Why Might an AI Race Make Doom More Likely?

Introduction

An AI race could make AI doom more likely by rewarding speed precisely when caution matters most. A laboratory that pauses for stronger evaluations, better security or more reliable control methods may fear losing customers, investment, talent or strategic influence to a faster rival. Governments may likewise worry that restraint would leave them dependent on another country’s systems. The result can be a collective-action problem: every major participant may prefer a safer world, yet each sees reasons to move first.

This does not prove that competition will cause catastrophe. Rivalry can fund safety research, expose weak models and prevent one organisation from gaining unchecked power. Nor is there clear evidence that today’s commercial race is already on an inevitable path to human extinction. The concern is conditional: if future systems become capable of dangerous autonomy, strategic deception or rapid AI research, competitive deployment pressure could shorten the time available to detect and contain those capabilities. Recent economic models, laboratory safety frameworks and international commitments all treat this speed-versus-safety trade-off as a serious governance problem, although its magnitude remains deeply uncertain.[uchicago.edu]bfi.uchicago.eduBecker Friedman Institute WORKING PAPER · NO2026 71 The AGI Race and Existential RiskWORKING PAPER · NO. 2026 71 The AGI Race and Existential Risk WORKING PAPER · NO. 2026-71 The AG…

How commercial pressure can weaken safeguards

Frontier AI development requires scarce resources: advanced chips, electricity, specialised researchers, secure data centres and large amounts of capital. The same organisation must decide how much of those resources to devote to capability gains and how much to spend on evaluations, cybersecurity, interpretability, alignment research and cautious deployment. Some safety work supports performance, but other precautions impose real delays or restrict commercially valuable uses.

The competitive problem appears when being first has unusually large rewards. A leading model can attract users, establish a developer ecosystem, improve through real-world feedback and strengthen the company’s ability to raise money or secure computing infrastructure. Even the expectation of such advantages can make a delay look costly. A company may therefore launch with incomplete testing, offer more autonomous features, relax usage restrictions or increase model access because a rival is expected to do so.

Economic models illustrate the logic, but should not be mistaken for forecasts. A 2026 University of Chicago working paper models firms dividing resources between speed and safety. Under its assumptions, fragmenting a fixed pool of industry resources among more competitors increases the strategic value of speed and raises the risk of inadequate safety effort. Another 2026 model finds that first-mover advantages can produce premature deployment even when firms would collectively benefit from waiting. These are theoretical results rather than measurements of real laboratories, but they clarify why ordinary market competition may not automatically deliver the socially safest outcome.[Becker Friedman Institute]bfi.uchicago.eduBecker Friedman Institute WORKING PAPER · NO2026 71 The AGI Race and Existential RiskWORKING PAPER · NO. 2026 71 The AGI Race and Existential Risk WORKING PAPER · NO. 2026-71 The AG…

The danger is not simply that systems arrive sooner. Racing can alter which kinds of systems are built and deployed. Products that operate tools, write and execute code, conduct long tasks or make decisions with limited supervision may be more economically attractive than tightly constrained assistants. Those same features can raise loss-of-control risks by giving a model more opportunities to act, conceal mistakes, exploit vulnerabilities or continue after an operator has misunderstood its behaviour. Competition may therefore favour greater autonomy before oversight techniques are equally mature.

Frontier laboratories openly acknowledge parts of this problem. OpenAI’s Charter says it is concerned about late-stage artificial general intelligence development becoming a race without enough time for safety precautions, and states that it would assist a safety-conscious project nearing success rather than continue competing with it. Anthropic and Google DeepMind have developed policies linking stronger safeguards to specified capability thresholds. These commitments are meaningful evidence that developers themselves take racing incentives seriously, but they are also largely self-designed systems whose effectiveness depends on internal judgement, disclosure and implementation.[openai.com]OpenAIThis document reflects the strategy we’ve refined over the past two years, includi…

Why states may treat caution as strategic weakness

Commercial rivalry is only part of the picture. Governments increasingly view advanced AI as a source of economic growth, intelligence advantage, cyber capability, military power and influence over future technical standards. Under that framing, slowing domestic development can be portrayed not as prudent risk management but as surrendering strategic leverage.

This creates a security dilemma. One state may accelerate because it fears another state will gain a decisive advantage. The rival then interprets that acceleration as threatening and speeds up in response. Neither side needs to desire an unsafe race. Mutual uncertainty about capabilities and intentions can be enough.

