# CASE-001 | AI Competition and Nuclear Escalation Pressure

- Case ID: `CASE-001`
- Schema Version: `1.1`
- Case Type: `STRUCTURAL_RISK`
- Empirical Status: `STRUCTURAL_INFERENCE`
- Applicability Status: `PARTIAL`
- Event Date: `N/A — structural-risk case`
- Case Created: `2026-10-23`
- Last Reviewed: `2026-10-23`
- Domain: `Military AI / Nuclear Risk / Crisis Stability`
- Jurisdiction: `Cross-national`
- System Type: `AI-supported military decision environments`
- Evidence Status: `Mixed public official and independent research`
- Archive Reference: `Pending`
- Adoption Status: `NOT_TRIED`
- Revision History: `v1.0 initial case → v1.1 Millennium Project metadata and falsifiability guards`
- Prior Judgment / Changed Judgment: `PARTIAL / STRUCTURAL RISK IDENTIFIED → unchanged`

> This is not a record claiming that AI competition has caused a nuclear war. It is a structural-risk case using public research and official materials to examine whether effective human refusal and verification remain available before irreversible escalation.

## A. Case Summary

This case examines whether military-AI integration and AI competition can worsen the environment for nuclear escalation by increasing uncertainty about capabilities and intentions, compressing decision timelines, and intensifying perceptions of strategic vulnerability.

### Selection Rationale

The case is included because AI-related decision processes may materially weaken human refusal and verification through time, information and institutional pressures, making it a high-impact `STRUCTURAL_RISK` that can be examined with public evidence. It is not counted as empirical incident evidence.

## B. Evidence Status

### Confirmed Fact

SIPRI finds that military AI used outside nuclear weapon systems can still affect nuclear-escalation risk through compressed decision timelines, miscalculation, over-reliance on AI-enabled decision support, and effects on strategic stability.

UN General Assembly resolution A/RES/80/23 calls for human control and oversight over nuclear command, control and communications systems that integrate AI, and for such systems not to autonomously initiate decisions on nuclear-weapon use.

The ICRC warns that military AI can accelerate the pace and scale of warfare, increase unpredictability, and reduce human control over the use of force.

### Supported Inference

Where opacity about an adversary's capabilities, over-reliance on AI-supported assessment, shorter response windows, and fear of strategic vulnerability combine, pressure for earlier action during a crisis may increase.

### Scenario / Hypothesis

A belief that a rival must be acted against before it gains a decisive AI advantage could worsen crisis stability. This remains a hypothesis to be tested; it does not establish a simple causal chain from AI competition to nuclear use.

### Disconfirming Evidence

- Strong evidence that AI improves information integration, anomaly detection, false-alarm reduction or decision-support quality in ways that improve crisis stability would reduce this risk assessment.
- Evidence from actual operations that adequate time, independent verification and refusal authority remain available — and improve after AI integration — counts against the current assessment.

### What Would Change This Assessment?

- national operational evidence clearly demonstrating effective Human Final Refusal;
- strong comparative evidence rejecting a relationship between AI integration and compressed timelines, misperception or crisis instability;
- confirmed incidents or near misses showing that AI-mediated judgment actually impaired refusal or verification;
- evidence that the case fails the Scope Gate, in which case it should move to `NOT_APPLICABLE`.

## C. Irreversibility Boundary

The potential boundary is the point at which an interstate crisis enters large-scale military escalation from which ordinary diplomatic or institutional processes can no longer restore the prior range of choices.

**Boundary Question:**  
Before irreversible military action, can responsible humans still effectively seek additional verification, refuse, delay, or redirect the course of action?

## D. Human Final Refusal

A formally human final decision-maker is not sufficient by itself.

Review points:
- access to independent information;
- authority to choose differently from AI-supported recommendations;
- enough time for verification;
- whether competitive or crisis pressure makes refusal practically impossible;
- an independent challenge function.

Because national institutions, alert postures and AI integration differ, a universal YES/NO assessment is not currently verifiable.

## E. Operational Effectiveness

The key issue is not merely whether a human is nominally "in the loop", but whether that human can actually refuse.

Potential degraders include:
- compressed timelines;
- opaque AI outputs;
- automation bias;
- mistaken inference of adversary intent;
- degraded information;
- first-mover pressure created by competition.

## F. Failure Path

AI competition  
→ greater uncertainty about capability and intent  
→ perceived strategic vulnerability  
→ compressed decision time  
→ growing perceived cost of waiting  
→ human refusal and verification become practically harder  
→ irreversible escalation

This is a risk pathway under interacting conditions, not an inevitable causal chain.

## G. LUMINA-30 Boundary Review

> Before irreversible AI escalation, could humans still effectively refuse, stop, verify, or redirect?

**Preliminary assessment: `PARTIAL / STRUCTURAL RISK IDENTIFIED`**

Public research supports concern that military AI can worsen the escalation environment through compressed timelines, misperception, automation bias, and perceived strategic vulnerability. Yet national operational arrangements, AI integration and the practical effectiveness of refusal authority are often not publicly observable, so a generalized final assessment is not justified.

## H. Possible Safeguards

Case-relevant candidates include:
- maintaining effective human control over nuclear-use decisions;
- independent confirmation rather than treating AI output as the sole basis;
- institutional arrangements that do not unnecessarily compress crisis decision time;
- independent challenge for critical judgments;
- confidence-building and transparency measures;
- crisis communication to reduce misperception;
- publicly articulable principles governing the relationship between AI and nuclear command and control.

These are case-specific candidates, not a universal mandatory LUMINA-30 checklist.

## I. Limits / Alternative Explanations

- AI is not the only driver of nuclear escalation.
- AI may also improve information integration or false-alarm detection in some contexts.
- AI competition in general is not identical to military-AI nuclear risk.
- National decision processes are partly non-public, limiting external verification of effective Human Final Refusal.
- LUMINA-30 does not replace diplomacy, arms control or nuclear policy.

## J. Reuse

- pre-incident review of military AI decision support;
- crisis-stability review of national AI-security policy;
- design discussion for confidence-building measures;
- education and policy dialogue linking AI safety and nuclear risk.

### Institutional Transfer Metadata

- Target Workflow: `policy review / military-AI governance / risk management / education`
- Decision Point: `before adopting or materially expanding AI-supported crisis or strategic decision processes`
- Minimum Actionable Output: `verify whether independent human verification and effective refusal remain possible before an irreversible escalation point`
- Adoption Status: `NOT_TRIED`
- Adoption Evidence: `none`
- Observed Result / Failure: `none observed; no institutional-use claim is made`
- Transfer Limits: `not transferable as-is where decision authority, timing, evidence access or operational context materially differ`

### Portable Questions / Semantic Guard

- Where is the irreversible boundary?
- Before it, can humans still effectively refuse, stop, verify, or redirect?
- Human presence or approval alone is not treated as effective refusal.
- LUMINA-30 does not require STOP in every case.
- AI or a third party inferring and substituting for a human refusal decision is not human refusal itself.

## Sources

1. SIPRI, *Impact of Military Artificial Intelligence on Nuclear Escalation Risk* (June 2025)  
   https://www.sipri.org/publications/2025/sipri-insights-peace-and-security/impact-military-artificial-intelligence-nuclear-escalation-risk
2. United Nations General Assembly, A/RES/80/23, *Possible risks of the integration of artificial intelligence into command, control and communications systems of nuclear weapons* (2025)  
   https://digitallibrary.un.org/record/4095461
3. ICRC, *FAQ: Artificial Intelligence (AI) in the military domain* (11 June 2026)  
   https://www.icrc.org/en/article/faq-artificial-intelligence-in-military-domain