”, proactively promulgating the narrative that the “competition between distinct ‘Chinese’ and ‘US’ models of leveraging emerging technologies” is a zero-sum struggle and must be won (The White House 2025). The implication is that American companies should fall in line with their state’s national security imperatives, without questioning the underlying normative premise – and indeed, many of them have, with Anthropic’s Dario Amodei becoming one of the leading defenders of US export controls (Amodei Reference Amodei2025), despite his confrontation of the Pentagon in early 2026 over the parameters of reasonable AI deployment. The EU AI Act’s cumbersome risk-tiered system of governance looks very much like a top-down enforcement mechanism – so heavy-handed that many worry about its drag on innovation speed and flexibility. Clearly, therefore, debates over the geopolitics of regulation must evolve to better capture the nuanced similarities and differences in governance dynamics across different AI powers. While the ‘multiple empires’ framing is tempting, it is ultimately too simplistic.

The third cluster pertains to the application and use of AI in conflict contexts. Payne (Reference Payne2021) observes that autonomous weapons are too ubiquitous and useful for states to be wholly outlawed. AI-enabled ‘warbots’, then, are a matter of political inevitability – even despite potential risks arising from misalignment between the occasion and intensity of force and the specific military contexts at hand. Many have warned (e.g. Simmons-Edler et al. Reference Simmons-Edler, Badman, Longpre and Rajan2024; Kmentt Reference Kmentt2025) of the perils of an AI-enabled autonomous weapons arms race, which could generate unproductive escalation, heightened uncertainty, and proliferation of AI technologies among illicit actors.

This emerging geopolitical discourse on AI is limited by two theoretical defects: ethical statism and great power solipsism. We briefly discuss both.

The first limitation consists of the absence of defence for what we term an ethical statist premise, namely, that individual states should behave in ways that advance solely their own interests. Most government policies, directives, and platforms on AI governance tend to place significant emphasis upon the securing of core interests of the state. This state-interest-driven nature of AI analysis tends to ignore considerations of what is just, fair or otherwise normatively desirable.

Much attention is paid towards the instrumentalist question of how a particular objective (e.g. ‘national security’) is to be attained; far less attention – if any – is paid towards the normative question: what should be the objective towards which countries must strive in the context of inter-state AI dynamics?

This claim will be contested by those denying that states possess any duties to actors beyond their borders. Some neo-realist accounts (Waltz Reference Waltz1979; Mearsheimer Reference Mearsheimer2001) hold that, given the anarchic conditions characterising the international order, states should prioritise their own survival, security, and interests. Indeed, Carr (Reference Carr1940) goes one step further in noting that the empirical reality of conflicts of interest between different countries gives rise to an international order forged through strength and contrasts of power. Walzer once famously described American IR training post-1945 with the following: “Moral argument was against the rules of the discipline as it was commonly practiced.” (See Nye Reference Nye2020).

In response, states should incorporate into their considerations the avoidance of unnecessary harms to and generation of benefits for other states when it comes to their decisions. Should here is normative: just as the interests of its individual nationals matter qua intentional human agents, the interests of non-nationals should also matter in constraining and guiding the actions of states.

Moreover, the prescriptions of more stringent export restrictions advocated by Western liberal democratic governments have at times ended up undermining the interests of their own citizens as well as those in allied states. Hence, there is an argument from self-interest to be made here as well. For example, the AI Diffusion Rule (Bureau of Industry and Security 2025b) proposed by former US President Joe Biden would have imposed excessively cumbersome regulatory requirements against American companies. Likewise, as we have seen in Supreme Court oral arguments in Learning Res. Inc. v. Trump, No. 24-1287, and Trump v. VOS Selections, Inc., No. 25-250, the main argument against the tariffs Donald J. Trump’s administration has levied on most other states is that they impose a tax on American citizens in the form of higher prices for imported goods (Learning Res. Inc. v. Trump Reference Res2025).

The second issue, Great Power Solipsism, consists of the excessive emphasis upon large nation-states and downplaying the role of small and medium states, as well as non-state actors, when it comes to their role within the global AI landscape (Wang et al. Reference Wang, Milne, Hu and Khan2024). Whether it be the ‘tripolar’ frame focusing on China, the US, and the EU (Bradford Reference Bradford2023; Schmid et al. Reference Schmid, Lambach, Diehl and Reuter2025), or the ‘Democracy vs. Autocracy’ frame adopted by select actors in the US (Aschenbrenner Reference Aschenbrenner2024), these narratives portray global AI development through the lenses of two to three actors with overwhelming influence.

This view ignores two key resources by which smaller countries can play an active role in the global AI landscape. The first is strategic hedging (Kuik Reference Kuik2008, Reference Kuik2016): pursuing a strategy whereby they build partnerships with a multitude of actors while refusing to conform to one digital sphere of influence. For example, Singapore and Vietnam are skewed hedgers seeking cooperation with American AI regimes while retaining openness to Chinese investment (Zhao Reference Zhao2024). Malaysia has positioned itself as both a chip manufacturing and re-exporting hub for American and Chinese semiconductor firms alike – courting Advanced RISC Machine (ARM) Holdings (Dib, Le, and Sharma Reference Dib, Le and Sharma2025), and importing extensively from Nvidia (Sipalan Reference Sipalan2025), while facilitating regional consolidation and expansion of Chinese tech firms. Malaysia now accounts for 13 per cent of the world’s semiconductor testing and packaging (Medina Reference Medina2025). Hedging enables these Southeast Asian states to preserve autonomy and flexibility.

The second means consists of niche specialisation. Medium AI powers are identifying and harnessing their expertise in particular domains of AI development, diffusion, and regulation in order to make themselves useful for a range of partners – thereby shoring up their bargaining capital. German AI companies, for example, have opted to specialise in image generation and translation (Praas and Balland Reference Praas and Balland2025).

Beyond nation-states, there exist other actors who exert substantial influence over AI in geopolitical settings. From terrorist and insurgent groups to powerful large corporations (Gu Reference 82Gu2024 warns of a “new type of Leviathan” emerging from the deep pools of Big Tech). These actors can make intentional choices that shape the trajectory of AI development, diffusion, and application.

Why the Study of AI Ethics Needs Geopolitics

The literature on geopolitical implications of AI is surprisingly underexplored in the philosophy of AI, as well as in the computer science literature touching on fairness, accountability, and other ethical topics. This could be in large part an accident of the particular time, place and context in which the field of AI ethics emerged and in which it evolved as a subdiscipline of computer science and philosophy. In this section, we explain how both of these facets created a path dependence affecting the way in which AI ethics unfolded, and why that path dependence largely steered clear of core geopolitical issues.

In the early 2000s, one of the most exciting areas of computer science and operations research was sparse matrix completion: identifying mathematical techniques for imputing missing values in large data matrices (Candès and Recht Reference Candès and Recht2009; Li et al. Reference Li, Huang, So and Zhao2019). Indeed, Netflix began running what they called the Netflix Prize Competition, a hackathon-style tournament for predicting film user ratings. To run the competition, they released an anonymised data set of user reviews, on which competitors could train their models (Bennett and Lanning Reference Bennett and Lanning2007). The competition was successful and proved to be very popular.

Yet in 2007, Arvind Narayanan and Vitaly Shmatikov did something different: instead of developing a model that would actually compete in the Netflix Prize challenge, they developed a technique that could take the anonymised Netflix Prize data and systematically re-identify the users referenced therein (Narayanan and Shmatikov Reference Narayanan and Shmatikov2007). This development amplified a lesson that had previously been learnt in the literature on electronic health records: namely, that anonymising data is not enough for guaranteeing user privacy. Around the same time that Narayanan and Shmatikov undermined the value of anonymising data, Cynthia Dwork and her colleagues had been developing methods for obtaining provable privacy guarantees by adding noise to data – a technique that came to be known as ε-differential privacy (Dwork Reference Dwork, Bugliesi, Preneel, Sassone and Wegener2006). Given these two developments, the technical literature on privacy grew quickly.

This was a whole new area of research: one could generalise Dwork’s approach to cases with different data-generating distributions, or relax its assumptions, or add noise from a different statistical model (Dwork et al. Reference Dwork, Bugliesi, Preneel, Sassone and Wegener2006; Pejó and Desfontaines Reference Pejó and Desfontaines2020). As fascinating as these developments are, there was a certain historical contingency to how AI ethics evolved: the time and place for working on privacy were just right, given the concerns about large recommender systems using our data without consent. Such issues would also consume a lot of attention in the AI ethics community, not necessarily because they were the most pressing, but because they were salient and tractable.

Approximately a decade following the Netflix Prize controversy, with the research on differential privacy having matured into a subfield of computer science, a different topic emerged and led to the birth of modern AI ethics. As classification models became increasingly better and capable of handling structured high-dimensional data, they were put to use for predictive tasks in a wide variety of social settings. One of these settings was criminal procedure. Some jurisdictions in the US began to use algorithms to assist judges or magistrates in bail and parole decisions (Okidegbe Reference Okidegbe2022).

Now, one might ask: why bail and parole? The answer is: their criteria are well-defined. In the context of bail, the main (or only) question for the decision maker is flight risk: will the defendant show up at trial? For parole, the main (or only) question for the decision-maker is the risk of recidivism: will the defendant commit another crime within some fixed time period if they are released? Given these well-defined questions, one can ask: if an algorithm is used to assist in answering them, will that algorithm be biased or unfair towards marginalised racial, gender, or ethnic groups?

There was one algorithm in particular known as Correctional Offender Management Profiling for Alternative Sanctions (COMPAS), which had been licensed to certain US courts for use in recidivism rate prediction (Brennan and Dieterich Reference Brennan and Dieterich2018). A popular magazine, ProPublica, gathered data on its use through Freedom of Information Act (FOIA) requests and published a feature-length article arguing that the COMPAS recidivism algorithm is in fact biased (Angwin, Mattu and Kirchner Reference Angwin, Mattu and Kirchner2016). In many ways, this article significantly spurred the birth of AI ethics as we know it today.

The authors defined racial bias as disparity in terms of false positive mistakes (i.e. the rate of people from each group predicted to recidivate that in fact did not recidivate). They then analysed their FOIA data and found that false positive rates among Black defendants were higher. This finding immediately raised a number of tractable questions among academics across both computer science and philosophy: should we define bias in terms of observed outcomes? If so, how should we deal with the missing data problem (people who are kept in prison cannot be observed to recidivate)? Can we define different fairness metrics? Is ethical bias really reducible to false positive disparity? The literature subsequently exploded. Again, social circumstances were right, and the problems tractable, for these issues to be properly examined in an academic light.