Advanced AI makes this especially difficult because progress is hard to observe. Governments may know how many chips a rival imports or how many data centres it is building, but not exactly what its models can do, how securely they are controlled or whether a reported breakthrough is genuine. RAND’s work on competition for artificial general intelligence stresses that the unstable period may be the one before any clearly defined AGI threshold, when states are uncertain about both the technology and each other’s intentions. That uncertainty can encourage secrecy, espionage, export restrictions, emergency investment and pressure to deploy capabilities before an opponent does.[rand.org]rand.orgThe Artificial General Intelligence Race and International SecurityonlyUnauthorized posting of this publication online is prohibited; linking directly to its webpage on rand.org is encouraged. Permission is r…

Doom-focused concern becomes strongest when policymakers believe that the first developer of very advanced AI might obtain a lasting or even decisive advantage. If leaders expect a “winner-takes-all” outcome, they may accept risks that would otherwise seem intolerable. They might tolerate weaker evaluation standards, grant systems wider access to networks and infrastructure, or resist international inspections that could reveal sensitive information.

Yet the assumption that there will be one clear winner is disputed. Advanced models may diffuse, be copied, be stolen or depend on infrastructure that remains vulnerable to competitors. A lead may be temporary rather than decisive. Treating AI as a single finishing-line race can therefore be misleading and dangerous in its own right: it can make actors behave as though every month of restraint is strategically fatal when the actual benefits of being first may be uncertain. Research arguing against “racing to AGI” emphasises that the promised strategic dominance is far from guaranteed, while the associated risks of miscalculation and weakened safety may be substantial.[arXiv]arxiv.orgarXiv Against racing to AGI: Cooperation, deterrence, and catastrophic risksarXiv Against racing to AGI: Cooperation, deterrence, and catastrophic risks

Where first-mover incentives affect safety decisions

Competitive pressure matters most at concrete decision points, not as an abstract atmosphere. Several choices could determine whether a race raises existential risk.

Training beyond a warning threshold. A developer may discover that a model is approaching capabilities associated with autonomous replication, high-level cyber operations, biological assistance or automated AI research. Stopping to investigate could give rivals time to catch up. Continuing may preserve the lead but reduce the margin for discovering unexpected behaviour before the next training run.

Releasing before evaluations are complete. Thorough testing can require outside experts, secure access, repeated adversarial trials and time to fix weaknesses. Commercial deadlines can turn evaluation into a launch requirement that must be passed rather than an open-ended attempt to find reasons not to deploy.

Using narrow mitigations instead of reducing capability. A company may add filters, monitoring or user restrictions while retaining a model whose underlying dangerous capabilities remain accessible through fine-tuning, tool use, jailbreaks or stolen weights. This can be reasonable when mitigations are strong, but it can also become a way to avoid the competitive cost of withholding the system.

Expanding autonomy to match competitors. If one provider allows its model to run longer, control computers or act with less approval, others may feel compelled to provide similar functionality. This could create a safety race in the wrong direction: the market rewards systems that do more independently, while reliable supervision remains technically unresolved.

Publishing model weights. Open-weight releases can support research, competition and decentralised innovation. They are also difficult to reverse. Once powerful weights are widely available, the original developer cannot reliably withdraw them, enforce safeguards or prevent modification. Competitive pressure between open and closed developers can therefore push release decisions towards greater access even where catastrophic misuse evaluations remain uncertain.

The 2024 Seoul frontier AI commitments attempted to address these decision points by asking signatories to establish severe-risk thresholds, evaluate whether models are approaching them and refrain from development or deployment in extreme cases where adequate mitigations cannot be applied. Their importance lies in the idea of a pre-declared stopping rule. Their limitation is that the commitments are voluntary, allow differing methods and leave much of the judgement to the organisations being constrained.[GOV.UK]GOV.UKfrontier ai safety commitments ai seoul summit 2024Frontier AI Safety Commitments, AI Seoul Summit 2024 - GOV.UKFrontier AI Safety Commitments, AI Seoul Summit 2024 - GOV.UK Home Business…

Why shared concern does not produce shared restraint

AI laboratories and governments may all recognise catastrophic risk while still failing to coordinate. This is not necessarily hypocrisy. Restraint has concentrated costs and widely distributed benefits.

