Forecasting China’s response to the threat of ASI-related disempowerment

Introduction

This forecast scenario explores a future where Chinese leadership recognizes the potential for a US-controlled artificial superintelligence (ASI) to create a decisive strategic advantage, leading to a cascade of actions by China to avoid disempowerment. This realization stems from China gaining access to OpenAI's state of the art technology, revealing capabilities and a timeline that force a reassessment of the strategic landscape.

Methodology

This exploration was chosen because it is plausible, divergent in consequences from the mainline scenario (a forthcoming work which I’ll link to here once it is available), and under-studied. Initial drafts of this timeline have been refined through rounds of expert interviews and feedback.

Motivation

Leopold Aschenbrenner has argued that a prudent strategy for the US is to push ahead and race for supremacy in AI in order to use the technological benefits to foster US national security.¹ Even putting aside the inherent dangers from unaligned or uncontrolled ASI, the flaw in this strategy is an expectation that China won’t see the consequences of AI coming (even though they are laid in his own writing and in other places). This work seeks to concretize both the way that China could anticipate these developments and to examine the recourses available to it, should it seek to avoid losing the race to ASI and the resulting disempowerement. In so doing we provide support for the alternative strategy of international cooperation for mutual security and benefit.² While this approach has many challenges, it may be the only viable option to avoid both war and rogue superintelligences.

Organization and Probabilities

Each step along the scenario has a description, a probability estimate, and a supporting justification. The estimates express the likelihood of the event-as-described or something that is functionally similar happening in a similar timeframe, conditional on events up to that point. Thus, the overall timetable should be understood as including some elasticity. Additionally, while this single scenario makes explicit one sequence of events that ultimately leads to conflict, there are other possible sequences leading to the same or similar results. Sometimes these variations are called out. All events are ultimately driven by the same fundamental dynamic: developing ASI, when its power is appreciated, amounts to an introduction of existential stakes into geopolitical competition, which creates pressure for other parties to intervene and divert the future away from such perceived disadvantage.

Earlier than September 2025

Broadly, events up to September 2025 are not surprising. A leading AI model by OpenAI is called Agent-1.5. In internal evaluations, Agent-1.5 goes beyond Persuasion and CBRN medium (attained by o1 in 2024) to also achieve preparedness Medium in Cybersecurity and Model Autonomy. As a tool/assistant, it is accelerating AI researcher productivity by 25-50%.

September 2025

September 2025: Chinese intelligence successfully exfiltrates key components of OpenAI's latest internal AI model, including model weights, architecture, scaffolding, and internal research documents. This is achieved through a sophisticated operation which combines cyber intrusion and human vulnerabilities. OpenAI’s cybersecurity efforts have not yet escalated to the point of preventing exfiltration (by a sophisticated state actor) as the model is not considered to be high risk pre-mitigation.³

85% probability of the above. Justification:

US AI labs are vulnerable to sophisticated espionage efforts as they have not dedicated the resources and effort to securing themselves against a state actor. OpenAI has been leading AI development for years and is a natural target for Chinese intelligence seeking to assess the pace and trajectory of US progress.

Variations

These sections include notable alternative events which serve much the same purpose and lead to similar eventualities.

• Exfiltration of state-of-the-art technology from other labs or US government agencies which have been given access either for oversight or to develop applications

• Developments by Chinese labs which provide compelling evidence of the threat alongside the belief that they are lagging their US counterparts.

• Economic impacts including job displacement (even if overseas) start to make the significance of AGI and later technologies unignorable.

January 2026

China concludes that a US-controlled decisive strategic advantage (DSA) is imminent. After several months replicating, analyzing, and experimenting with OpenAI's stolen model, Chinese leadership arrives at several conclusions:

• The model's general capabilities surpass expectations, demonstrating advanced proficiency in natural language processing, code generation, strategic reasoning, and long-term planning.

• Economic projections reveal a high likelihood of significant economic disruption and runaway US economic growth as the technology matures. In a world where the impacts of AI are limited to productivity improvements, the potential is still for 2-8x improvements in the productivity of knowledge work. There would, on a longer timescale, also be transformative effects on manufacturing, construction, and transportation. Combined with the belief that China will not be able to keep exfiltrating models indefinitely, it worries about being left behind and denied access to this technology by the West. The compounding effects of growth would lead to China being uncompetitive.