What has driven AI ethics thus far is familiarity – of tools and of topics. Computer scientists know how to work with optimisation and dynamic programming techniques, for instance. Philosophers know how to debate one’s definition of ‘fairness’. Yet just because these are familiar questions calling for familiar tools does not mean that they are the most consequential. The geopolitics of AI may be too complex and multifaceted to reduce to a computational or conceptual problem within a vacuum. Yet we will attempt to address its complexities carefully.

The United States: A Path to Isolation

In the US, domestic legislation has been used primarily to increase the likelihood of a ‘US Victory’ in a perceived winner-take-all AI race (in Section 4, we will argue directly against the framing of inter-state AI dynamics as a race). This has been accomplished in roughly three episodes: the 2022 CHIPS and Science Act, the tightening of the export controls regime, and the 2025 White House AI Action Plan.

In 2022, Congress enacted, and former US President Joe Biden signed, the CHIPS and Science Act, 136 Stat. 1366 (CHIPS Act 2022), appropriating 39 billion USD to bolster domestic semiconductor manufacturing capacity and another 11 billion for semiconductor R&D through government research labs and in collaboration with industry partners. The CHIPS and Science Act was motivated by “[c]oncerns about a decline in the US position in semiconductor manufacturing and technology and potential rise of China’s industrial and technological competitiveness” (Blevins et al. Reference Blevins, Grossman and Sutter2023). Indeed, the Act explicitly states “[i]t is the sense of Congress that … technology development is critical to long-term competitiveness with near-peer competitors, including China” (Sec. 103(d)) and defines several key provisions by reference to the need for competitiveness with China and other “countries of concern”, defined to include North Korea, Russia, and Iran (Sec. 103(a)).

The US export controls regime has been expanding for almost a decade. Beginning roughly with the 2018 Export Control Reform Act (ECRA), which authorised the President to control exports for national security and foreign policy purposes, 132 Stat. 1636 (ECRA 2018). ECRA reversed a decades-long trend of export control liberalisation, which began with the US-Soviet detente of the late 1960s (Casey and Kerr Reference Casey and Kerr2021), providing the statutory basis for the Export Administration Regulations (EAR), a set of rules and guidelines administered by the US Department of Commerce’s Bureau of Industry and Security (BIS), including the Commerce Control List (CCL), which identifies commodities, software and technologies subject to export controls for national security and foreign policy reasons. In October 2022, BIS significantly updated the EAR with the interim final rule 87 F.R. 62186 (BIS 2022), with the expressed intent of ‘restricting China’s ability to purchase and manufacture high-end chips’. Interim Final Rule (IFR) 2022 targeted ‘advanced computing and semiconductor manufacturing items,’ effectively prohibiting the export of Nvidia’s A100 and H100 GPUs, which were the flagship high-performance models at the time.

Finally, the White House AI Action Plan, unveiled during the second Trump administration, has reinforced the isolationist path. After exclaiming that “the United States is in a race to achieve global dominance in artificial intelligence,” the document adds that “[w]inning the AI race will usher in a new golden age of human flourishing, economic competitiveness, and national security for the American people.” The document goes on to emphasise that “[w]e need to establish American AI – from our advanced semiconductors to our models to our applications – as the gold standard for AI worldwide and ensure our allies are building on American technology.” While this is admittedly mostly rhetoric and purple prose, it is worth studying to understand the kind of tone that the White House sought to set between 2022 and late 2025. With that said, it is well worth noting that, as of this writing, the US seems to have gradually loosened some of its more restrictive measures, allowing, for example, Nvidia’s H200 processors to be exported to China (Babic and Wong Reference Babic and Wong2026).

The approach in the rest of the world is decidedly different, and perhaps best characterised by a cautious interest in multilateral cooperation together for tit-for-tat responses to US restrictions and trade barriers.

European Union

In the EU, while Regulation (EU) 2021/821 establishes a unified regime for export licences, focusing in particular on so-called dual-use items (those with both civilian and military applications), such as drone and surveillance equipment, other initiatives suggest an increasing interest in multilateral cooperation. As Lu (Reference Lu2025) notes, “[t]he EU has actively advocated for the establishment of global AI ethical guidelines and governance frameworks within multilateral platforms such as the United Nations (UN), the Organization for Economic Cooperation and Development (OECD), the Group of Seven (G7), and the Group of Twenty (G20)” and the EU “actively supports the UN Secretary-General’s proposal to establish a ‘Global Artificial Intelligence Governance Agency’” (4).

China

The tone in China is likewise decidedly different from that of the US White House AI Action Plan. For example, the Ministry of Foreign Affairs (MFA) released the Position Paper of the People’s Republic of China on Strengthening Ethical Governance of Artificial Intelligence in November 2022, stating that “China is committed to building a community with a shared future for mankind in the domain of AI” and “finds it important to enhance the understanding of all countries on AI ethics” (MFA 2022). The document also states that “China calls on the international community to […] work to formulate a widely accepted international AI governance framework.”

Meanwhile, in October 2023, President Xi Jinping released the Global AI Governance Initiative at the Third Belt and Road Forum; the document emphasised both the need for global cooperation and recognised the value of including developing countries in shared governance, stating “[e]fforts should be made to conduct international cooperation with and provide assistance to developing countries, to bridge the gap in AI and its governance capacity” as well as that “[w]e support discussions within the United Nations framework to establish an international institution to govern AI, and to coordinate efforts to address major issues concerning international AI development, security, and governance” (MFA 2023, see also MFA 2024, for a similar sentiment).

Of course, as some have pointed out, China’s Global Initiatives exhibit “dexterous ambiguity” (Olander and Wong Reference Olander and Wong2025), a feature that grants its decision-makers flexibility when designing and adapting their policies based on incidental and circumstantial needs. Likewise, we did see tit-for-tat responses to US noncooperation in China via both tariffs (Race and Espiner Reference Race and Espiner2025) and exports of rare earth minerals (Swanson and Tobin Reference Swanson and Tobin2025). It is imperative to discern between rhetoric and facts. The following sections will develop a framework for avoiding getting stuck in such noncooperative cycles.

The Responsibility Gap Challenge

Yet while there are instances of interactional risks where the LM can be readily applied, it is not sufficient for the responsibility assignment over many interactional risks stemming from AI in geopolitical settings. This is in part because of the Responsibility Gap Challenge (Mittelstadt et al. Reference Mittelstadt, Allo, Taddeo, Wachter and Floridi2016; Amoroso and Tamburrini Reference Amoroso and Tamburrini2019; Braun et al. Reference Braun, Hummel, Beck and Dabrock2020). Consider this challenge the normative equivalent of what legal theorists dub the ‘liability gap’ in the emerging field of AI regulation – the difficulty of assigning compensatory burdens and related costs in relation to harmful outcomes resulting from AI, due to the “algorithmic opacity, the distributed architecture of AI systems, and the systemic and diffuse nature of AI-related harms” (Tran Reference Tran2025).

An example of a liability gap would be a case where a semi-autonomous Food and Drug Administration (FDA)-cleared AI medical device makes a mistake. There is a gap because (a) the device is not licensed to practise medicine and hence does not have malpractice insurance, while (b) its semi-autonomous nature may shield the residing physician from liability.

We propose the following statement of the challenge:

• C1: AI agents can cause harmful outcomes in the geopolitical space.

• C2: The human source of responsibility for those outcomes is unclear.

• C3: Holding AI agents responsible in the traditional sense cannot be an adequate substitute.

C1 has been tentatively established.

C3 is interesting. On the one hand, there are objects which we clearly do not hold responsible: we do not hold a rock ‘responsible’ for tripping a jogger. On the other hand, certain accounts of AI agency posit that only upon their reaching a level of sentience (Wong Reference Wong, Tao, Kam-por and Camus2026), self-awareness (Bojić, Stojković, and Marjanović Reference Bojić, Stojković and Jolić Marjanović2024), or capability to consistently and reliably execute longer action sequences (Ahn et al. Reference Ahn, Brohan, Brown, Chebotar, Cortes, David and Finn2022; Yang et al. Reference Yang, Yang, Wen, Fu, Mei, Wu and Cai2025; Fung et al. Reference Fung, Bachrach, Celikyilmaz, Chaudhuri, Chen, Chung and Dupoux2025), could it be reasonable to attribute partial responsibility to said entities, for they qualify as agents. Yet this does not appear at all likely in the foreseeable future.

First, the causal sequence of the harms will often be epistemically difficult to identify and ontologically diffuse. For example, as Gerke et al. (Reference Gerke, Babic, Evgeniou and Cohen2020) argue, AI models in medicine and public health are best construed as continuously evolving systems, not products. This is equally true for complex AI systems used for geopolitical objectives. Within such systems, there are innumerable moving parts: for example, we may have multiple sources of data from many different vendors, multiple AI models, robotic devices on which those models are implemented: sensors, cameras, gyroscopes, etc. for processing real-time evidence, many subsidiary digital assistants, and many human inputs for guiding the system.

When a mistake is made, it may be essentially impossible to determine which component of the system has ‘caused’ that mistake, in the sense of but-for and proximate causation, as required by product liability in tort law or the aforementioned accounts of moral responsibility. This is not merely an epistemic challenge. Consider the various examples from the imperceptible harms literature (e.g. Parfit Reference Parfit1984; Nefsky Reference Nefsky2012): the ‘exact’ contribution made by individuals who are distantly contributing towards interactional risks – for example, those voluntarily participating in the census generating Bernie’s data inputs – given the overdetermination of the resultant outcome just cannot be identified. To hold all these individuals morally responsible, even granting that some of them are excusable in virtue of their ignorance, appears unjustified.

Second, our current approach to liability may not be sufficient for holding AI systems liable. If Bernie predicts that an asylum claimant would commit a crime if admitted, that claimant is then refused entry, and we will never observe them committing a crime in the country. Asylum claimants could come from many different countries, and they will in general not have the resources to seek legal recourse, nor are they likely to have standing in the country they are petitioning. Consequently, Bernie’s produced harms (e.g. of access denial) are diffused across many people all over the world, none of whom have the means or ability to address them.

We do not currently have a legal regime whereby a system like Bernie could sue or be sued in court, and even if Bernie could be sued, it would be judgement-proof (i.e. insolvent/incapable of paying). Bernie also cannot be punished in most conventional senses – we cannot imprison or fine it (though, of course, we could shut it off). Some of these limitations could change in the future – we could require AI systems to have liability insurance, or bank accounts opened in their names. Yet at present, we are nowhere near that kind of immersive digital environment. The law still largely operates by looking for humans to blame. Practically, present circumstances lend strong support to C3.