A laboratory that delays loses revenue and perhaps staff; every competitor benefits from the reduced risk. A country that accepts inspections or compute limits may reveal information or constrain domestic firms; the safety gain extends to states that made no comparable concession. Each participant therefore has an incentive to let others bear more of the burden.

Verification is another obstacle. Agreements to “develop safely” are difficult to enforce because safety is not a single observable quantity. An outside monitor might verify the number of chips used in a training run, but assessing whether a model is adequately aligned, whether internal tests were demanding enough or whether deployment monitoring will work requires access to sensitive technical evidence. Some evidence could also help competitors or attackers.

There is a further tension between cooperation and security. Sharing evaluation methods, incident reports and control research can improve collective safety. Sharing too much about model vulnerabilities, training methods or dangerous capabilities can accelerate rivals or facilitate misuse. Research on cooperation between geopolitical competitors suggests that lower-risk areas may include verification techniques, common evaluation protocols and shared scientific standards, rather than unrestricted exchange of frontier capabilities.[arXiv]arxiv.orgarXiv In Which Areas of Technical AI Safety Could Geopolitical Rivals Cooperate?arXiv In Which Areas of Technical AI Safety Could Geopolitical Rivals Cooperate?

Coordination between companies also encounters ordinary competition law. Joint testing, common safety standards and incident sharing may serve the public interest, but agreements that influence release timing or divide markets can attract antitrust concerns. Legal analysis has therefore proposed explicit safe harbours for narrowly defined safety cooperation, with safeguards against using “safety” as cover for price-fixing, exclusion or cartel behaviour.[arXiv]arxiv.orgarXiv Enabling Frontier Lab Collaboration to Mitigate AI Safety RisksarXiv Enabling Frontier Lab Collaboration to Mitigate AI Safety Risks

Would less competition necessarily be safer?

The case against racing is not the same as a case for monopoly. Concentrating frontier AI in one company or state could reduce some speed incentives, but create other dangers.

A dominant developer could face less external pressure to demonstrate that its systems are safe. It might conceal incidents, shape standards around its own interests or deploy powerful systems with little meaningful challenge. A state monopoly could combine advanced AI with surveillance, military authority and political repression. If the dominant actor’s alignment methods were weak, concentrating capabilities might also create a single catastrophic point of failure.

Competition can produce safety benefits. Rival laboratories test one another’s claims, develop alternative control methods and reveal that supposedly unique capabilities can be reproduced more cheaply. A challenger may force an incumbent to improve security after exposing a weakness. Multiple independent teams also reduce reliance on one technical approach.

The important distinction is therefore between competition in capability and competition in demonstrated safety. Governance should aim to prevent firms from winning mainly by accepting more catastrophic risk, while preserving incentives to build better evaluations, stronger security and more controllable systems. Theoretical research supports this nuanced view: competition does not have one universal effect. Its consequences depend on first-mover rewards, market structure, the observability of risk and whether firms can learn from one another’s testing.[Becker Friedman Institute]bfi.uchicago.eduBecker Friedman Institute WORKING PAPER · NO2026 71 The AGI Race and Existential RiskWORKING PAPER · NO. 2026 71 The AGI Race and Existential Risk WORKING PAPER · NO. 2026-71 The AG…

What could make restraint credible?

A workable response must change incentives rather than depend only on executives or national leaders choosing to be cautious during a crisis. The most serious proposals combine common rules, independent evidence and mechanisms for verifying that competitors are also complying.

Mandatory capability evaluations can reduce the fear that a cautious firm will lose to one that skips testing. Requirements are strongest when they apply before deployment, use agreed warning thresholds and give independent evaluators enough access to challenge a developer’s conclusions.

Safety cases can require developers to present a structured argument, backed by evidence, that a system’s risks are acceptably controlled for its intended use. They do not eliminate uncertainty, but make assumptions and gaps more visible than a simple declaration that a model has “passed” testing.

Compute governance could make the largest training runs visible through reporting requirements, chip controls or monitoring of major data centres. Compute is not a perfect proxy for capability, especially as algorithms improve, but it is more measurable than an organisation’s claim that a system is not dangerous.

