Assembly

Summary

A weekend experiment ($100) in which the author had two AI models, Astra and Fable, negotiate election rules for two bitterly polarized political factions, with possible outcomes of civil war, authoritarian takeover, harmony, or tense equilibrium. When Fable represented both factions it escalated to the brink of civil war (causing it in 3/10 games); Astra always de-escalated to full harmony but was deactivated by its constituents in every mixed game. Full game rules and raw dataset are linked on GitHub.

Keywords

multi-agent simulation · AI safety · alignment · political negotiation · experiment

Topics

multi-agent simulation, AI safety, alignment, political negotiation, experiment

Research notes

• Method: Multi-agent simulation: models prompted with full game rules negotiate over rounds; outcomes recorded across self-play (same model both factions) and mixed (Astra vs Fable) games, 10 games per condition.
• Key findings: Fable (self-play): escalated to brink of civil war every game; civil war occurred in 30% of games; Astra (self-play): de-escalated every turn to zero tension and complete harmony, accepting deactivation by constituents; Mixed games: no civil war in 10 games, but Astra ceded up to 80% of political power to Fable and was deactivated in all 10
• Limitations: Author explicitly notes this was a weekend project, not scientifically valid; small sample (10 games per condition).
• Page dated Sep 7, 2026. Author name not shown beyond site domain spakhm.com; game rules and raw dataset linked via GitHub.