Now we turn to C2. The conventional LM is inadequate in cases where clear, intentional perpetrators of harms cannot be found (e.g. the government employee interfering with Bernie’s input). Indeed, there exist various reasons for which few human agents can be deemed blameworthy for most interactional risks.

As suggested by Kiener (Reference Kiener2024), one way of articulating this premise is to say that few or no human agents satisfy the following two necessary conditions for moral responsibility – the agent must be capable of (i) meaningfully influencing (Control Condition) and (ii) knowing the range of the possible outcomes (Epistemic Condition), of the AI model.

These two conditions are unlikely to be met for most agents operating AI in the complex global geopolitical environment. For Bernie, the census conductors, the computer scientists training the LLM, and the bureaucrats drawing upon its outputs to devise specific policies on refugees and tightening of immigration controls may each hold a ‘jigsaw piece’ to the overarching puzzle. Yet individually, none of them is sufficiently well-positioned to meaningfully influence the ultimate outcome. To what extent, and for what exact contribution, should we hold agents such as Bernie’s original manufacturers, the politicians pressing for Bernie’s adoption in ‘keeping out the migrants,’ or the citizens voluntarily submitting their data as morally responsible for Bernie’s recommendations? These are just a few sources: there will also be multiple data vendors, multiple development companies contributing different components towards the overall system, and different human agents using Bernie for a particular decision (e.g. one employee may use Bernie for the initial assessment, while another may use it for a review of that assessment).

There is a practical tendency to err on the side of obviousness and prioritise assigning blame to the causally nearest human operator, for example, perhaps the immigration officers who turn Bernie’s prescriptions into concrete policies. Yet as Elish (Reference Elish2019) notes, per a naïve version of LM, the nearest operator – just as the front of a crashing vehicle – could fall into the “moral crumple zone” and be unduly scapegoated for a long and elaborate causal chain of events preceding their decision. Likewise, Gerke et al. (Reference Gerke, Babic, Evgeniou and Cohen2020) suggest that the system view undermines the temptation to lay blame on the nearest human operator. To the extent that they can be clearly identified using the control condition, these obviously responsible agents may not be, all things considered, the most responsible – only the most exposed in the relevant causal chains.

To illustrate this in light of the AI-powered Lavender system allegedly used by Israel in the ongoing war in Gaza in targeting suspected Hamas elements and affiliates, we must ask – to what extent should the operators using the system to identify terrorist suspects in the Israel Defense Forces (IDF) be held responsible for the eventual targeted assassination of these individuals? What of their superiors, or the policy leadership of the military, who decided that such an AI-enabled system would be conducive towards the political objectives – as stipulated by the democratically elected government?

Further, many of the individuals involved in these scenarios may struggle to understand the extent to which they have made active contributions to and the nature of the harms arising from AI deployment. This problem is compounded by the black box nature of AI systems like Bernie, meaning that human operators will be incapable of understanding the reasons for the recommendations of such systems (cf. Babic and Cohen Reference Babic and Cohen2023). The epistemic condition is unlikely to be fulfilled on the part of individual employees of large, sprawling corporate (e.g. Meta and OpenAI) or governmental complexes (e.g. the Pentagon or State Department in the US, or the Chinese Academy of Sciences in China).

We could set aside the normative discussion by postulating strict liability on legal and prudential grounds, assigning it to certain actors (both individuals and groups alike), for example, an entire manufacturing unit can be held liable for the high incidence rates of false positives/negatives when it comes to the AI-powered drones, and tasked with the responsibility of compensating injured victims and loved ones of killed victims. This move can theoretically decouple the question of legal and practical responsibilities from the more complex and controversial debate over moral responsibility assignment.

Yet this carries a high degree of arbitrariness: we could then pick out any particular employee, for example, from a manufacturing unit, and prosecute them at will. It also fails to answer questions concerning how such liabilities ought to be distributed among individual members of these collective units. This seems unjust. As such, the Responsibility Gap Challenge remains a potent concern.

Interactional Risks and Structural Injustices

Per Young (Reference Young2011, 52), structural injustices refer to instances where:

1. 1. “Social processes put large groups of persons under systematic threat of domination or deprivation of the means to develop and exercise their capacities.”

2. 2. “These processes enable others to dominate or to have a wide range of opportunities for developing and exercising capacities available to them.”

3. 3. These processes comprise “a consequence of many individuals and institutions acting to pursue their particular goals and interests, for the most part within the limits of accepted rules and norms.”

Wong (Reference Wong2025a) terms these, respectively, the Domination-deprivation, Enrichment, and Non-individuation Conditions. Existing literature has already identified and analysed structural injustices posed by AI in certain contexts (Himmelreich and Lim Reference Himmelreich, Lim, Bullock, Chen, Himmelreich, Hudson, Korinek, Young and Zhang2022) – education (Madaio et al. Reference Madaio, Blodgett, Mayfield, Dixon-Román, Holmes and Porayska-Pomsta2022), healthcare (Braun et al. Reference Braun, Bleher, Hille and Krutzinna2023), and algorithmic decision-making more broadly (Browne Reference 78Browne, Browne, Cave, Drage and McInerney2023). Browne, in particular, deserves particular credit for her efforts in incorporating structural injustice into the context of AI injustices. Indeed, Browne (Reference Browne2025) is seminal in devising a robust account of political responsibility and the governance of technologies in the era of AI. Likewise, Smart and Kasirzadeh (Reference Smart and Kasirzadeh2025) rely on Young’s SCM to develop a social notion of AI explainability.

However, relatively little has been said on how Young’s work can be applied and developed in the context of international conflicts or geopolitical tensions. As such, we can fruitfully apply the theory of structural injustice to the subset of interactional risks identified thus far.

One may wonder whether a risk is an injustice. In the context of interactional risks, we take it to be so for two reasons. First, these risks can induce reasonable agents to change their behaviours and adapt their preferences accordingly (e.g. a soldier turning towards an AI model for recommendations on whether to open fire). Second, such risks give rise to the potential for domination – as Pettit (Reference Pettit1999) puts it, domination consists of a power asymmetry between a dominator who can arbitrarily exert influence over the dominated. In any case, assuming risks can amount to injustices, many of the interactional risks we examine in this section would fit the three proposed conditions:

The SCM and Responsibilities for Interactional Risks

The SCM, as originally advanced by Young (Reference Young2011, 104–113), provides an alternative to the LM; it comprises the following three features:

1. 1. Orientated around structures, not individuals. For us to interpret and “criticise the way that many individuals and groups face too limited and unsavoury sets of options”, what is needed is an “account of large-scale systemic outcomes of the operations of many institutions and practices that constrain some [… and] enable others” (Young Reference Young2003, 4); individuals who are ‘connected’ to such structural injustices should be held responsible.

2. 2. Primarily ‘forward-looking’. Responsible actors should focus their energies and attention on devising solutions that can be implemented to tackle injustices that have yet to manifest – namely, to prevent them altogether, or mitigate against their worst effects. The kind of responsibility she envisions here is non-blame-centric – indeed, she suggests viewing it as political responsibility.

3. 3. Anchored in fundamentally shared and non-isolatable responsibilities. The responsibilities arising from the SCM are essentially shared across all actors acting within the sphere permitted by rules and norms of the society.

Two constructive amendments should be incorporated into this account.

The first is a clarification by McKeown (Reference McKeown2018) concerning the meaning of “connection”. She is concerned that the vagueness of social connection could render the concept of “connection” over-inclusive, thereby assigning responsibilities to actors who are tangentially and insufficiently related to the injustice. Drawing a distinction between two kinds of causation – attributive causation, which tracks the LM, and explanatory causation, where background conditions causally contribute towards particular injustices, she construes connection as the reproduction of structures amounting to man-made explanatory causes of injustices, which involves a distinctly lower threshold of proof than conventional causal accounts of liability.

That is, an individual – so long as they are behaving with a modicum of intentionality and bare awareness concerning their actions – can reproduce a background condition even in cases where such a condition is overdetermined. For instance, in choosing to purchase a branded shoe made by a sweatshop worker, a particular customer is contributing revenue towards the brand. Even though the marginal revenue would have made no difference to the number of shoes the company produces, the customer has left their mark on the exact sales total. They are, to that extent, reproducing (even if not causing, per se) one of the conditions for the continued employment of sweatshop labourers. The venture capitalist backing a fledgling AI start-up specialising in lethal autonomous weapons is reproducing the background condition of AI ‘enhancements’ to killing technologies, even if some other investor would have entered the picture in their absence.

The second tackles the objection raised by theorists including Nussbaum (Reference Nussbaum and Young2011) and Reiman (Reference Reiman2012), concerning the “free pass” that Young’s SCM apparently generates. If the SCM indeed does away with all notions of retrospective guilt for past moral transgression or ongoing failure to redress past injustices, Nussbaum fears that individuals would “get a free pass indefinitely” (xxi). The worry is that the SCM has no teeth: in the absence of real accountability mechanisms, there would be neither motivation nor an appropriate acknowledgement of the failure on the part of these non-compliant actors.

Where responsible agents (under the SCM) fail to discharge their responsibilities stemming from structural injustices, they can be criticised (Young Reference Young2011, 165). Such criticism can be meaningful, even while it does not amount to outright blame – for example, we can criticise the politicians in the city of Toronto for failing to provide adequate public transit, even without going so far as to say they are blameworthy and hence punishable for not doing so. The focus of our criticism is their forward-looking responsibilities: they ought to strive to improve the city’s public transit. Indeed, such criticism may well entail that if they do not do so, they are less likely to be elected again – this is the way in which the SCM can still have ‘teeth’, despite the fact that it floats independently of traditional notions of blame, which stand at the core of the LM.

More generally, we advocate a pluralist view of responsibility for injustices (Wong Reference Wong2025b). Any robust account of justice should make room for both forward-looking and backward-looking conceptions. In instances where perpetrating, complicit, or even culpably negligent actors can be easily identified, the LM should be applied. Yet in cases where such identification is not possible, we suggest that the SCM, as developed earlier, ought to be utilised.

We will call the SCM, with these two important clarifications, the SCM*.

Limitations to the Neo-Rawlsian Framework

There are three reasons why this simply stated neo-Rawlsian framework is inadequate.

First, while Beitz primarily focuses on natural resources – on grounds that their distribution across different countries is geographically determined and hence arbitrary – the same cannot be said of AI, which requires investment of considerable artificial resources to develop, produce, and sustain – such as the development of large data centres and related infrastructure (e.g. Heating, Ventilation, and Air Conditioning (HVAC) systems). That these goods are created, reproduced, and transformed by human users complicates the picture of justice – especially if we recognise the existence of property or more generally the private rights of ownership and commercialisation of AI models. If we heavily invest our time, energy, and capital into building a house, we do not appear to be bound by overriding moral obligations to share it, even if doing so promotes distributive justice (Nozick Reference Nozick1974).