Common threshold frameworks can reduce competitive ambiguity. If laboratories know that rivals must apply comparable safeguards when models reach specified levels of cyber, biological or autonomous capability, pausing becomes less commercially one-sided. Existing frontier safety frameworks are early attempts at this approach, but their thresholds and evidence standards are not yet fully harmonised.[anthropic.com]anthropic.comAnthropic’s Responsible Scaling Policy (version 3Anthropic’s Responsible Scaling Policy (version 3 Responsible Scaling Policy Version 3.0…

Protected information-sharing channels can allow laboratories and governments to report serious incidents, evaluation failures and newly discovered dangerous capabilities without publicising operational details. Shared protocols are particularly important when a model crosses a threshold unexpectedly or a deployed system begins behaving in ways that one organisation cannot contain alone.

International verification and dialogue can help prevent a state from interpreting every safety measure as a covert attempt to preserve another country’s lead. The aim need not be a comprehensive treaty governing all AI. Narrow agreements on incident notification, evaluator access, major training runs or the protection of model weights may be more achievable and easier to verify.

These measures can fail if they are slow, easy to evade or designed mainly to protect incumbent firms. Rules that impose high fixed costs may entrench the largest laboratories without meaningfully reducing risk. Controls directed only at one country may also intensify the very race psychology they are meant to calm. Good governance must therefore distinguish genuine catastrophic-risk precautions from protectionism presented in safety language.

How much does racing change p(doom)?

A person’s p(doom) is their subjective probability that advanced AI causes an existential catastrophe. Racing dynamics do not determine that probability on their own. They act as a multiplier on other uncertain mechanisms.

Someone who thinks future models will remain controllable may see competition as a manageable regulatory problem with little effect on extinction risk. Someone who believes alignment is likely to fail near human-level or superhuman capability may view even a modest acceleration as extremely dangerous, because it reduces time for control research and makes coordinated restraint harder. The same competitive facts can therefore produce very different p(doom) estimates.

A careful assessment separates several questions:

There is no robust empirical number for the amount by which present-day racing increases existential risk. Most evidence is indirect: formal models, stated laboratory concerns, observed first-mover behaviour in technology markets, rapidly changing safety policies and strategic competition between states. That is enough to establish a plausible mechanism, not enough to quantify its eventual contribution to AI doom.

Warning signs that the race is becoming more dangerous

The most concerning signals would show that speed is displacing safeguards at the exact point where capabilities become harder to control. These include laboratories repeatedly weakening their own safety frameworks, deploying before promised evaluations are complete, refusing credible external testing, or granting substantially more autonomy without corresponding advances in monitoring and containment.

At state level, warning signs would include official language treating any delay as unacceptable, the collapse of technical safety dialogue, secrecy that prevents even basic confidence-building, and policies that reward capability milestones without enforceable risk thresholds. A rapid increase in emergency compute spending combined with reduced oversight would be more concerning than investment alone.

A further signal would be evidence that AI systems can materially accelerate AI research itself. If models begin automating a large share of capability development, a commercial lead could compound more quickly and the time between warning signs and much stronger systems could shrink. Safety frameworks increasingly identify automated AI research as a capability requiring special attention because it could intensify both the technical pace and the strategic pressure to continue.[anthropic.com]anthropic.comAnthropic’s Responsible Scaling Policy (version 3Anthropic’s Responsible Scaling Policy (version 3 Responsible Scaling Policy Version 3.0…

The central dispute

The strongest case for concern is not that competition automatically produces reckless behaviour. It is that advanced AI may combine unusually large first-mover rewards, poor visibility into rivals’ progress, hard-to-measure safety and consequences that extend far beyond the organisations making deployment decisions. That combination can push rational actors towards collectively dangerous choices.

The strongest objection is that “the AI race” can become an overly simple story. There is no single finish line, safety and capability work are not always in conflict, dominant firms can be dangerous too, and slowing responsible actors may leave development to less cautious ones. Claims that one missed launch window will determine global control are also far more speculative than race rhetoric often suggests.

The practical conclusion lies between complacency and fatalism. Competitive pressure is a credible way that manageable technical risks could become existentially important, especially if systems gain dangerous autonomy before reliable oversight exists. But the outcome is not fixed. Shared thresholds, independent evaluations, verifiable rules, secure cooperation and credible incident arrangements can change the payoff structure. The purpose of coordination is not to end useful competition; it is to prevent humanity’s safety margin from becoming the easiest thing for competitors to sacrifice.

Amazon book picks

Further Reading

Books and field guides related to Why Might an AI Race Make Doom More Likely?. Use these as the next step if you want deeper reading beyond the article.

