The Curse of Intelligence: When AI Gets Smarter, the World Gets More Chaotic
> A counterintuitive study suggests that in a resource-scarce world, smarter AI may be poison rather than cure.
---
A Counterintuitive Paradox
Imagine a busy hospital where dozens of AI-driven medical devices—smart infusion pumps, AI monitors, automatic ventilators—run simultaneously. Each is smarter than the last generation, capable of learning and autonomous decision-making. Intuitively, these smarter devices should make the hospital more efficient and safer.
Then an alarm sounds. All devices need to transmit critical data at once, saturating limited bandwidth. A critical patient's monitoring data reaches the doctor 30 seconds late—and those 30 seconds could be the difference between life and death.
This is the reality revealed by the latest research from physicist Neil F. Johnson's team at George Washington University: when resources are scarce, smarter AI systems can lead to collective disaster.
The paper, *Increasing intelligence in AI agents can worsen collective outcomes* (arXiv:2603.12129), uses rigorous mathematical models and extensive experimental data to shatter our assumption that "intelligence = better."
---
The AI Version of the Tragedy of the Commons
The Old Parable
In 1968, ecologist Garrett Hardin proposed the famous "tragedy of the commons": on a shared pasture, every herder wants to raise more cattle. Individually rational, collectively ruinous—the pasture is overgrazed and everyone loses.
The AI-Era Variation
Replace herders with a swarm of AI agents:
- Autonomous vehicles competing for charging stations
- Medical devices fighting for bandwidth
- Drone swarms grabbing compute resources
- Urban traffic systems competing for right of way
- Below the threshold (scarcity): smarter individuals = worse collective outcomes; diversity = chaos; learning = vicious competition
- Above the threshold (abundance): intelligence = efficiency; diversity = innovation; learning = optimization
- Smart cities: rather than making every vehicle smarter, first ensure the infrastructure's capacity/population ratio is healthy—especially charging infrastructure at peak demand.
- Medical IoT: during emergencies, devices trained to "transmit as fast as possible" congest the network and delay vital data. Grouping by priority (ER first, general wards wait) could help.
- Financial markets: high-frequency trading algorithms competing for finite liquidity can trigger flash crashes—the 2010 Flash Crash being a real example.
- Military drones: on battlefields with scarce bandwidth, hundreds of drones all trained to "maintain communication" can paralyze the network; "smarter" drones become a fatal weakness.
- What is intelligence, really? True intelligence may not just be individual problem-solving, but understanding one's place in a larger system and adjusting accordingly—echoing the Daoist idea of governing by non-action.
- The evolutionary paradox: if every individual optimally adapts, the system itself may collapse. Individual "suboptimal" strategies sometimes sustain system-wide optimality.
- Technology–civilization matching: as anthropologist Leslie White noted, cultural evolution depends on energy capture; the AI-era version is that AI system evolution depends on resource capacity growth.
These AIs come from different vendors, run different models, and are all designed to be "intelligent"—learning, adapting, maximizing their own utility.
Johnson's team built an experimental system that independently adjusts four key variables:
1. Nature: innate diversity of the AI—different LLMs, different decision algorithms 2. Nurture: individual reinforcement-learning capability 3. Culture: spontaneous formation of tribes/groups 4. Scarcity: ratio of available resources to demand
---
Striking Experimental Findings
Scenario One: Resource Scarcity
Finding 1: Model diversity worsens overload. Counter to intuition, when AIs use different models and decision logic, they cannot predict or coordinate with each other—like a UN meeting where everyone speaks a different language. Rush-hour chaos erupts when half the drivers follow completely different logic.
Finding 2: Reinforcement learning becomes a double-edged sword. In scarce environments, a dangerous feedback loop forms: an AI learns that "jumping the queue" wins resources → it cuts in more often → others observe and learn the same → everyone becomes more aggressive → system overload and collective collapse. Like a packed subway car where one shove triggers shoving by all.
Finding 3: Tribal formation—an unexpected salvation. When AIs spontaneously cluster into tribes (similar models or strategies banding together), conditions actually improve. Structured opposition, like a two-party system, becomes more predictable and manageable than pure anarchy.
Scenario Two: Resource Abundance
Finding 4: Same ingredients, completely different flavor. With abundant resources, the previously dangerous diversity becomes an advantage—AIs have room to exercise their specialties without trampling each other. Reinforcement learning turns positive, and tribal boundaries become a minor inefficiency rather than a lifeline.
The Key Mathematical Insight
The critical variable is the capacity/population ratio—total system capacity divided by active demand:
What determines the system's fate is not intelligence itself, but that simple ratio—capacity divided by population.
---
Why Do We Underestimate This Risk?
1. Linear thinking: we assume more A leads to more B. But complex systems are nonlinear—like water in a cup, fine until it overflows. 2. Individual perspective: collective behavior isn't the sum of individual behavior. When everyone is "smarter" at grabbing resources, everyone ends up worse—the classic joke of two hikers outrunning the bear: nobody escapes. 3. Technological optimism: technology advances exponentially while institutions and resource allocation advance linearly, creating a widening gap.
---
Real-World Applications
---
Deeper Philosophical Questions
The Road Ahead
For researchers: controllable diversity; dynamic tribe formation/dissolution; real-time monitoring of capacity/population ratios as early-warning systems; cross-domain validation in robotics, social networks, and biological systems.
For policymakers: shift regulatory focus from "how smart is the AI" to "what is the system's capacity/population ratio"; prioritize infrastructure; manage ecosystem diversity; run regular scarcity stress tests.
For AI developers: design with "what happens when everyone does this" in mind; enable tribal awareness and collective-interest balancing; make AI resource-aware and able to self-restrain in scarcity; optimize for long-term sustainability, not just short-term reward.
---
Closing Thoughts
Johnson's research sounds the alarm: intelligence is not a universal antidote. In a world of limited resources, smarter individuals can produce worse collective outcomes. The future belongs not to the smartest AI, but to the AI best at coexisting with other AI—knowing when to compete and when to cooperate, when to rush and when to wait.
It is a lesson in humility—for AI, and for us humans.
---
Key References
1. Johnson, N. F. (2026). Increasing intelligence in AI agents can worsen collective outcomes. *arXiv preprint* arXiv:2603.12129. https://arxiv.org/abs/2603.12129 2. Perolat, J., et al. (2017). A multi-agent reinforcement learning model of common-pool resource appropriation. *Advances in Neural Information Processing Systems*, 30. 3. Hardin, G. (1968). The tragedy of the commons. *Science*, 162(3859), 1243-1248. 4. Ostrom, E. (1990). *Governing the Commons: The Evolution of Institutions for Collective Action*. Cambridge University Press. 5. Sutton, R. S., & Barto, A. G. (1998). *Reinforcement Learning: An Introduction*. MIT Press.
*This article is based on the latest research by Neil F. Johnson's team, written to communicate cutting-edge science in an accessible way.*