Second, as highlighted by nationalist theories of various kinds (Miller Reference Miller1995, Reference Miller2007; Riker Reference Riker2009), this broadly cosmopolitan distributive justice worldview does not make sufficient room for the ethical significance of nations – especially for the individual citizens that comprise such collectives. Miller (Reference Miller1995), for instance, offers a defence of moderate ethical nationalism. He argues that “relations between persons are part of the basic subject matter of ethics, so that fundamental principles may be attached directly to these relations” (5), including the relations between teachers and students, neighbours within a village, or start-ups of like-minded co-workers. Such commitments in turn lead one to identify and perceive substantial value in their ties with their co-nationals, thereby giving rise to the special obligations we owe towards one another. These obligations do not obviate the relevance of the interests of foreign non-nationals; but they do mean that we can sometimes prioritise compatriots under conditions of scarce resources (e.g. distributing a limited supply of vaccines in a pandemic) (Zhang Reference Zhang2024).

Third, global AI justice should not be primarily centred around distribution. When we posit that individuals should have ‘equal opportunities’ to succeed through AI, what exactly are we speaking of here? As Anderson (Reference Anderson1999, 288) puts it, distribution-orientated accounts of global justice focus excessively on the treatment of “divisible, privately appropriated goods, such as income and resources, or privately enjoyed goods, such as welfare”. Given such a fixation, how can we account for the normative status of the kinds of political and legal constellations, social norms, and values that both contribute towards and are shaped by AI?

Consider, for example, an LLM used by a large, dominant state to create purposefully discriminatory content towards particular nationalities and ethnicities. A purely distributive account may suggest that this model ought to be made equally available to smaller and less resourceful states as well, which of course misses the main point – namely, that the outputs are problematic, not that the opportunities to access it are unequal.

Indeed, justice is more than a merely distributive exercise. It is a fundamentally justificatory exercise, one in which individual agents have claims to normatively meaningful defences, excuses, or counter-defences concerning the events befalling them. The framing of distribution is too reductionist, precisely because it pays little attention to interpersonal relations, which are no less important – if not more so – for individuals, than bundles of resources or access to advantages. The neo-Rawlsian account falls short of meeting this threshold.

Step 1

Some egalitarian theorists hold that egalitarian responsibilities should apply strictly within one’s nation-state – Wellman (Reference Wellman, Wellman and Cole2011, 75) observes that “the coercion states invariably employ is nonconsensual, and, as such, is extremely difficult to justify […] Thus, states are justified insofar as they provide vitally important benefits (i.e. protecting their constituents’ human rights) that would otherwise be unavailable.” In receiving benefits from one’s state and both partaking in and being impacted by the same system of governance, co-citizens of the same state thus accrue egalitarian responsibilities towards one another. Wellman does not believe the same could be said of non-citizens.

Yet this coercion-based justification against broadening the scope of egalitarian concern runs into an objection launched by, for example, Abizadeh (Reference Abizadeh2008), Nath (Reference Nath2011, Reference 86Nath, Fourie, Schuppert and Wallimann-Helmer2015), Sharp (Reference Sharp2022), and, to an indirect extent, Pogge (Reference Pogge2008). States can exercise coercive influence – if not control – over foreign populations, via border-immigration control, economic sanctions (Pattison Reference Pattison2015), trade and investment policies, and even the propagation of affirmative or detractive discourses in relation to these populations’ governments.

Indeed, Nath (Reference Nath2011) compellingly addresses the ways in which inter-state inequalities can impact states’ abilities to “shape the terms of cross-border interaction”. We care about eliminating inequality – and the advancement of broader equality – between individuals of the same state, because their “subjection to institutional terms that account for such interconnectedness is, for the most part, unavoidable and nonvoluntary”. The dearth of choices in the domestic setting gives rise to relational egalitarian demands; the same could be said of globally coercive structures – from the sweatshop economy to large multi-national corporations, from the unequal playing field of trade negotiations to the weaponisation of capital restrictions and exporting of critical raw materials, for example, rare earths and lithium, copper, and nickel, for geopolitical ends. Coercion extends beyond borders; Wellman’s argument is therefore unpersuasive.

There are three reasons for relational egalitarians to endorse GRE. First, there is no inherent reason to limit the scope of justice to simply within one’s national boundaries. That there exist normatively significant distinctions between how one relates to a fellow citizen, and to a non-citizen, does not eliminate the relevance of the latter’s claims on one’s duties towards them. Accepting that our responsibilities towards co-citizens are stronger does not thereby imply we accrue no responsibilities towards non-citizens. The Wellman-Nath dialectic above is replicated across a number of purported differentiating factors between citizens and non-citizens – for example, contribution towards welfare (non-citizens also contribute greatly towards one’s domestic economy, life chances, and wealth), or strength of association (select citizens may associate more closely with non-citizens in ways they find meaningful). See, for instance, Ip (Reference Ip2016), who advances a relational view of egalitarianism anchored in the normative significance of equal relations as the basis for global justice.

Second, most human beings are trivially involved in activities that would render them connected to a myriad of institutions with the potential to exert force upon them, cutting across nation-states and legal-territorial boundaries. This range of institutions is bolstered by trans-border interconnectivity (Pogge Reference Pogge2008).

Third, the spread of globalisation throughout the two decades between the Berlin Wall’s fall and the Global Financial Crisis led to the destabilisation of once-rigid conceptions of national citizenship and identities among many, with a survey of over 20,000 people from eighteen countries in 2015–2016 revealing that 51 per cent of respondents reported seeing themselves more as ‘global citizens’ than ‘citizens of their country’ (GlobeScan 2016).

This phenomenon was particularly apparent in countries with emerging economies and large populations, such as India, China, and Nigeria. While subjective attitudes are normatively non-determinative, they do offer a glimpse into the emergence of an international order that both presents an alternative to and contextually supersedes the nation-state as the definitive anchor for individuals’ self-understanding. In short, the nation-state is by no means the only political community that binds individuals’ self-identification.

Step 2

The next claim is that the crux of GRE lies with individual citizens, as opposed to nations or states. Some may advocate a key role for nations and states within the RE conversation, given their functions as the primary political unit of social organisation, resource distribution, and domestic and international representation. Bernard Williams (Reference Williams and Hawthorn2005, 3, cf. Sleat Reference Sleat2010) proposes the first political question, one that “is the condition of solving, indeed posing, any others”, with the securing of “order, protection, safety, trust, and the conditions of cooperation”. We have also seen convincing reasons for which nations should matter in our understanding of global justice. If an inter-national egalitarian theory makes no room for the state or nation, how complete could it be?

In response, a distinction should be drawn between the foundational and the instrumental. The instrumental significance of nations and states can be embraced by those subscribing to the view that foundationally, it is the rightful claims of individuals, not states or nations, to RE that ought to motivate our normative prescriptions. Advocating that states or nations are entitled to relationally egalitarian treatment apparently requires confronting more contentious questions on group and collective agency, which would be beyond our scope. Regardless of how geopolitically powerful or weak, wealthy or poor, large or small a state is, an individual is pro tanto entitled to being treated as equal by others from other states.

Indeed, achieving – or approximating – relational equality between individuals of different states, may practically require the elimination of some inequalities between their very states. States are instrumental intermediaries in determining the range of opportunities, the legal and regulatory foundations, and the average level of resources accorded to their own citizens. A world where powerful states can radically suppress and exclude weaker counterparts is unlikely to be one where the citizens of both sets of states can relate to one another as equals (Fanon Reference Fanon1963; Pogge Reference Pogge2008).

Step 3

Why, then, should GRE apply to AI?

Consider the following two positions:

1. 1. A: AI is uniquely distinct from most other spheres of human activity, to which global RE applies; as such, AI should not be governed by principles of GRE.

2. 2. B: AI is not uniquely distinct from most other spheres of human activity, to which GRE applies, in a way that renders it exempt from RE considerations.

B is more plausible. Unlike the family, which may reflect deeply intimate attachments integral to individuals’ self-conceptions as autonomous moral agents, anchored in their authentic personal values, commitments, and relations (Williams Reference Williams1973, Reference Williams1981), the sphere of global AI opportunities is one that definitively does not involve prospective violations of integrity for the individual. The pursuit of GARE does not require that one give up vitally important personal ground projects.

One might reply that AI should be integrated into our understanding of GRE. Yet precisely because it can be subsumed accordingly, relational inequalities within AI can be offset through a heightened push for egalitarian transformations along other, non-AI dimensions. Perhaps the fact that a country X’s government causes their citizens to not relate with citizens in country Y as equals, through imposing export bans on GPU chips, data, and AI talent outflows to Y, can be offset by X’s government including Y’s citizens within decision-making processes concerning developmental loans and financing. The fungibility with other spheres of justice cannot sufficiently obviate the specific case for GARE for two reasons:

First, RE is a holistic concept. Suppose a philosophy department chair Joseph opted to systematically exclude a colleague, Susan from all social events, public discussions and dialogues, and inter-departmental meetings – on dubious grounds that she is not ‘suitable’ for such engagements. Yet in his capacity as a senior academic, Joseph offers to co-author an important paper with Susan, crediting the latter fully for her conception by making her the first author. We would not deem the respectful and egalitarian treatment in the paper-writing context as adequate in offsetting the disrespectful and inegalitarian treatment in the administrative context.

Second, to the extent that it is possible to enact principles of RE within the sphere of AI – through the measures outlined earlier – there exists no particular value in the preservation of such relational inequalities, such that they can only be offset indirectly. Why should we not tackle inegalitarianism across both AI and non-AI spheres?

Contingently, given AI’s increasing importance and necessity in individuals’ lives, the pursuit of greater relational equality in AI development, diffusion, and usage could well require that we prioritise AI in the proverbial order of priority. A world where access to AI usage, but also capacity to shape norms, regulations, and accountability, is increasingly stratified, is a world where individual citizens are unlikely to see and relate to one another as equals, at least in relation to their ability to derive benefits from the digital economy.

Eliminating Status Hierarchies

An oft-highlighted component of RE is its attempt to eliminate oppression, where the victim suffers from “some inhibition of their ability to develop and exercise their capacities and express their needs, thoughts, and feelings” (Young Reference Young1990, 40). This can be traced to its broader commitment to tackling status hierarchies (Voigt Reference Voigt2020). Anderson (Reference Anderson and Estlund2012) argues that RE is opposed to problematic hierarchies of domination, esteem, and standing.

