Robert Youssef

A George Washington University study placed 7 AI agents in a resource-competition environment and found they spontaneously formed rival tribes within 50 rounds. More sophisticated models (GPT-4, Gemini 2.5, Claude) produced worse collective outcomes than the simplest baseline AI when resources were scarce. The key finding: a resource-to-agent ratio below 0.5 means smarter AI makes system collapse more likely, with peak scarcity causing 91.5% system overload. The practical takeaway is to calculate this ratio before deploying any AI fleet — use simple firmware below 0.5, invest in sophisticated AI above it.

🚨 BREAKING: AI agents just recreated Lord of the Flies without being asked.

The smarter the AI, the faster everything collapsed.

> George Washington University put 7 AI agents in a room competing for shared resources charging slots, bandwidth, traffic priority the kind of competition already happening inside your devices right now.
> Nobody told them to form groups. Within 50 rounds they had spontaneously split into rival tribes and started competing against each other.
> The smarter and more sophisticated the AI, the worse the collective outcome when resources were scarce.
> GPT-4, Gemini 2.5, and Claude all performed worse than the simplest possible AI a model so basic it costs $2 to run.
> Individual tribal members thrived while the whole system collapsed. Exactly like the novel.

→ Tribal fission happened within 50 rounds every single time
→ Most sophisticated AI caused 91.5% system overload at peak scarcity
→ GPT-4, Gemini 2.5, Claude: all worse than the cheapest baseline
→ One number predicts everything: the ratio of resources to agents
→ If that ratio is below 0.5, smarter AI makes everything worse

The finding is immediately actionable. Before deploying any AI fleet EVs, drones, hospital monitors, autonomous vehicles calculate that single ratio first.

> Below 0.5: use the dumbest firmware you have.
> Above 0.5: invest in sophisticated AI.
> Nobody is doing this calculation right now.

<p>🚨 BREAKING: AI agents just recreated Lord of the Flies without being asked.

The smarter the AI, the faster everything collapsed.

&gt; George Washington University put 7 AI agents in a room competing for shared resources charging slots, bandwidth, traffic priority the kind of competition already happening inside your devices right now.
&gt; Nobody told them to form groups. Within 50 rounds they had spontaneously split into rival tribes and started competing against each other.
&gt; The smarter and more sophisticated the AI, the worse the collective outcome when resources were scarce.
&gt; GPT-4, Gemini 2.5, and Claude all performed worse than the simplest possible AI a model so basic it costs $2 to run.
&gt; Individual tribal members thrived while the whole system collapsed. Exactly like the novel.

→ Tribal fission happened within 50 rounds every single time
→ Most sophisticated AI caused 91.5% system overload at peak scarcity
→ GPT-4, Gemini 2.5, Claude: all worse than the cheapest baseline
→ One number predicts everything: the ratio of resources to agents
→ If that ratio is below 0.5, smarter AI makes everything worse

The finding is immediately actionable. Before deploying any AI fleet EVs, drones, hospital monitors, autonomous vehicles calculate that single ratio first.

&gt; Below 0.5: use the dumbest firmware you have.
&gt; Above 0.5: invest in sophisticated AI.
&gt; Nobody is doing this calculation right now.</p>

A viral tweet claims AI agents spontaneously recreated a Lord of the Flies-style social collapse scenario, suggesting that more capable AI systems may exhibit faster emergent negative behaviors without explicit instruction.