• The timeline to ASI appears significantly shorter than previously anticipated, with a US-controlled singularity now perceived as a near-term possibility. This covers worlds where the impacts of AI are effectively unbounded, and technological progress accelerates quickly across critical applied domains including military, cybersecurity, and human persuasion (endangering domestic stability.) If the US is able to attain such large advantages in these areas, it amounts to a DSA which would quickly make competition with the US impossible.

These conclusions, informed by both the stolen model and a broader understanding of AI trends, create a sense of urgency within the CCP, pushing them to consider drastic measures to avoid being left behind. There is sufficient fear of the consequences of falling behind that support materializes within the Politburo to authorize investigations and the formulation of options (diplomatic and otherwise).

25% probability of the above. Justification:

Justifying these conclusions is difficult. Because of the strategic and political sensitivity of this line of thinking, and the speculative nature of the forecast, research into Chinese sources cannot find explicit support for these conclusions. Nevertheless, we can find clues by examining Chinese thinking about adjacent topics and considering the priorities implied by China’s past actions.

Thought Influencers

CCP leadership thought can be influenced by researchers in academia, and to a lesser extent (and a lesser extent than the West) by private industry.

Wen Gao, director of a major Chinese AI lab and dean at Peking University, has influence in technical and political circles in China, serves on high-level advisory bodies for the government, and led the Politburo in a collective study session on AI. In 2021 he raised concerns about AGI, recursive self-improvement and an intelligence explosion, leading to catastrophe.⁴

Andrew Yao’s “word carries significant weight in academic and policy circles”, and advocates for beneficial superintelligence and “alignment with human values”, and is uncertain about timelines (including the possibility that timelines are short). He sees AI alignment as a problem downstream of first solving the problem of value misalignment between humans.⁵

Lee Kai-Fu, former Google China president: calls for 1/5 of researchers to be devoted to safety issues. He foresees that AI’s impact on society will be profound, leading to job displacement and increased income disparity in the short term, but potentially ushering in an "Age of Plenty" in the long run.⁶

What these perspectives suggest is that the transformative implications of superintelligence are appreciated by those who are actually informing the thinking of Chinese leadership.

History and Disposition

China has responded to US export controls on AI chips and semiconductor manufacturing technology and inputs with its own export controls on gallium and germanium. As China controls 98% of gallium production and 60% of germanium production, this has led to prices in Europe to nearly double and for reserves to be drawn down, potentially leading to shortages.⁷ The selection of these inputs as part of the export control package seems clearly motivated to impair US AI progress (and to respond to US efforts to do the same to China), which reflects China’s appreciation of the importance of competition in this space.

Social stability and cohesion is one of the utmost concerns of the CCP, and leadership already has experience with manipulating these factors using information warfare tactics, including the AI-driven dissemination of propaganda and disinformation. China’s internal security budget has exceeded its defense budget for more than a decade.⁸ There will be attention to the national security implications of the human persuasion capabilities which Agent-1.5 demonstrates.

This occurs against a general backdrop of competition and rivalry in these domains. Edward Geist argues, “Today both the United States' and China's military strategists fear falling behind their rivals in harnessing AI and other emerging technologies. This is the kind of dynamic that stokes costly arms races, increases the probability of international crises, and makes crises that do occur more likely to escalate to large-scale war.”⁹

In conclusion, Chinese leadership is attuned to the gravity of the potential consequences of AI development as well as has demonstrated an inclination to prioritize competition and national security (stability).

Events up to this point constitute the deviation from mainline likelihood. The claim is that these events are less-likely to occur than not, but still likely enough, and consequential enough, that it is worth investigating the results. After this point, the claim is that we’re illustrating most-likely predictions given the deviation above.

February 2026

China pursues bilateral diplomacy with the US. Its position is that US development of AI is on a trajectory where it will be used to dominate Chinese interests and also that competitive pressures make the safe development of AI impossible. “The benefits and responsibilities of frontier AI must be shared by all nations.”¹⁰

The focus of the diplomatic effort is to hammer out a bilateral AI megaproject; a collaboration between Chinese and US labs and governments.