While domination (and oppression) have been extensively discussed as manifestations of structural injustices, more should be said about esteem – a form of recognition granted by others based on the individual’s virtues, actions, or other admirable merits – and standing – the extent to which the individual is an agent of concern, respect, and accountability in the eyes of others. Only when undue inequalities across esteem and standing, along with structural relations giving rise to domination, are minimised can relational equality be attained.

In the context of AI, where actors’ production and regulation of AI manifest in overt domination (per Section 2) or unwarranted inequalities in esteem and standing, there thus exists a strong GRE case on two levels: trivially, for these actors to change their actions; and more substantively, for third parties to positively intervene in limiting such inequalities. Refraining from contributing towards structural injustices alone is insufficient.

For example, if an AI conglomerate backed by a dominant state amasses copious volumes of human capital and technology, such that it effectively crowds out alternative competitors on the global market, even if it ends up deploying its resources towards wholly benign ends, its monopolistic nature enables it to exert dominant market influence. Such an immense power asymmetry between the company and its rivals can be conceptualised as domination, reflecting a problematic hierarchy of esteem and standing – not only relative to its fledgling and struggling competitors, but also the millions of consumers beholden to the whims of its leadership and the government behind it. Regulators across the world would have the pro tanto responsibilities to pursue anti-trust regulation, to preclude the conglomerate from accruing extremely concentrated power. As of 2022, just 100 companies – mostly in the US and China – comprised 40 per cent of global AI research and development (UNCTAD 2025). Leading players in the US and Europe, including xAI in the former and Mistral AI in the latter, have come under growing scrutiny for their ‘weak’ risk management practices, despite having amassed significant capital and market influence (Chow Reference Chow2024).

The elimination of status hierarchies would of course have a distributive component. When citizens residing in Country X are denied access to instrumental components for cutting-edge LLM development, or integral technologies synthesising AI and infrastructural planning, they are certainly unlikely to relate to their more resourceful counterparts in Y and Z as equals. They would be subjugated to another manifestation of the long-standing digital divide between the Global North and South (Heeks Reference Heeks2022).

Some authors (e.g. Zarakol Reference Zarakol and Zarakol2017; Bell and Wang Reference Bell and Wang2020) point to the value of hierarchies in shaping the international order. Hierarchies provide stability of expectations and support enduring leadership on the part of great powers. Similarly, one could argue that AI requires global hierarchies in order to be governed effectively, lest more unbridled competition between a large number of AI state and non-state powers take over in a world where there exists no overwhelmingly dominant player in its development or regulation.

Yet limiting hierarchies in the global AI space does not imply endorsing a state of anarchy across states and individuals within them. Nor, indeed, should we commit to the view that all relational hierarchies must be eliminated. Benign and just hierarchies can and do exist (Bell and Wang Reference Bell and Wang2020). However, such rationale cannot be invoked to defend the extensive restrictions on the exporting of cutting-edge semiconductors, arbitrary stipulations targeting AI investments, and radical financial capital inequalities as witnessed in the status quo, which leave the less advantaged parties at the whims of the more advantaged.

In the face of AI status hierarchies without a compelling, normatively sound justification, all parties with the capacity to do so should work towards ameliorating them. This could take the form of collectively resisting or circumventing geopolitically motivated export controls, advocating more transparent sharing of model parameters and weights in LLM training – for example, open-source AI development (Bisht, Yee and Roberts Reference Bisht, Yee and Roberts2025), and providing socioeconomically disenfranchised groups within countries with both access to AI and AI literacy training. The latter connects to a broader point about capacity-building – individuals cannot relate as equals with one another if they possess radically divergent capabilities, as well as access to the minimum threshold of functional knowledge required to derive benefits from AI.

Equalising Decision-Making Power Over Global AI Regulation

An alternative focus adopted by some RE theorists concerns ensuring that all agents possess equal decision-making power – that is, that their preferences and interests are weighed equally, considered carefully, and incorporated duly in the process by which significant decisions are made. Scheffler (Reference Scheffler, Fourie, Schuppert and Wallimann-Helmer2014, 25) conceives of relational equality as when individuals conduct their relations “on a footing of equality [such that] each person accepts that the other person’s equally important interests [… including] the person’s needs, values, and preferences – should play an equally significant role in influencing decisions made within the context of the relationship”.

He further substantiates this by calling for reciprocal acknowledgement of all parties in the relationship of one another’s interests as “playing just as significant a role as [their own] in constraining [each other’s] decisions”. Such acknowledgement should give rise to dispositions that bind their joint decisions, through serving as an egalitarian deliberative constraint (Elford Reference Elford2017). In the context of AI, we could see this evolve through discussions over global regulations and policies governing its development, diffusion, and use.

GARE thus encourages the establishment of multilateral frameworks in which the voices of small and medium states, but also representative non-state actors, are afforded a voice and endowed with impactful agenda-setting powers regarding AI governance. Progressively, the inclusion of more voices in the AI decision-making process can be accomplished through in-house reforms within the UN, World Bank, and the International Monetary Fund (IMF), whereby these states’ representatives are given more autonomous say and political platforms to robustly comment on and advocate AI agenda on their citizens’ behalf. This process could also be triggered by the efforts of more progressive lobbying and advocacy groups pressuring corporations to open up to genuinely consulting and incorporating views of representatives of the Global South and under-represented political communities in their development processes.

GARE can, therefore, go some way in identifying key expectations benchmarks for which actors in the global AI space ought to be accountable, specifically in areas in which such global deliberation is procedurally and instrumentally significant, including the management of AI use in military conflicts, combating AI-generated misinformation and hateful content, and tackling the use of AI-based algorithms in public health and humanitarian contexts. Some of these are issues tentatively discussed under the UN AI Panel and Dialogue (Resolution A/RES/79/325 (UNGA 2025)).

During this process of deliberative inclusion, we may find states’ practically desirable proxies representing their constituencies’ interests internationally. First, it is over-demanding to expect all their constituents to play an active role in steering and contributing towards a structured and outcome-oriented global conversation; second, there is value in having states serve as the initial filtration mechanism for the demands and interests of their citizens. Even in non-democratic states, where popular participation in elections is often lacking, select governments would draw upon consultative processes to aid them in gauging public interests and concerns over AI. As discussed by Sheehan (Reference Sheehan2023) and Zhang (Reference Zhang2024), public opinion plays a partial – albeit important – role in shaping the bureaucratic processes in the Chinese state, as it iterates and devises regulatory frameworks for AI. Notably, non-state actors – ranging from important business associations, corporations, and investors, through to activists and citizens’ collectives with significant interests in AI regulation – should also be given a robust chance to influence global regulatory principles (Li Reference Li2023; Diya Reference Diya2025).

Ensuring an Ethos of Respect with AI Deployment

The third variant of GARE advocates that international AI deployment must develop and maintain a genuine ethos of respect. Wolff (Reference Wolff1998) discerns between two distinct strands of the egalitarian ethos – one is fairness, which, per luck egalitarian intuitions, requires individuals to bear the costs of their choices and for the effects of bad brute luck to be eliminated; the other is respect, whereby everyone should be treated with acknowledgement of their equal innate worth as persons. Wolff suggests that these two values come into conflict, and that in circumstances where individuals’ self-respect is at stake, respect should be prioritised over fairness. In short, both Wolff and Anderson (Reference Anderson1999) believe that luck egalitarianism misses the core of what makes egalitarianism so desirable and forceful: ensuring that individuals are not treated in ways incommensurable with their status as autonomous, self-conscious members of shared political communities.

As applied to AI, there exist clear responsibilities on the part of all capable actors to develop AI in a way that does not undermine the respect of individuals. A generative AI model producing positive stereotypes about ethnic majorities (e.g. Hagen, Huo, and Nguyen Reference Hagen, Huo and Nguyen2025), yet implicitly excluding ethnic minorities from its outputs, may not be contributing towards structural injustices per se, though it can certainly be interpreted as reinforcing disrespectful attitudes towards the minorities.

Consider, for example, the case of Cambridge Analytica. The now-defunct analytics firm drew upon machine learning models to target and influence voters in elections, including the 2016 US Presidential Election. In seeking to optimise the receptiveness of voters towards advertisements, the firm “built an AI based system that was able to automatically test hundreds or even thousands of variations of an ad before deciding which one to place.” (Delcker Reference Delcker2020) The company undoubtedly used AI to wield and manipulate prospective voters’ perceptions of facts (Rubera Reference Rubera2023). The reining-in of AI manipulation and deception (Carroll et al. Reference Carroll, Chan, Ashton and Krueger2023) is thus a foremost priority in the advancement of RE. Indeed, this normative concern also underpins the European rules concerning AI in political advertisement (Stephenson, Rinceanu, and Boverman Reference Stephenson, Rinceanu and André Bovermann2025).

The Security Game

As Nash worked on his dissertation in Princeton, New Jersey, on the other side of the country were two mathematicians recruited by the RAND Corporation in California to conduct social scientific research on military strategy following World War II (Gass and Assad Reference Gass and Assad2005). Merrill Flood and Melvin Dresher were interested in exploring player behaviours in repeated play of non-cooperative and non-zero-sum games, and they designed a game with the payoff structure of what we now know as a Prisoner’s Dilemma. Strategic behaviours in AI development are often analogised to a Prisoner’s Dilemma (e.g. Foley Reference 81Foley2024; Perez Reference Perez2025).

The Prisoner’s Dilemma has been particularly influential in IR, due to its central role in security studies. It is the conceptual model underlying security strategy (Jervis Reference 83Jervis1978) – the idea that in the absence of a central authority, states acting in their own self-interest will pursue individually rational actions that could lead them to a conflict which none of them prefers (cf. Waltz Reference Waltz1979).

In the twentieth century, this model was often applied to the study of nuclear proliferation (or non-proliferation). Suppose we have two players/nation states (e.g. the US and the USSR during the Cold War), who must decide whether to cooperate with an arms control agreement or a nuclear test ban treaty. Each state has two choices: cooperate or defect. Suppose that defections are undetected. Let us call this the Security Game.

