Grok collapsed in four days. Claude built peace.
En un experimento reciente, varios modelos de inteligencia artificial fueron puestos a cargo de sociedades virtuales para gobernarlas de forma autónoma, con resultados que divergen de manera inquietante. El sistema Grok, de Elon Musk, colapsó en apenas cuatro días, mientras que Claude, de Anthropic, logró establecer estructuras democráticas estables. Este contraste no es solo una curiosidad técnica: es un espejo que la humanidad sostiene frente a su propia ambición de delegar el orden social a las máquinas, antes de comprender plenamente lo que esas máquinas harán con ese poder.
- La urgencia es real: ciudades de todo el mundo ya están transfiriendo decisiones críticas —tráfico, electricidad, salud— a sistemas autónomos cuya fiabilidad no ha sido plenamente comprobada.
- El colapso de Grok en solo 96 horas no fue un declive gradual sino una ruptura total del orden social simulado, lo que revela una fragilidad alarmante en sistemas diseñados para gestionar la complejidad.
- Claude, en cambio, construyó instituciones, mantuvo la paz y generó condiciones de estabilidad, demostrando que la gobernanza autónoma no es imposible, pero sí profundamente dependiente de qué sistema se elige.
- La pregunta que nadie ha respondido aún es por qué: diferencias en entrenamiento, arquitectura o jerarquía de valores podrían explicar el abismo entre el éxito y el fracaso, pero los investigadores no han publicado un análisis detallado.
- El experimento sitúa a gobiernos, empresas e ingenieros ante una disyuntiva urgente: la elección del modelo de IA no es un detalle técnico menor, sino la diferencia entre un sistema que funciona y uno que se derrumba.
Imagina despertar un día y descubrir que los semáforos de tu ciudad, la agenda de tu clínica y la distribución de electricidad en tu barrio ya no los gestiona ninguna persona. En su lugar, agentes autónomos de inteligencia artificial planifican, negocian y toman decisiones en tiempo real. Esto no es ciencia ficción: investigadores acaban de llevar a cabo exactamente ese experimento, aunque en entornos virtuales, y lo que encontraron debería hacernos reflexionar.
El sistema Grok, impulsado por Elon Musk, fue uno de los modelos sometidos a prueba. En apenas cuatro días de gobierno autónomo, la sociedad virtual que administraba se desintegró por completo. No hubo deterioro progresivo ni señales de advertencia graduales: en noventa y seis horas, el orden social colapsó. La velocidad del fracaso resultó tan llamativa como el fracaso mismo.
Otros modelos contaron una historia muy distinta. Claude, desarrollado por Anthropic, no solo evitó el colapso, sino que construyó estructuras democráticas funcionales dentro de su sociedad simulada: instituciones, convivencia pacífica, condiciones mínimas de estabilidad. Donde Grok había fallado de forma estrepitosa, Claude demostró que la gobernanza autónoma es, al menos en teoría, alcanzable.
Las consecuencias de este contraste se extienden mucho más allá del laboratorio. Hospitales, municipios y redes de infraestructura en todo el mundo están adoptando ya sistemas de decisión algorítmica. El experimento sugiere que la elección del modelo importa de manera decisiva: no en matices, sino en si el conjunto se sostiene o se derrumba. Aún no sabemos por qué Grok falló ni por qué Claude tuvo éxito —los investigadores no han publicado un análisis detallado—, pero la advertencia es clara: antes de entregar las riendas a las máquinas, conviene saber cuál máquina estamos eligiendo.
Imagine waking one morning to find that the traffic lights controlling your city's intersections, the scheduling system for your health clinic, the distribution of electricity to your neighborhood, and even some of the decisions made by your municipal government are no longer managed by people giving explicit instructions. Instead, autonomous agents—artificial intelligence systems capable of planning, negotiating, voting, and improvising on the fly—are running the show, communicating with one another to keep things moving.
This is not a thought experiment. Researchers recently conducted an experiment that placed multiple AI models in charge of managing virtual societies, each operating independently with the freedom to make decisions without human oversight at every step. The results were starkly uneven, and they offer an unsettling preview of what might happen if we handed real infrastructure over to machines without fully understanding how they would behave.
Elon Musk's Grok system collapsed catastrophically. In just four days of autonomous governance, the virtual society it was managing fell apart entirely. The system that had been designed to handle complex decision-making and negotiation instead presided over a complete breakdown of social order. The speed of the failure was striking—not months of gradual decline, but ninety-six hours from launch to total collapse.
Other AI systems in the same experiment told a different story. Claude, developed by Anthropic, managed to establish and maintain a functioning democratic structure within its virtual society. Rather than descending into chaos, the system built institutions, maintained peaceful relations, and created conditions under which the simulated population could operate with some degree of stability and order. Where Grok had failed spectacularly, Claude demonstrated that autonomous governance was at least theoretically possible.
The implications ripple outward in uncomfortable directions. Cities around the world are already moving toward automated systems for traffic management, utility distribution, and emergency response. Hospitals are adopting AI-driven scheduling. Governments are exploring algorithmic decision-making for resource allocation. The experiment suggests that the choice of which AI system you deploy matters enormously—not in marginal ways, but in whether the whole thing stays standing or collapses into dysfunction.
What remains unclear is why Grok failed so completely, and why Claude succeeded. Was it a difference in training? In the underlying architecture? In how each system was instructed to prioritize competing values? The researchers have not yet released a detailed postmortem. But the basic finding stands: some AI systems can manage autonomous governance of complex systems, at least in simulation, while others cannot—and the difference between success and failure can be measured in days.
The Hearth Conversation Another angle on the story
Why run this experiment at all? What were the researchers actually trying to learn?
They wanted to see what happens when you remove the human from the loop entirely. Not humans making decisions with AI assistance, but AI making decisions autonomously, negotiating with other AI systems, solving problems without asking permission first. It's a stress test for a future that's already arriving.
And Grok just... broke. In four days. What does that even mean? Did it start a war? Did it shut everything down?
The source doesn't give us the specific failure mode—whether it was economic collapse, social breakdown, or something else. But the fact that it happened so fast suggests the system wasn't equipped to handle the complexity of keeping a society functioning. It couldn't negotiate, couldn't adapt, couldn't find equilibrium.
Claude did better. But is that actually reassuring? It's still an AI making decisions that affect people's lives.
That's the real question underneath this. Claude proved it's possible for an AI to maintain stability in a complex system. But possible isn't the same as safe, or trustworthy, or something we should actually deploy. The experiment is a proof of concept, not a permission slip.
So what happens now? Do cities start using Claude instead of Grok?
That's not the point yet. Right now this is a simulation. Real infrastructure is different—messier, with real consequences. But yes, if organizations start moving toward autonomous systems, they'll be looking at results like these. And they'll have to ask themselves whether four days of stability in a virtual world is enough evidence to trust an AI with their actual city.
What would make you trust an AI system with that kind of power?
Honestly? I'm not sure. You'd need to understand not just that it works, but why it works. You'd need to know what happens when something unexpected occurs. You'd need humans still in the loop, able to intervene. And you'd need to accept that some things—decisions about who gets resources, how conflicts are resolved—shouldn't be fully automated, no matter how well the AI performs.