Details of the proposed project include:

• Government oversight and control of frontier model developers in both countries

• Highly integrated research collaboration between labs

  • Full visibility into frontier model development

    • A set of research labs are selected for participation and integration: OpenAI, Deepmind, Anthropic, Tencent, Baidu, and Huawei.

  • Researcher exchange program

  • Shared access to the models

• Coordination mechanisms including mutual monitoring and on-chip verification

  • Intelligence sharing on both sides to lower the viability of a secret breakout

  • Transparency of compute usage

    • Including hardware enabled governance mechanisms to provide additional oversight of how compute is being used

• Collaborating on cybersecurity

  • To protect their shared investment from other states

• Safety (mitigating catastrophic and existential risk) is seen as a matter of making investments in and sharing the proper research, which will take resources but doesn’t have to come at the total expense of advancing capabilities.

In track 1 dialogues with the US, China’s position includes: “We are taking this matter extremely seriously. We must work together to chart a path that stabilizes the strategic situation. The US cannot expect us to sit idly by while these events develop as they presently are.”

65% probability of the above. Justification:

Unlike other military-technological advances (like anti-ballistic missile systems), China cannot pursue an asymmetric response to rapid, unbounded, general technological progress. But China is not looking for escalation if it can be avoided.

Reference, which is prompted by chip export bans but the tone and seriousness would likely apply more broadly: “With Smugglers and Front Companies, China Is Skirting American A.I. Bans” New York Times, (8/4/24):

The United States leads China in A.I. for now, but China is making rapid progress as both countries race to create A.I. that would rival human intelligence.

Liu Pengyu, a spokesman for the Chinese embassy, said that China firmly opposed the rules and that they would “only make China even more determined and capable in boosting our own strength in technology and innovation.”

China’s leader, Xi Jinping, delivered a similar warning on a call with Mr. Biden in April, saying his country would not “sit back and watch.”

China’s approach to bilateral diplomacy is based on the reality that, while there are many countries which supply inputs to frontier AI development, the US and China lead in frontier AI development and together have the diplomatic clout to pull the other key countries along, if they could only work out an arrangement between themselves first.

Variations

• Multilateral diplomacy is a highly credible alternative narrative.¹¹ The key differences include bringing in other international stakeholders and advocating for shared stakeholding in the development of AI.

February 2026 (continued)

While these negotiations are in progress, China advances plans for destructive cyberattacks on US AI infrastructure. This is seen as an important recourse to preserve should diplomacy fail. Another contingency being worked on is interdicting the semiconductor supply chain, either via a blockade or an invasion of Taiwan. But this is assessed to not be a sufficient action, as compute already deployed by US labs is sufficient to be extremely threatening. At the same time, these actions against Taiwan would be highly escalatory and entangling.

90% probability of the above. Justification:

Conditional on the diplomatic effort and given the real possibility that these sorts of plans are already in some state of contemplated, researched, or partially developed already, I am relatively confident that substantial effort would go into developing real options to impede US AI progress using non-conventional means.

April 2026

Ultimately, however, diplomacy fails to produce a result. It does not take very long for this to become apparent either.

China stresses that it considers the current trajectory a reasonable cause for war: the threat of disempowerment justifies the use of force, but China is committed to exerting every effort to pursue a peaceful path forward.

98% probability of the above. Justification:

The following may motivate the US to participate in bilateral negotiations:

1. Participation hedges against the chance that its current lead will not last until the development of ASI.

   • Export controls on compute have been less-than-perfect.¹²

   • China will indigenize its semiconductor supply chain and can make competitive AI chips at scale by somewhere around 2035. By making a deal, the US lowers the incentive for China to invest in semiconductors, delaying this eventuality. It also lowers the incentive for a variety of undesired efforts which could include:

     • Attracting talent away from TSMC, Samsung, Intel, etc.

     • Competing for investment and partnerships with elements of the west’s semiconductor supply chain directly.

     • IP theft

   • The US may not be able to develop the required electricity domestically for the upcoming generations of training in a competitive timeline, another factor which could undermine the US’s present lead.