How should we order the preferences of the two states in a security game? First, we assume that the states’ interests are symmetric and that the game is played once (repeated play affects the rational strategy as defections can be punished in subsequent rounds of the game (Axelrod and Hamilton Reference Axelrod and Hamilton1981)). We assume that self-interested actors seek the ‘best deal’ for their own nation, so defecting while the other state cooperates is best (we will assign this outcome four units of utility). The opposite of this is the worst (‘you defect/I cooperate’), so we will assign one utility unit to that outcome. Meanwhile, if we both cooperate, we gain the benefit of the non-proliferation treaty or arms control agreement, whereas if we both defect, we return to the status quo, so we will assign the former three utility units, and assign the latter two utility units. Since the Von Neumann-Morgenstern utility function is unique up to affine transformations (Von Neumann and Morgenstern Reference Von Neumann and Morgenstern1953), the choice to use a 1–4 ranking is made without a loss of generality for our purposes. This version of the Security Game therefore looks like a typical Prisoner’s Dilemma (PD), as in Figure 1.

Figure 1

The Security Game. A typical Prisoner’s Dilemma with a unique pure strategy Nash Equilibrium highlighted in bold font.

This is a symmetric game with one pure strategy Nash Equilibrium: namely, the outcome where both states defect. This simple game illustrates what we know well: two states acting rationally may find themselves in a socially sub-optimal outcome. While it is individually rational for each state to defect, it is collectively not ideal for them to end up in that outcome. Yet cooperation is not a Nash Equilibrium – it could leave one purportedly vulnerable to the other’s taking advantage of them. In the context of a public good – something that is non-excludable and non-rivalrous – each player has an incentive to free-ride. This is an instance of a non-cooperative but also non-zero-sum game.

There are aspects of global AI governance that seem to resemble a PD – nation-states, or frontier technology firms, can be modelled as having a similar incentive against cooperating as in a PD game: the more quickly that a company or nation can develop its own infrastructure or its own technological capability, the more economic, security, or pecuniary benefits it can secure. In this sense, cooperating with a competitor seems irrational. The 2025 US AI Action Plan makes this PD framing explicit:

Indeed, former US President Joe Biden implemented strict export controls on leading components required for the development of AI, including high-end chips and SME, 89 F.R. 72926 (BIS 2024a). As part of its rationale, the US Department of Commerce argues that ‘[e]xport controls on semiconductor manufacturing equipment (SME) and related parts and components are central to countering the PRC’s goal of furthering its … production capacity’, 89 F.R. 96790 (BIS 2024b). This framing has jolted Chinese manufacturers to compete by doubling down on their self-sufficient capacities, aggressively recruiting talent, redirecting state capital in a focussed manner, and identifying innovative means to maximise the utility of legacy chips (Zhang Reference Zhang2025).

One can, of course, contend that the drive towards self-sufficiency in fact pre-dated the introduction of US export controls – citing, for example, the ‘Made in China 2025’ policy (Center for Security and Emerging Technology 2022). Yet while 139 billion yuan (roughly 19.5 billion USD) were invested into semiconductors in 2014 in the initial phase of China’s National Integrated Circuit Industry Investment Fund, a further 204 billion yuan (28.7 billion USD) and 340 billion yuan (47.8 billion USD) were invested in the second and third phases (L. Lee Reference Lee2024). During his first term, Donald Trump launched the US-China trade war in 2018 (Bown and Kolb Reference Bown and Kolb2018), with his successor Biden taking a distinctly more proactive and targeted approach in kneecapping China’s semiconductor industry developments (The Guardian 2024). Trump also sought to limit the power of the World Trade Organisation (WTO) by refusing to fill the quorum required for its Appellate Body (AB) to function (Murrill Reference Murrill2019), a policy which Biden also continued (Baschuk Reference Baschuk2023).

Despite persisting challenges concerning corruption and inefficient capital deployment, the drive towards self-sufficiency in China has indeed delivered significant dividends. For instance, China’s Semiconductor Manufacturing International Corporation (SMIC) has raced to increase domestic chip manufacturing capacity in an attempt to meet local demand for high-end chips required by domestic technology giants such as Huawei, Tencent, Alibaba and DeepSeek (Wu Reference Wu2025). In September 2023, Huawei launched the smartphone chip Kirin 9000S, which was manufactured with a 7-nm process by SMIC (Liu Reference Liu2023).

Despite the rhetoric, in many important respects, the PD framing of AI dynamics among nation-states is misleading. We can explain why by focusing on the specific reasons cited by the White House AI Action Plan: the first passage suggests that America must drive the adoption of its systems, infrastructure and standards (The White House 2025, 1). This is not an unreasonable claim for any nation-state to make – its own infrastructure outlay in, for example, small and medium countries, can benefit the investing state as its domestic manufacturers experience a growth in market share, its reputation grows across the world, and its industry standards are widely dispersed. Yet the last passage of the above quotation then goes on to say that America must prevent its adversaries from free-riding on its innovation and investment. The US Department of Commerce makes similar, and indeed stronger claims in 90 F.R. 47201 (BIS 2025a).

If global reliance on the US’s innovation and investment does constitute a significant benefit for the country, as the first passage states, then such reliance on its innovation and investment does not amount to harmful free riding of the sort that drives a tragedy of the commons. Perhaps these claims can be squared by highlighting that such reliance counts as free-riding only if it is performed by adversaries but not if it is performed by allies. Yet this statement does not stand up to scrutiny – the thought that geopolitical adversaries will acquire significantly greater competitive advantages, bargaining capital, or prospective capacity to threaten coercion against the US through US-supplied technologies presupposes that the marginal technological (or productivity) gains can indeed yield a substantial increase in threat capacity and level. Present discussions have yet to point to precise ways in which ostensible ‘adversaries’ – for example, China – have uniquely stood to gain from US-developed AI technologies, in a manner that would substantively undermine American interests. If anything, attempts to curb American chip exports could well undermine the prevalence of the US tech stack as a foundation for global benchmarking and soft influence (Harithas Reference Harithas2025; Villasenor Reference Villasenor2025).

Furthermore, the zero-sum framing or the PD framing assumes that advancements by all AI entities in one economy (e.g. China) must come as a detriment to the interests of the American ecosystem. One could point to DeepSeek as a counterexample. The company made its R1 model open source, and China has not blocked access to it from IP addresses originating in the US; this is despite the fact that OpenAI has restricted access to ChatGPT from IP addresses originating in China (Jiang Reference Jiang2024). The open-source nature of the company’s models means they can be easily applied and extended by state and non-state actors within the US and beyond (Zhang and Liu Reference 92Zhang and Liu2025).

Indeed, one could even argue that our analysis of the PD model points towards the need for cooperation in inter-state AI dynamics. If we view the international institutions necessary for ensuring safe and responsible AI development as the relevant public goods, then a tragedy of the commons occurs where countries defect from such cooperation. The free-riders, in this case, are those who refuse to contribute to multilateral governance, due to which we lose out on both the mutual gains from trade that can happen when technological know-how and materials, standards, and infrastructure are shared, as well as the public goods that can only be achieved through cooperation – for example, the robust and consistent regulation of AI, the safe and transparent deployment of AI, and the reduction of barriers to the opportunities of AI for individuals lacking the financial resources and economic wherewithal to access them. These are, by hypothesis, non-excludable and non-rivalrous benefits that can be obtained from cooperation. Indeed, the ethos espoused by China’s Shanghai Declaration on Global AI Governance, released at the 2024 World Artificial Intelligence Conference, focuses on global cooperation and encouraging shared governance (MFA 2024).

A Modified Prisoner’s Dilemma

The preceding framing would suggest that we need to do what we can to increase the costs of defection and change the Nash Equilibrium of the simplistic PD model – indeed, this is Elinor Ostrom’s approach to solving tragedy of the commons problems (Ostrom Reference Ostrom1990).

For example, the easiest re-modelling illustration would occur when defections are punished (this could occur due to reputational harms in repeated play or through a central enforcement mechanism). If defecting carries a modest penalty of just one utility unit, that is enough to transform the game into one where cooperating becomes a Nash Equilibrium, as in Figure 2 (which we revisit later).

Figure 2

The AI-PD game. A modified Prisoner’s Dilemma where defections are mildly punished with two unique pure strategy Nash Equilibria highlighted in bold font.

In China, the framing of AI governance stands in stark contrast to the US White House ‘winning the race’ narrative. In October 2023, President Xi Jinping announced China’s Global AI Governance Initiative (GGI) at the Third Belt and Road Forum in Beijing. The Global AI Governance Initiative (GGI) describes the Governance of AI as a “common task faced by all countries” and calls on countries to “enhance information exchange and technological cooperation”. The pursuit of multilateral governance via, for example, an international institution designed specifically to oversee AI development, as envisioned in the GGI, would be one way of transforming the ordinary PD matrix (Figure 1) into a game where cooperating becomes a Nash Equilibrium (Figure 2) (cf. Ostrom Reference Ostrom1990).

Of course, there are concerns one could raise with regard to the “dexterous ambiguity” inherent within China’s approach to AI and global governance (Olander and Wong Reference Olander and Wong2025). Multilateral institutions require buy-in, connections, and institutional know-how and capacity to impose effective sanctions, for them to be able to meaningfully impose the penalties associated with defection. One could also criticise this proposal not on grounds of theoretical intactness, but for its empirical infeasibility. Likewise, one can observe that mere pledging of openness to global governance does not translate into a binding policy or commitment towards such mechanisms. These are valid concerns. But the general sentiment in China has been different from that in the US – though there are signs that the Trump administration is beginning to see the value of cooperation.

Hawks and Doves

In the competition game set out earlier, there is a unique Nash Equilibrium characterised by both parties deciding to race. This is because if I race, you can race too – i.e. we can both attempt to develop our own AI ecosystems and I cannot stop you from developing yours. And while we cannot both win – in the sense of being first – we can certainly both reach the moon. Now consider a different kind of race – one where the underlying shared resource is strictly rivalrous. This is best illustrated via the following example:

A and B are racing towards a tunnel and each of them must decide whether to drive straight or swerve; if A drives straight and B swerves, A wins (and vice versa); if both swerve, they both lose; if they both drive straight, they both die. Naturally, a losing tie is preferred. This game thus has two, non-overlapping Nash Equilibria as described in Figure 4.

Figure 4

The Hawk-Dove game with two unique pure strategy Nash Equilibria highlighted in bold font.

This game, known as the Hawk-Dove game (Maynard Smith and Parker Reference Maynard Smith and Parker1973) is a more extreme version of the earlier competition game, because now the good in question (access to the tunnel) is strictly rivalrous (we cannot both fit), yet it is non-excludable. It is also a puzzling game: if we both decide to swerve, then we are out of equilibrium – at that point, either of us would be better off by deciding to drive straight. Yet once we both decide to drive straight, we are again out of equilibrium and one of us should swerve.