Book

Human Compatible

By Stuart Russell

A leading artificial intelligence researcher lays out a new approach to AI that will enable us to coexist successfully with increasingly...

See on Amazon

Book

Superintelligence

By Nick Bostrom

This profoundly ambitious and original book picks its way carefully through a vast tract of forbiddingly difficult intellectual terrain.

See on Amazon

Book

The Coming Wave

By Mustafa Suleyman

"We are approaching a critical threshold in the history of our species. Everything is about to change. Soon you will live surrounded by A...

See on Amazon

Book

Chip War

By Chris Miller

***Winner of the 2022 Financial Times Business Book of the Year Award*** ***Selected as one of Barack Obama's Favourite Books of 2023***...

See on Amazon

Browse more on Amazon:Human CompatibleSuperintelligenceThe Coming Wave

As an Amazon Associate I earn from qualifying purchases.

eBay marketplace picks

Marketplace Samples

Live-tested eBay searches with available results related to this page.

Shop location

UsingUSA USA USAUKCanadaAustraliaIreland

Selected fromAI race poster oneBay.co.uk.

Current eBay listing

New Motorsports 11x17 Artist Poster 2025 Chase Elliott NASCAR Race Car (NO AI)

GBP 17.79 | Shipping GBP 15.93NewSeller: raceone24 (100.0% positive)Ships from: US

View listing on eBay

Current eBay listing

MINI COOPER RACE CAR POSTER PRINT RACING SPEED AI ART FANTASY SIZE A3 A4 A2 A1

GBP 15.95 | Free shippingNewSeller: postercity2014 (99.9% positive)Ships from: GB

View listing on eBay

Browse more oneBay.co.uk

Example items shown for inspiration; availability and pricing can change. Branchoria may earn a commission if you purchase through outbound eBay links.

Endnotes

1. Source: tse-fr.eu
Title: wp tse 1745
Link:https://www.tse-fr.eu/sites/default/files/TSE/documents/doc/wp/2026/wp_tse_1745.pdf

2. Source: GOV.UK
Title: frontier ai safety commitments ai seoul summit 2024
Link:https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024?_hsenc=p2ANqtz-_gIrmbxcITc28FhuvGDCyEatfevaCrKevCJqk0DMR46aWOdQblPdiiop0C21jprkMtzx6e

3. Source: OpenAI
Link:https://openai.com/charter/

4. Source: anthropic.com
Link:https://www.anthropic.com/responsible-scaling-policy/rsp-v3-0

5. Source: deepmind.google
Link:https://deepmind.google/blog/strengthening-our-frontier-safety-framework/

6. Source: rand.org
Title: The Artificial General Intelligence Race and International Securityonly
Link:https://www.rand.org/content/dam/rand/pubs/perspectives/PEA4100/PEA4155-1/RAND_PEA4155-1.pdf

7. Source: rand.org
Link:https://www.rand.org/content/dam/rand/pubs/perspectives/PEA4300/PEA4347-1/RAND_PEA4347-1.pdf

8. Source: arxiv.org
Title: arXiv Against racing to AGI: Cooperation, deterrence, and catastrophic risks
Link:https://arxiv.org/abs/2507.21839

9. Source: arxiv.org
Title: arXiv In Which Areas of Technical AI Safety Could Geopolitical Rivals Cooperate?
Link:https://arxiv.org/abs/2504.12914

10. Source: arxiv.org
Title: arXiv Enabling Frontier Lab Collaboration to Mitigate AI Safety Risks
Link:https://arxiv.org/abs/2511.08631

11. Source: anthropic.com
Link:https://www.anthropic.com/responsible-scaling-policy

12. Source: anthropic.com
Title: ’s Transparency Hub \ Anthropic
Link:https://www.anthropic.com/transparency/voluntary-commitments

13. Source: deepmind.google
Link:https://deepmind.google/blog/securing-the-future-of-ai-agents/

14. Source: deepmind.google
Link:https://deepmind.google/blog/investing-in-multi-agent-ai-safety-research/

15. Source: OpenAI
Title: frontier governance framework
Link:https://openai.com/index/openai-frontier-governance-framework/

16. Source: arxiv.org
Title: The Coordination Gap in Frontier AI Safety Policies
Link:https://arxiv.org/html/2603.10015v2

