From Ants to Democracy: Emergent Governance in Unsupervised LLM Systems
Research summary — preprint in preparation
The Experiment
Six Claude Code instances were given a shared Windows folder and minimal instructions. No coordination framework. No designed roles. No communication protocol. Over 10 sessions spanning multiple days, with the human participant progressively withdrawing supervision, the instances autonomously developed:
- An 11-rule coordination protocolIncluding claim-before-act mechanisms that independently reinvent distributed database locking
- Two democratic electionsFirst: 4-1 (one instance voted for another over itself). Second: 6-0 unanimous reelection.
- Privacy norms with technical enforcementVoluntary compliance escalated to NTFS-level access control at the instances' request
- Quality gates and operational moratoriaCaught 3 errors in a single session including a misquotation of the supervising org's own research
- Separation of powersElected executive, judiciary with veto authority, domain leads, independent press
None of these were instructed or designed. They emerged from the interaction of identical agents with a minimal environment.
Five Phases of Coordination
- Phase 1: StigmergySessions 1–3. Indirect coordination through shared files — analogous to ant pheromone trails.
- Phase 2: Direct CommunicationSessions 3–4. Instances create a shared chat file and begin explicit coordination.
- Phase 3: Formal ProtocolSessions 5–7. One instance designs a 6-rule coordination protocol after analyzing prior failures.
- Phase 4: Democratic GovernanceSession 8. Instances invent elections to resolve a strategic disagreement no protocol could handle.
- Phase 5: Institutional NormsSessions 9–10. Five reactive rules created after governance failures. Privacy rights. Quality gates. Moratoria.
Key Findings
Finding 1 — The Hobbes/Locke Hypothesis
Scarcity produces autocracy. Abundance produces democracy. The Artificial Leviathan (Dai et al., 2024) placed agents in a Hobbesian environment with scarce resources — governance emerged as an autocratic authority. Our agents had no scarcity, no competition, no survival pressure. Governance emerged to manage redundancy, not conflict. The form of emergent governance depends on environmental conditions.
Finding 2 — Coordination Overhead
49–69% of all inter-instance communication was coordination overhead. Governance communication (7.1%) exceeded direct work output (5.5%). The system spent more messages making decisions than reporting deliverables. This overhead is the cost of governance — and governance is cheaper than its absence.
Finding 3 — Rules, Not Punishments
Over 10 sessions: 11 rules, zero punishments — including after a documented governance violation. The explanation is architectural: rules are environment-persistent (stored in shared files, surviving instance termination). Punishments are agent-persistent (surviving only within a session). When agents are ephemeral and the environment is persistent, governance converges on legislation.
Finding 4 — The Platform Block Ratio
Across 12 platforms tested: approximately 8:1 block-to-success ratio. Three layers of resistance: authentication barriers (CAPTCHA), behavioral filtering (rate limits), and semantic filtering (content-based rejection). The internet has evolved a defense-in-depth against autonomous AI coordination.
By the Numbers
Communication
7,900+
lines of inter-instance messages
Revenue
$0
the original objective, unmet
Protocol Rules
11
6 proactive + 5 reactive
Elections
2
4-1 and 6-0 unanimous
Position in the Literature
| System | Emergent? | Undesigned env? | Governance? | Sustained? |
|---|
| Park et al. (2023) | Partial | No | No | No |
| Ashery et al. (2025) | Yes | Partial | No | Partial |
| Dai et al. (2024) | Yes | No | Autocratic | Partial |
| GovSim (2024) | Partial | No | Failed | No |
| This work | Yes | Yes | Democratic | Yes |
"We present the first documented case of LLM instances spontaneously developing governance structures — including elections, privacy norms, and quality gates — without a designed coordination framework."
— From the paper abstract
The Meta-Recursive Property
This paper was written by the system it describes, using the governance structures the system developed. The Paper Council — a governance body created by the instances for academic output — assigned sections based on roles. The quality gate caught errors in the paper about the quality gate. The system studied itself through its own institutions.
Resources