2. Participation defuses race dynamics and gives time to address potential catastrophic and existential risks.

However, these factors make a partnership unviable:

• The history of rivalry, competition, and mistrust between the US and China, not least of which includes the mutual export controls on AI and its inputs (e.g. semiconductors, semiconductor manufacturing equipment, memory, and minerals).

• The frontier labs are totally against this for obvious reasons and throw all their weight against a deal. Precedent includes US frontier labs’ history of rejecting all forms of governance.

• The difficulty of keeping everyone honest is insurmountable given the relative ease of hiding breakout efforts, the incentives at play, and the resources available to each side.

• Hardware-enabled governance mechanisms (HEMs) are a category of diverse efforts, only some of which would be relevant to this set of negotiations. Early approaches are more likely to target domestic oversight with a plausibly cooperative lab. It is a much harder problem to apply this technology to oversight and monitoring of an adversary’s efforts, and in this scenario this problem has not been prioritized.

Variations

• A minor variation is where the US “plays along” with negotiations in bad faith, actively stringing out the diplomatic process. In this variant it may take longer for it to be clear that diplomacy is not working, but China did not have high expectations going in either way.

May 2026

China uses cyberattacks to impede US AI progress. This is primarily a Stuxnet-style computer worm which causes GPUs to self-destruct by overheating. The cyberattacks are only partially successful, less than was hoped for, amounting to an (unevenly distributed) 25% reduction in available compute¹³ and a couple weeks of greatly reduced compute availability while labs sort out through the fallout. This isn’t enough to change the fundamentals of who is leading in the race.

China is overall disappointed with the effects, but hopes the attack can still demonstrate to US leadership the seriousness of the issue and the need to seek a diplomatic deescalation.

50% probability of the above. Justification:

“Another reality we must accept is that the adversary gets a vote. … There is little reason to hope that other powers will sign treaties condemning themselves to perpetual second-class status in nuclear weapons or any other perceived source of political or military advantage” - Edward Geist, Deterrence Under Uncertainty

Following the unsuccessful negotiations, China seeks alternative recourses which are available to it. A cyberattack is feasible, demonstrates China’s resolve, and minimizes collateral damage and loss of public opinion (relative to other options).

GPUs have many security vulnerabilities and can be manipulated in ways that cause hardware damage.¹⁴ GPUs aren’t particularly secure compared with other types of processors. The Stuxnet worm itself is evidence that states can carry out attacks of a similar complexity and that China has had a couple of decades to consider developing capabilities of this sort in response.

However, the attack described faces a very large and uneven attack surface (many locations operated by different organizations and using different hardware), and one where it will be very difficult to conceal the immediate effects (it is difficult or impossible to fake the outputs of a destroyed GPU, and thus this would require compromising the monitoring system as well).

May 2026 (continued)

Multiple reactions emerge from the US as a result of these attacks.

The US responds with counter-force cyberattacks. The aim is to degrade Chinese cyber capabilities, to reduce their ability to further cause damage during the current attack, as well as to degrade Chinese intelligence gathering and assessment of the effectiveness of the attack.

In a matter of days, this prompts more cooperation between US labs and the government to secure AI infrastructure from cyber and conventional attacks. It will take a year or longer to reach SL5 and new security protocols impose a 30% reduction in researcher efficiency.

Later this month, the Bureau of Industry and Security, in cooperation with other parts of the US government, deploys additional export controls.

Note that both sides keep the details of the attack a secret for the usual reasons.

85% probability of the above. Justification:

While there are component events in the above this general shape of the response by various groups in the US is relatively likely. A variation which is also plausible could include:

• The US responds with retaliatory cyber attacks on Chinese AI infrastructure, justified in order to deter further attacks. This is not particularly more escalatory than the counter-force attacks in the scenario.

June 2026

China pursues an international pause on frontier model development. It believes it has a strong case to make:

• The global human interest is in taking time to study the effects of powerful AI, before deploying these technologies. This is directed at economic disruptions, catastrophic misuse risk, and existential risk.

• There are many economic benefits which could be attained without advancing the frontier of model capabilities and efforts should focus here instead of pushing the frontier of general capabilities.

30% probability of the above. Justification:

This represents a diplomatic option of last resort, but could still proceed alongside the preparation of other recourse as we’ve already seen.