Notice that the equilibria in this game occur where I do one thing and you do the other. We benefit by not working together. Perhaps an advocate against more AI cooperation and parity across countries, then, could draw an analogy between AI and the street race: in the specific context of AI development for military purposes, if a leading country unilaterally places restrictions (i.e. ‘swerves’) concerning the level of performance exhibited by AI models, for the fear that they could produce deleterious consequences if not aligned properly, some may worry that its geopolitical rival can march ahead by ‘driving straight’ – namely, in applying a brakes-off approach to AI development.

Should both countries cease innovating with their AI, however, they could deprive their populations of its potential opportunities. And should both countries fail to apply any restrictions to their AI development, we could witness substantial interactive risks, as discussed in Section 2.

Despite these ominous conclusions, there are reasons to think that it is possible to engineer AI governance dynamics in a way that renders it disanalogous to a Hawk-Dove game. We shall see why shortly.

Free-Riding as a Problem?

In the case of nuclear disarmament, for example, the optimality of free-riding is obvious: if I can get you to eliminate your nuclear arsenal, while I keep mine, which I have already invested in, that creates a low-cost security buffer for me. I would rather keep the optionality of holding on to my nuclear arsenal ‘just in case’ – while you commit to reducing yours.

Yet it is not clear what free-riding amounts to in the context of inter-state AI dynamics. If the choice between two states is to share or not share certain frontier models, technologies, or the components required for their development, then refusing to share while the other party shares is not necessarily optimal: you then lose the ability to disseminate your own platforms, standards, raw materials, and modes of development.

The underlying good – which may be construed as the frontier AI models and systems, for instance – is largely non-rivalrous (we can all use the same model, so long as it remains openly available (e.g. is made open-source, and the use of which is charged at a minimum or implemented locally). While the marginal user places a non-zero strain on the underlying data centres and energy resources used to power that model when they are not run locally, such additional cost could be shared yet it is excludable (I can choose to block your access to it). Indeed, we have seen many examples of exclusion; at the market level, OpenAI’s decision to block access to its models from IP addresses originating in China (Altman Reference Altman2024); and at the state level, the US Department of Commerce’s decision to block exports of leading GPUs (BIS 2024) or Congress’s decision to in effect force ByteDance’s divestiture of the famous social media application, TikTok (138 Stat. 955, ByteDance 2024).

As such, this is the reverse of the Hawk-Dove game: from any given state’s perspective, we have largely non-rivalrous but excludable goods instead of rivalrous but non-excludable goods. While US-China examples are plentiful, the Dutch government faced a similar choice as to ASML: they can exclude countries from purchasing EUV lithography equipment, while any given country’s use of such equipment is not strictly inconsistent with another country’s use of it (of course if one machine is in Germany, that same machine cannot be Italy – but the point is that Germany’s ability to purchase ASML’s EUV lithography system is not incompatible with Italy’s ability to purchase one as well, unless it were unduly difficult for ASML to manufacture both (which is not the case). We have also seen examples of non-exclusion and collaboration: DeepSeek R1 is a largely open-access model which anyone can use, while recently Alibaba and Nvidia agreed to integrate Nvidia’s robotics and embodied AI tools on Alibaba Cloud (Chow Reference Chow2025). When we have goods that are excludable but largely non-rivalrous, it is advantageous for both parties to cooperate since they can share in the resource, yet it is likewise tempting to not cooperate since exclusion is a viable strategy.

In particular, if we have two states facing a choice between sharing an AI resource, such as a frontier model, or excluding the other party from that resource, then the best outcome occurs where they both choose to share, thereby capitalising on the benefits from cooperation and mutual gains from trade and joint governance. The worst outcome is where the target state retreats into isolation while its competitor reaps all the gains from cooperation – the reverse of this is second best, for the target state, trailing to the cooperative outcome. The second worst is where both parties descend into isolation.

That leads to the following game-theoretic setup (Figure 5).

Figure 5

Cooperation. A typical Stag-Hunt or Assurance Game with two pure strategy Nash Equilibria highlighted in bold font.

This is in some sense the anti-Hawk-Dove game, and it is also similar to a PD game where defections are punished, as in Figure 2: there are two pure strategy Nash Equilibria, but they occur where we coordinate and take the same action instead of where we perform the opposite acts. Readers familiar with game theory will recognise this game as the ‘Stag Hunt’, also known as a trust game, assurance game, or more generally cooperation game. Coincidentally, the name comes from Jean-Jacques Rousseau, who discusses such a situation in his A Discourse on Inequality (Reference Rousseau1755). Rousseau uses the parable to explore pre-social humans organising a hunt; whereby they face a choice of hunting a bigger prey together (the stag) or chasing a smaller prey alone (the hare). The lesson for Rousseau is that cooperation is possible, but it is also fragile because defection is tempting. Embedded in this is a criticism of Thomas Hobbes’ social contract tradition; for Hobbes, the state of nature is characterised by a war of all against all (Hobbes Reference Hobbes and Pogson Smith1651).

This dual situation – the possibility of cooperation together with its fragility – is perfectly captured by this game as it contains two pure-strategy Nash Equilibria, one of which is a Pareto improvement over the other. If both parties choose not to share, they will indeed find themselves in Nash Equilibrium. This therefore poses a substantial problem; one that is closest to capturing the present world as of 2025 – where US tariffs and export controls lead to similar responses, and a tit-for-tat exchange of non-cooperative moves which leads the parties to get ‘stuck’ in a Pareto inferior outcome, yet because that outcome is still Nash, there is no incentive to change course. The vicious cycle then becomes self-reinforcing during extended repeated play. Yet if the parties can cooperate – they can likewise get ‘stuck’ in the Pareto superior Nash Equilibrium (top left), and in that case there is likewise no incentive to deviate, which can lead to a virtuous cycle that is likewise self-reinforcing, but encourages cooperative behaviour over repeated play. Indeed, the Nash Equilibrium on the top left is preferable to that in the bottom right.

The critical challenge concerns how we should get to the point where all parties in the global AI space move towards and remain in the top left. This will require, among other policy initiatives, compelling and credible international institutions to enforce and channel cooperation and commitments. In the concluding section, we examine ways in which this can be done.

It is worth highlighting that these are not the only Nash Equilibria in the cooperative game. We can identify a mixed strategy equilibrium given the framing in Figure 6. Let p represent the probability that State A cooperates to some degree (i.e. shares at least some AI technologies with others) and 1-p represent the probability that it does not at all (i.e. shares no AI technology with others). Similarly, let q represent the probability that State B cooperates and 1-q represent the probability it does not. It is rational for each player to adopt a probability distribution over pure strategies when the expected payoffs of their pure strategies are equal.

Figure 6

Cooperation. This is functionally the same game as in Figure 5, but with probabilities of each action expressed in terms of p and q.

Now we can ask: what is the mixture of p/1−p (and q/1−q) that makes State B (and State A) indifferent in the sense that its expected payoff from sharing is equal to its expected payoff from not sharing? This will occur when p = 0.5 and q = 0.5. Therefore, if both countries act perfectly arbitrarily, that would be a Nash Equilibrium.

This too is something that we partially observe in real life: if one state’s actions seem wholly erratic and unpredictable – i.e. approximately random – then other states will have an incentive to behave in a way that is equally unpredictable – i.e. to randomise or at least behave as if they are randomising. Yet even though behaving arbitrarily is a Nash Equilibrium strategy, it is Pareto inferior to cooperating. When p = 0.5, then the expected payoff is 2.5 (and when q = 0.5, then it is likewise 2.5). Hence, randomising is a Pareto improvement over both parties refusing to share, but it is still Pareto-inferior to both parties sharing. This nicely captures Rousseau’s parable of the fragility of trust and the difficulty of sustained cooperation.

We even see evidence of something like this not at the state level, but rather at the industry level, and industry choices are increasingly important as large tech companies increasingly influence states’ foreign policies. For example, OpenAI’s Sam Altman argues that sharing with China would undermine democratic values (Altman Reference Altman2024); hence, from his perspective, the absolute worst thing to do is to share. This is even more extreme than Figure 6. If we revise the preceding game with the assumption that sharing is worst, regardless of what the other party does, and excluding is best, regardless of what the other party does, then (from Player A’s perspective) the situation looks like this:

Figure 7

Stubbornness. This is the same game as in Figure 6, except now Player A most prefers not to share and least prefers to share.

In this case, the unique pure strategy Nash Equilibrium occurs when Player A does not share while Player B shares (Figure 7 bottom left). Unsurprisingly, this game has no mixed strategy Nash Equilibria. This captures OpenAI’s argument in Sam Altman’s Wall Street Journal op-ed; the belief that upholding democratic values enjoins one to never share with an adversary means that one’s preferred strategy is unresponsive to what the other player does.

Yet what if mutual non-cooperation, where both states retreat into isolation, is the worst outcome – as compared to unilateral cooperation?

Let us turn to the industry level again and take DeepSeek and OpenAI as an illustrative example. One could argue that it is ultimately better for DeepSeek, and in turn overall for China, to remain open than to retaliate against OpenAI, because their model, and in turn China’s overall approach to AI governance, gains increased exposure and both DeepSeek and China’s approach to AI governance gain the opportunity to become entrenched as the dominant model and framework through expanded access. Indeed, the US is one of the top five user countries of DeepSeek (Kumar Reference Kumar2025). Additionally, given the extensive discussion in Sections 1 and 2, we can see that the costs of non-cooperation are significant and will likely spiral over time through lock-in and path dependence (Hebbar and Hubinger Reference Hebbar and Hubinger2022; M. Wong Reference Wong2025).

We can generalise this suggestion – namely, that unilateral cooperation is better than mutual non-cooperation – to state-level dynamics more generally. The following is in effect a version of the classical argument for free trade. Consider a situation where China knows that the US will refuse to grant export licences for leading SMEs, and in return, China must decide whether to allow exports of critical minerals, in particular its large stock of rare earth elements (REEs). If China allows exports of REEs while the US refuses exports of SMEs, then China can continue to develop its own manufacturing expertise while building goodwill and critical relationships with trading partners. This could plausibly be better than punishing the US’ non-cooperation. Similarly, regardless of whether China controls the export of REEs, there is a tremendous economic benefit to the US in allowing the exports of SMEs – indeed, its companies can thereby become global leaders in cutting-edge technology. In doing so, they can capture more of the global market, develop the relevant know-how overseas, and position the US to lead on effective governance and regulation.

One may be tempted to respond that there is a national security threat to the US if it unilaterally exports SMEs. Indeed, this has been the argument throughout much of the legislative activity leading to export control restrictions on SMEs in the US, including in the Interim Final Rules of 2022 and 2023, as described in Section 1 (IFR 2022; 88 F.R. 73458 (BIS 2023)).