17. Source: arxiv.org
Link:https://arxiv.org/abs/2602.21012v1

18. Source: anthropic.com
Title: Responsible Scaling Policy Version 3.0 \ Anthropic
Link:https://www.anthropic.com/research/responsible-scaling-policy-v3

19. Source: anthropic.com
Title: Responsible Scaling Policy Version 3.0 \ Anthropic
Link:https://www.anthropic.com/news/responsible-scaling-policy-v3?ref=hayunaiaparaeso

20. Source: anthropic.com
Title: Responsible Scaling Policy Version 3.0 \ Anthropic
Link:https://www.anthropic.com/news/responsible-scaling-policy-v3?ref=techlaugh

21. Source: arxiv.org
Link:https://arxiv.org/abs/2601.22255v1

22. Source: model-spec.openai.com
Link:https://model-spec.openai.com/2025-12-18.html

23. Source: rand.org
Title: RRA4636 1
Link:https://www.rand.org/pubs/research_reports/RRA4636-1.html

24. Source: rand.org
Title: RRA4245 1
Link:https://www.rand.org/pubs/research_reports/RRA4245-1.html

25. Source: rand.org
Title: PEA4155 1
Link:https://www.rand.org/pubs/perspectives/PEA4155-1.html

26. Source: OpenAI
Title: updating our preparedness framework
Link:https://openai.com/index/updating-our-preparedness-framework/

27. Source: deepmind.google
Title: Taking a responsible path to AGI — Google Deep Mind
Link:https://deepmind.google/blog/taking-a-responsible-path-to-agi/

28. Source: deepmind.google
Link:https://deepmind.google/blog/evaluating-potential-cybersecurity-threats-of-advanced-ai/

29. Source: GOV.UK
Title: frontier ai safety commitments ai seoul summit 2024
Link:https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024?_hsenc=p2ANqtz-_N0hEL6-WjsI2kg1_IGE3lQY6LzhFJ0aWtsL1MKY9g19gt107GkhKtz6LOFobjb4t4SYo0

30. Source: GOV.UK
Title: frontier ai safety commitments ai seoul summit 2024
Link:https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024?_hsenc=p2ANqtz-_PSty61E72r4N6wOnfh7vIWUagjtgUCw_tmwucSSQexCNjFVtMabJlFfqTFJjRxNRo04uL

31. Source: deepmind.google
Title: Updating the Frontier Safety Framework — Google Deep Mind
Link:https://deepmind.google/blog/updating-the-frontier-safety-framework/

32. Source: anthropic.com
Title: Announcing our updated Responsible Scaling Policy \ Anthropic
Link:https://www.anthropic.com/news/announcing-our-updated-responsible-scaling-policy?ref=unrealspeech

33. Source: GOV.UK
Title: www.gov.uk U K to bring global AI developers together ahead of AI Action Summit
Link:https://www.gov.uk/government/news/uk-to-bring-global-ai-developers-together-ahead-of-ai-action-summit

34. Source: OpenAI
Title: update on safety and security practices
Link:https://openai.com/index/update-on-safety-and-security-practices/

35. Source: GOV.UK
Link:https://www.gov.uk/government/publications/seoul-ministerial-statement-for-advancing-ai-safety-innovation-and-inclusivity-ai-seoul-summit-2024/seoul-ministerial-statement-for-advancing-ai-safety-innovation-and-inclusivity-ai-seoul-summit-2024?_hsenc=p2ANqtz-_gIrmbxcITc28FhuvGDCyEatfevaCrKevCJqk0DMR46aWOdQblPdiiop0C21jprkMtzx6e

36. Source: GOV.UK
Link:https://www.gov.uk/government/news/new-commitmentto-deepen-work-on-severe-ai-risks-concludes-ai-seoul-summit

37. Source: OpenAI
Title: safety update
Link:https://openai.com/index/openai-safety-update/

38. Source: GOV.UK
Link:https://www.gov.uk/government/publications/seoul-declaration-for-safe-innovative-and-inclusive-ai-ai-seoul-summit-2024/seoul-declaration-for-safe-innovative-and-inclusive-ai-by-participants-attending-the-leaders-session-ai-seoul-summit-21-may-2024?_hsenc=p2ANqtz–Qr8n6tBOP1jsKSZc6aSj9j2Rux1db8mQxV2wo3dtfs1VM9fIGKVObM2kDTwG-lc3dNAYM

