Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 25·18 min read
Strategic Coercion Within Alliances: The Greenland Sovereignty Game as an AI Stress Test
TL;DR
Large language models (LLMs) lack tested frameworks for modeling coercive diplomacy within alliances—situations where a dominant power pressures weaker allies. Researchers simulated the Greenland sovereignty crisis using eight frontier LLMs playing six geopolitical roles across 3,604 games, recovering structural utility parameters (self-interest, reciprocity, inequality aversion, norm respect, commitment) via inverse game theory.
✦ Why It Matters
Engineers can use inverse game theory to audit LLM geopolitical reasoning and detect escalation vulnerabilities before deployment in advisory systems.
Key Takeaways
- Test LLMs under coercion scenarios; baseline escalation rates may triple under pressure framing.
- Extract utility weights for reciprocity and norm respect; Chinese and Western models differ systematically.
- Use jus cogens and self-determination prompts to reduce escalatory behavior in geopolitical contexts.
- Benchmark LLMs on structural game-theoretic outcomes, not just action frequencies, for safety validation.
Full Summary
Current LLM benchmarks measure action frequency but lack structural understanding of how language models reason about strategic coercion—pressure exerted by dominant alliance members on weaker partners. Researchers developed three game-theoretic models (asymmetric coercion, NATO assurance with tipping points, triadic extensive-form games incorporating social preferences) and deployed them in multi-agent simulations where eight frontier LLMs (including GPT-4, Claude, DeepSeek V3.2) played the United States, Denmark, Greenland, NATO, Russia, and Canada across 3,604 completed games.
Using inverse game theory, they extracted each model's utility weights for five behavioral dimensions: material self-interest, reciprocity, inequality aversion, norm respect, and commitment consistency. Key findings: escalatory four-action sequences rose from 10.7% to 28.6% under coercion framing; Chinese-origin models weighted power differently than Western models when playing the U.S. role; only 3 of 8 models ever achieved peaceful acquisition (1.9% success rate), with DeepSeek V3.2 executing a stable five-round diplomatic sequence.
Prompts emphasizing jus cogens (peremptory international norms) and self-determination reduced escalation back to baseline levels.