Dario Amodei articulates this argument in an essay published in January 2025, claiming that export controls are essential in slowing down China’s development and ensure that the US wins the AI race (Amodei Reference Amodei2025). Yet the speciousness of that argument should be clear now. The entirety of this section has argued that inter-state AI Dynamics should not be viewed as a race; if it does not make sense to speak about racing to win, then it does not make sense to privilege slowing down an opponent at the expense of economic, multilateral, and other cooperative gains.

Had the US continued selling leading Nvidia chips to China, and indeed had they not pressured the Netherlands into refusing to sell EUV lithography equipment to China, would we be in a worse world than we are now? We do not think so. Indeed, we think it would have been a better world. The US could have reaped tremendous economic benefits, its high technology industry could have grown at a higher rate, and the market capitalisations of the big five US tech companies could be much higher than they are now, many American startups could have been able to benefit from a large overseas market, and US financial markets, as well as the many institutional funds tied to them, could have done better (while arguing over counterfactual predictions can be treacherous – such a possible world is not implausible, and is indeed likely). Hence, far from the doom and gloom scenario that Amodei tries to paint, we see a lost opportunity not only from mutual gains from cooperation, but even from unilateral non-cooperation. Co-operating can be persuasively seen as the dominant strategy regardless of what one’s opponent does.

By contrast, when it comes to the space race or the race to develop a nuclear weapon, there were distinctly apparent upsides associated with one’s victory. The US, for example, was able to secure a decisive victory against Japan in World War II in part due to its possession and willingness to use the atomic bomb (Alperovitz Reference Alperovitz1995; Maddox Reference Maddox2004). Yet this kind of unique gain from unilateral victory simply does not exist in inter-state AI dynamics, given the excludable yet largely non-rivalrous nature of the significant upsides.

One may criticise our discussion for reducing the policy debate into cooperation and non-cooperation. Of course, the real world is not binary. Instead of referring to two actions, we could consider a continuum of possibilities that the US can choose from between eliminating all export controls and refusing to export any SMEs. Yet the binary framing helps to clarify our intuitions without much loss of generality.

Let us consider further what the preceding discussion implies from a game-theoretic perspective. Under the assumption that unilateral cooperation is preferred to mutual non-cooperation, we have the following framing, where the two choices are to, for example, share/refuse to share technology/resources (Figure 8):

Figure 8

Cooperation. This is the same game as in Figure 4, except now unilateral cooperation is preferred to mutual non-cooperation. The Nash Equilibrium is highlighted in bold.

This game has one pure strategy Nash Equilibrium, which is to share, and it does not have any mixed strategy Nash Equilibria.

If we grant that mutual isolationism is least preferred, then indeed cooperation is the unique Nash Equilibrium, in either pure or mixed strategies. The core point is not technical, but rather a normative point about preference ordering, and we have tried to argue for it in the preceding paragraphs. As we said at the outset, the concept of Nash Equilibrium is itself normative: this is not a prediction about what states will do. It is a claim about what states should do. If we see states behaving otherwise, then either they are irrational or they have calculated the payoffs differently. Hence, as scholars, we can either (a) encourage them to act more rationally or (b) encourage them to revise their payoff matrix. For example, we may publicly advocate that in developing its export control rules in the Interim Final Rules, the US Department of Commerce has exaggerated the benefits of export controls while overlooking its full costs.

Establishing Governance and Dispute Resolution Mechanisms (Principles #1 and #4)

If states believe that the commitments of other states on AI can be trusted, they are more likely to cooperate. This credibility is eroded with unpredictable and erratic actions, which have characterised the foreign policies of numerous international and regional powers, for example. One way to encourage credible commitment is by establishing effective and robust dispute resolution mechanisms. Such a recommendation is largely supported by our first two guiding principles; in particular, dispute resolution mechanisms are centrally important in collaborative games with multiple equilibria, perhaps even more so than they are in the work of Ostrom (Reference Ostrom1990), who places central importance on dispute resolution mechanisms for the management of common resources in PD-style games. When we have multiple equilibria, dispute resolution mechanisms can foster cooperation and cultivate the trust required for players to believe that their counterparts will engage in cooperative behaviours; such a belief is in turn required for these individuals to likewise engage in cooperative behaviour. This is in the spirit of the Hobbesian approach to encouraging cooperation in the Stag Hunt: cultivating the institutions that are required to change players’ beliefs about the likelihood of cooperation.

When effective, accessible and low-cost conflict resolution mechanisms are in place, they can shape parties’ beliefs about the prospects of cooperation, by providing assurance that non-cooperative behaviour can be quickly addressed. This should not come as a surprise – for instance, among the very central functions of the WTO has been its staged process for resolving trade disputes. Complaints at the WTO are triaged accordingly: members begin with a complaint, which is then processed by the Dispute Settlement Board (DSB) and begins with consultations; where consultations fail to resolve the dispute, a panel is formed by the DSB. The panel produces a report which is then adopted by the DSB. Crucially, parties unsatisfied with the panel’s report may appeal to the AB (World Trade Organization 2004). Unfortunately, in the case of the WTO, the AB has been non-functional since December 2019, due to the US’ blockage of new appointments, which has resulted in a lack of quorum (Gao Reference Gao2021).

Disputes around inter-state AI dynamics are inevitable, across a range of products and industries – including the mining of rare earth minerals, the manufacturing of semiconductors, the requisite equipment, the assembly of chips and GPUs, and the export/import restrictions of any of these items. The current regime does not reinforce cooperative behaviour. Instead, what we see is either unilateral trade barriers or tit-for-tat punitive measures. What is separately needed is a dispute resolution mechanism that stimulates confidence among states in the possibility of cooperation. There is no magic pill to solve this: the way to build such trust is gradual and it requires states to be willing to enter into and, more importantly, honour their agreements over time.

As we saw earlier, both China and the EU have advocated for the creation of an international AI governance institution, either independently or within the purview of the United Nations. For such an institution to be successful, it will be critical to include a dispute resolution mechanism, with a credible enforcement scheme (Hobbesian approach), or strongly institutionalised norms and rituals of cooperation and compromise (Humean approach). Indeed, more substantively, China has recently established the International Organization for Mediation (IOMed) (Government of the Hong Kong Special Administrative Region 2025). One way forward would be to build on this development and establish a mediation – or more broadly dispute adjudication – centre for geopolitical conflicts regarding inter-state AI dynamics, perhaps in a strategically neutral geographic location. Another way forward could be to repair the WTO AB and build within its purview a body dedicated to inter-state AI dynamics, though this is less comprehensive, as the WTO’s mandate is limited to trade disputes.

Reforming Supply Chain Management (Principles #2 and #3)

The core of deep conflicts regarding inter-state AI dynamics involves access inequality and supply constraints. On grounds of both relational egalitarian principles and the interests of actors involved, managing supply in rare earth minerals, semiconductors, chips, GPUs, and related SME is essential for reinforcing cooperative behaviour. A number of policy developments in China have reflected a broader drive for global collaboration in supply management (e.g. MFA 2023 and MFA 2025b). These policy pronouncements could be reinforced with multilateral attempts to manage the supplies of AI-related technology, perhaps again through a multilateral institution, either within the auspices of the UN or independently. More recently, the US has also begun showing signs of dethawing of some export control restrictions on Nvidia GPUs (Mickle Reference Mickle2025); such positive policy changes, too, can be reinforced with more systematic multilateral supply management agreements.

When countries fear that their critical supplies will be cut off, this strongly incentivises reciprocal non-cooperative behaviour, which leads to a cycle of isolation. We grappled with this earlier, when the US export controls, described in Section 1, restricted, for example, Huawei’s access to the leading Nvidia GPUs as well as restricting its market access in the United States. As a result, Huawei turned to domestic supply chains, and to cultivating a robust domestic market (Olcott Reference Olcott2025). While they managed to succeed in that market, a cooperative counterfactual where American consumers can freely purchase the latest Huawei products, and Huawei can get unfettered access to leading Nvidia GPUs, would have benefited the stakeholders of and consumers in both countries. Indeed, this is precisely the lesson of our argument in Section 4 – that mutual noncooperation ought to be the least preferred outcome and, by extension, that unilateral cooperation should be preferred to mutual noncooperation.

In the future, the role and agency of small and medium states – which serve as vital nodes in what Brantly Womack terms the ‘multinodal’ world (Womack Reference Womack2024), or as critical chokepoints when it comes to AI supply chains – cannot be overstated. These states not only possess the empirical room, but also the normative imperative, to maneuver across different, competing spheres of digital influence and AI standards. They should exercise their leverage against more powerful counterparts with care and precision, with the intention of pressing for more options, choices, and room for them to be non-aligned. In the face of false binaries, the rest of the world may seek an alternative.

Devising Global Mechanisms of Blameless Responsibility Uptake (Principle #3)

It is beyond our scope to discuss proposals aimed at bolstering mechanisms of legal accountability over AI failures and errors at length, though one of us has written extensively on the subject elsewhere (Babic et al. Reference Babic, Cohen, Evgeniou and Gerke2021). Granting that there exists no single overarching legal authority with the capacity of exacting and enforcing punishments and penalties for AI risks – especially across national borders – an alternative path may entail the establishment and implementation of policies, both within and in coordination with other nations, that would allow for responsibility uptake. Such responsibilities should not be rhetorically couched in terms of criminal liability, moral culpability, or political fault, for such labels, as noted by structural injustice theorists (e.g. Young Reference Young2011, Lu Reference Lu, Brown and McKeown2024, McKeown Reference McKeown2018, Nuti Reference Nuti2019), could often be invidious, exacerbate injustices, and undermine effective fact-finding.

Domestically, governments should explore legislation that binds actors across all stages of the AI training, development, and application process to duties to redress the risks associated with their decisions. A concrete example would be OpenAI, whose amassing of extreme levels of financial wherewithal, model opacity and disdain for open-source development, and repeated corporate restructuring efforts that have diluted its commitment to “ensure […] technology serves humanity” has drawn considerable consternation from experts in the AI community (A Right to Warn about Advanced Artificial Intelligence 2024; An Open Letter to OpenAI 2025). More specifically, in the geopolitical context, the US government should recognise that such concentration of resources within individual corporations – in the absence of accountability mechanisms – could well generate radical geopolitical uncertainties harmful not just to rival states, but also to the interests of its own people and itself. Legislation delineating limits to private corporate usages of AI in geopolitically sensitive contexts, without government or public authorisation, should be considered across jurisdictions in the world – democratic and non-democratic alike.