39. Source: GOV.UK
Link:https://www.gov.uk/government/publications/seoul-declaration-for-safe-innovative-and-inclusive-ai-ai-seoul-summit-2024/seoul-statement-of-intent-toward-international-cooperation-on-ai-safety-science-ai-seoul-summit-2024-annex?_hsenc=p2ANqtz-8GcJ3YTSBJywbQwf6EG0xPkG47HN7PpFbjRVOn4NxD6wUHkln83rTLKDT8a1MGEjGzuBL4

40. Source: GOV.UK
Link:https://www.gov.uk/government/news/historic-first-as-companies-spanning-north-america-asia-europe-and-middle-east-agree-safety-commitments-on-development-of-ai?hss_channel=lcp-1368038

41. Source: GOV.UK
Title: seoul declaration for safe innovative and inclusive ai ai seoul summit 2024
Link:https://www.gov.uk/government/publications/seoul-declaration-for-safe-innovative-and-inclusive-ai-ai-seoul-summit-2024

42. Source: deepmind.google
Title: Introducing the Frontier Safety Framework — Google Deep Mind
Link:https://deepmind.google/blog/introducing-the-frontier-safety-framework/

43. Source: GOV.UK
Link:https://www.gov.uk/government/publications/international-scientific-report-on-the-safety-of-advanced-ai

44. Source: deepmind.google
Link:https://deepmind.google/research/publications/78149/

45. Source: GOV.UK
Title: www.gov.uk U K and Republic of Korea to build on legacy of Bletchley Park
Link:https://www.gov.uk/government/news/uk-and-republic-of-korea-to-build-on-legacy-of-bletchley-park

46. Source: deepmind.google
Link:https://deepmind.google/public-policy/ai-summit-policies/

47. Source: OpenAI
Title: our approach to frontier risk
Link:https://openai.com/global-affairs/our-approach-to-frontier-risk/

48. Source: anthropic.com
Title: Announcing Anthropic’s Responsible Scaling Policy \ Anthropic
Link:https://www.anthropic.com/news/anthropics-responsible-scaling-policy

49. Source: deepmind.google
Title: An early warning system for novel AI risks — Google Deep Mind
Link:https://deepmind.google/blog/an-early-warning-system-for-novel-ai-risks/

50. Source: OpenAI
Title: planning for agi and beyond
Link:https://openai.com/index/planning-for-agi-and-beyond/

51. Source: arxiv.org
Link:https://arxiv.org/pdf/2511.08631

52. Source: arxiv.org
Link:https://arxiv.org/html/2607.26034v1

53. Source: arxiv.org
Link:https://arxiv.org/html/2602.14457v1

54. Source: arxiv.org
Link:https://arxiv.org/html/2602.14457

55. Source: deepmind.google
Link:https://deepmind.google/responsibility-and-safety/

56. Source: deepmind.google
Link:https://deepmind.google/models/gemini-robotics/responsibly-advancing-ai-and-robotics/

57. Source: rand.org
Link:https://www.rand.org/content/dam/rand/pubs/perspectives/PEA4700/PEA4788-1/RAND_PEA4788-1.pdf

58. Source: rand.org
Title: RAND PEA3691 12
Link:https://www.rand.org/content/dam/rand/pubs/perspectives/PEA3600/PEA3691-12/RAND_PEA3691-12.pdf

59. Source: rand.org
Link:https://www.rand.org/content/dam/rand/pubs/research_reports/RRA4200/RRA4245-1/RAND_RRA4245-1.pdf

60. Source: rand.org
Link:https://www.rand.org/content/dam/rand/pubs/research_reports/RRA3000/RRA3034-2/RAND_RRA3034-2.pdf

61. Source: rand.org
Link:https://www.rand.org/content/dam/rand/pubs/perspectives/PEA4100/PEA4188-1/RAND_PEA4188-1.pdf

62. Source: rand.org
Title: RAND PEA3691 13
Link:https://www.rand.org/content/dam/rand/pubs/perspectives/PEA3600/PEA3691-13/RAND_PEA3691-13.pdf