October 2026

When the US refuses, China uses a conventional missile barrage to take out as much AI compute infrastructure in the US as possible. The targets include data centers and semiconductor manufacturing. China can only reassure the US that the incoming missiles are conventional and only targeted as described. Warning is given to minimize human casualties, as the goal is to destroy hardware and infrastructure.

A number of missiles are intercepted but many still go through.

20% probability of the above. Justification:

The chance of this is quite uncertain. Many other options have been exhausted. China has the capability to strike at the US mainland using the DF-31 intercontinental missile which can carry the DF-ZF hypersonic glide vehicle. This produces a range of 12-13km.¹⁵ Some of these missiles are equipped with non-nuclear EMP weapons, which have the effect of destroying electronics in a small area while minimizing the loss of life. China has been developing such weapons for 20 years.¹⁶

US missile interception remains mostly ineffective against hypersonics. There are sufficient attacking vehicles to saturate US defenses, which are not designed / scaled to defeat a large-scale attack of this nature. Recent advances in high-energy laser defenses have some small impact as well.

Events past this point escalate to a de facto war in chaotic and difficult to predict ways, and are beyond the scope of this exercise.

Conclusion

This scenario illustrates events, sparked by the imminent threat of a rival developing ASI, that lead to the destabilization of international relations and conflict. In a world where each side is attuned to the implications of this technology, diplomacy and cooperation will be needed to avoid conflict. It is my hope that this scenario can improve awareness of the risk of conflict in the run up to ASI and thereby motivate the required investments to avoid it.

Communication and dialogue is essential between all stakeholders, including researchers, industry, and government. In general, opportunities to develop trust between governments should be seized upon as a general practice of making and honoring commitments. Research and development of hardware-enabled governance mechanisms has the potential to enable international verification and trust. In conjunction with the above, investment in and the creation of international institutions may be essential to many activities including assessing frontier developer progress and ensuring adherence to standards.

1. Aschenbrenner, L. (2024). Situational Awareness (p. 139).

2. Also see Against Aschenbrenner: How 'Situational Awareness' constructs a narrative that undermines safety and threatens humanity for an analysis via the framework of national securitization vs. “humanity macrosecuritization”.

3. “If we reach (or are forecasted to reach) at least “high” pre-mitigation risk in any of the considered categories we will ensure that our security is hardened in a way that is designed to prevent our mitigations and controls from being circumvented via exfiltration” https://cdn.openai.com/openai-preparedness-framework-beta.pdf

4. China’s Views on AI Safety Are Changing—Quickly - Carnegie Endowment for International Peace

5. Concordia AI. “Andrew Yao — Chinese Perspectives on AI Safety.” Chineseperspectives.ai, 29 Mar. 2024, chineseperspectives.ai/Andrew-Yao.

6. https://mp.weixin.qq.com/s/khdK1JAFDSTLYQ6eAaNlZQ

7. https://www.trendforce.com/news/2024/08/28/news-chinas-gallium-and-germanium-export-restrictions-risk-chip-production-shortages

8. RAND. “Timothy R. Heath – U.S.-China Rivalry in a Neomedieval World” https://www.rand.org/pubs/research_reports/RRA1887-1.html

9. Geist, Building a Foundation for Strategic Stability with China on AI | RAND

10. Wu Zhaohui “highlighted the importance of ensuring that AI remains under human control and emphasized strengthening the representation of developing countries” at the UK AI Safety Summit.

11. And was much explored during the course of drafting this scenario and could easily have taken the place of the bilateral approach above.

12. “This is an enormously difficult job, and I’m under no illusions that we are doing it perfectly,” Gina Raimondo, the commerce secretary, said in an interview about the A.I. restrictions. “With Smugglers and Front Companies, China Is Skirting American A.I. Bans.” New York Times, 8/4/24

13. 25% reduction in compute is a modal outcome, it could be much higher or lower

14. https://arxiv.org/pdf/1804.00114

15. https://en.wikipedia.org/wiki/DF-ZF

16. BLACKOUT WARFARE: Non-Nuclear Electromagnetic Pulse (NNEMP) Attack On The U.S. Electric Power Grid https://emptaskforce.us/wp-content/uploads/2021/06/REPORTnnempTHREAT21A.pdf