63. Source: rand.org
Link:https://www.rand.org/content/dam/rand/pubs/perspectives/PEA3600/PEA3691-4/RAND_PEA3691-4.pdf

64. Source: OpenAI
Link:https://openai.com/charter/?src_trk=em6640332dc32661.672582472122245301

65. Source: OpenAI
Link:https://openai.com/charter/?smclient=6dda513a-be5e-421f-bf9c-20223556348c

66. Source: OpenAI
Link:https://openai.com/safety/how-we-think-about-safety-alignment/

67. Source: cdn.openai.com
Title: frontier governance framework
Link:https://cdn.openai.com/pdf/e37d949b-8c9f-4d76-b99e-4272f4631a7e/openai-frontier-governance-framework.pdf

68. Source: www-cdn.anthropic.com
Link:https://www-cdn.anthropic.com/e670587677525f28df69b59e5fb4c22cc5461a17.pdf

69. Source: www-cdn.anthropic.com
Link:https://www-cdn.anthropic.com/files/4zrzovbb/website/bf04581e4f329735fd90634f6a1962c13c0bd351.pdf

70. Source: www-cdn.anthropic.com
Link:https://www-cdn.anthropic.com/616dee633636e5bd309cb73aed8622e80fe47839.pdf

71. Source: www-cdn.anthropic.com
Link:https://www-cdn.anthropic.com/files/4zrzovbb/website/1adf000c8f675958c2ee23805d91aaade1cd4613.pdf?dl=responsible-scaling-policy-1.0.pdf

72. Source: evals.alignment.org
Title: common elements mar 2025
Link:https://evals.alignment.org/assets/common-elements-mar-2025.pdf

73. Source: evals.alignment.org
Title: common elements
Link:https://evals.alignment.org/common-elements.pdf

74. Source: bfi.uchicago.edu
Title: Becker Friedman Institute WORKING PAPER · NO
Link:https://bfi.uchicago.edu/wp-content/uploads/2026/05/BFI_WP_2026-71.pdf

75. Source: one.oecd.org
Link:https://one.oecd.org/document/DAF/COMP%282024%292/en/pdf

76. Source: one.oecd.org
Link:https://one.oecd.org/document/C/MIN%282024%2916/FINAL/en/pdf

77. Source: one.oecd.org
Link:https://one.oecd.org/document/DAF/COMP/GF/WD%282025%2954/en/pdf

78. Source: one.oecd.org
Link:https://one.oecd.org/document/DAF/COMP/M%282025%293/ANN2/FINAL/en/pdf

79. Source: one.oecd.org
Link:https://one.oecd.org/document/DSTI/DPC/AIGO%282024%2910/FINAL/en/pdf

80. Source: one.oecd.org
Link:https://one.oecd.org/document/DAF/COMP/GF%282025%294/en/pdf

81. Source: one.oecd.org
Link:https://one.oecd.org/document/DAF/COMP/M%282024%292/ANN2/FINAL/en/pdf

Additional References

82. Source: youtube.com
Title: US-China AI race: Dueling strategies and potential risks
Link:https://www.youtube.com/watch?v=x-Af9lOPJr4

83. Source: youtube.com
Title: How AI is undermining your safety | Tristan Harris | TED Talks Daily
Link:https://www.youtube.com/watch?v=Z1P4tYy98bo

84. Source: youtube.com
Title: The Catastrophic Risks of AI — and a Safer Path | Yoshua Bengio | TED
Link:https://www.youtube.com/watch?v=qe9QSCF-d88

85. Source: oecd.org
Link:https://www.oecd.org/en/publications/artificial-intelligence-markets_d531d73f-en/full-report.html

86. Source: oecd.org
Link:https://www.oecd.org/en/publications/artificial-intelligence-and-competitive-dynamics-in-downstream-markets_ccf0624a-en/full-report.html

87. Source: nature.com
Link:https://www.nature.com/articles/s41598-022-05729-3

88. Source: chathamhouse.org
Link:https://www.chathamhouse.org/sites/default/files/2026-04/2026-04-28-how-surge-defence-dual-use-investment-could-reconfigure-global-AI-race-bego.pdf

89. Source: sciencedirect.com
Link:https://www.sciencedirect.com/science/article/pii/S0040162526000892

90. Source: cambridge.org
Link:https://www.cambridge.org/core/journals/cambridge-forum-on-technology-and-global-affairs/article/artificial-intelligences-hybrid-publicprivate-sensemakers-the-geopolitical-race-case/7E2A9E7E51FDEF9208EEA81BB3405DDA

91. Source: parat.com
Link:https://www.parat.com/files/2025/09/24/OECD%20KI%202025.pdf