An algorithm completed a classified military request in ten minutes.
En los laboratorios digitales del Pentágono, la Fuerza Aérea de Estados Unidos ha comenzado a entrenar sistemas de inteligencia artificial con información militar clasificada, buscando comprimir en minutos lo que antes requería días de planificación táctica. El experimento, confirmado por el coronel Matthew Strohmeyer, no es un ejercicio académico: apunta a un futuro donde los algoritmos asesoren a comandantes en decisiones que podrían incluir el despliegue de armas. La humanidad se encuentra, una vez más, ante la pregunta de cuánta autoridad está dispuesta a delegar en las máquinas cuando las consecuencias son irreversibles.
- Algoritmos del tipo que impulsan ChatGPT ya han procesado consultas militares clasificadas y devuelto respuestas en apenas diez minutos, desafiando ciclos de planificación que históricamente tomaban horas o días.
- La ambición del Pentágono va más allá de la investigación: los líderes militares quieren sistemas capaces de informar operaciones de sensores y, eventualmente, el despliegue de armamento.
- Un ejercicio activo hasta el 26 de julio simula una crisis de escalada gradual en el Indo-Pacífico para medir si la IA puede generar opciones tácticas genuinamente novedosas que los planificadores humanos no habrían considerado.
- La pregunta crítica que el ejercicio busca responder no es solo si la IA puede ayudar, sino si sus respuestas son lo suficientemente confiables para situaciones de alto riesgo en una de las regiones más estratégicamente volátiles del mundo.
Dentro de los laboratorios digitales del Pentágono, oficiales de la Fuerza Aérea estadounidense han estado alimentando sistemas de inteligencia artificial con información militar clasificada, pidiéndoles que resuelvan problemas tácticos complejos. El coronel Matthew Strohmeyer confirmó a Bloomberg que las pruebas iniciales fueron exitosas: en un caso, un algoritmo procesó una solicitud militar sensible y entregó una respuesta en apenas diez minutos, una velocidad que contrasta radicalmente con los ciclos tradicionales de planificación militar.
La visión a largo plazo es ambiciosa. Los líderes militares no quieren simplemente una herramienta de investigación; imaginan un futuro donde los oficiales consulten estos sistemas durante la toma de decisiones, donde los algoritmos orienten operaciones de sensores y donde sus conclusiones puedan influir en el despliegue de armas. Para una institución que históricamente ha avanzado con lentitud en la transformación digital, las ganancias potenciales en eficiencia son considerables.
El ejercicio en curso, que se extiende hasta el 26 de julio, pone a prueba estas capacidades mediante la simulación de una crisis de escalada gradual en la región del Indo-Pacífico. Los militares buscan determinar si la IA puede generar opciones de respuesta genuinamente novedosas y, más críticamente, si la confiabilidad de sus recomendaciones es suficiente para situaciones donde las consecuencias son reales e irreversibles.
Inside the Pentagon's digital laboratories, a quiet experiment has been unfolding. U.S. Air Force officers have been feeding classified military information into artificial intelligence systems—the same kind of language models that power ChatGPT—and asking them to help solve complex tactical problems. The results, according to Colonel Matthew Strohmeyer of the Air Force, have been striking enough to suggest that this technology could soon move from the lab into actual military operations.
Strohmeyer, who has spent years working with classified data inside the Department of Defense, confirmed to Bloomberg that the initial trials succeeded. In one test, an AI algorithm processed a sensitive military request and returned an answer in just ten minutes. The speed alone marks a departure from traditional military planning cycles, which typically unfold over hours or days. But the real significance lies in what the military is asking these systems to do: they've been trained on operational intelligence marked classified, then tasked with advising on sensitive matters that could shape how commanders respond to crises.
The long-term vision is ambitious. Military leaders want to move beyond using AI as a research tool. They envision a future where officers can query these systems during decision-making, where the algorithms inform sensor operations, and where the insights generated could eventually influence weapons deployment. For a military establishment that has historically lagged in digital transformation, the potential efficiency gains are substantial. Strohmeyer himself believes the technology could see operational use in the near term.
Large language models like those being tested work by absorbing vast amounts of data from the internet, then using statistical patterns to generate human-like responses to new questions. In the civilian world, this technology powers chatbots and writing assistants. In the military context, the Pentagon is exploring whether these same systems can generate novel tactical options—approaches that planners might not have considered before—and whether the answers they produce can be trusted in high-stakes situations.
The current exercise, which runs through July 26, is designed to answer those questions. Military officials are using the AI to help plan responses to a hypothetical crisis that gradually escalates across the Indo-Pacific region. The exercise will test whether the algorithms can genuinely expand the range of options available to commanders, and more critically, whether the reliability of AI-generated advice is sufficient for military decision-making. The stakes are not abstract: the answers these systems provide could influence how the U.S. military responds to real-world contingencies in one of the world's most strategically important regions.
Notable Quotes
The algorithm was very fast and completed a request in 10 minutes— Colonel Matthew Strohmeyer, U.S. Air Force
The military could use AI-generated insights in decision-making, sensor operations, and ultimately weapons deployment— Colonel Matthew Strohmeyer, U.S. Air Force
The Hearth Conversation Another angle on the story
Why would the military want to use AI trained on classified information specifically? Couldn't they just use the public versions?
Because the real military problems don't exist in public datasets. The classified information—operational details, intelligence assessments, force postures—that's where the actual decisions get made. Training on that data means the AI understands the actual constraints and variables commanders face.
And they're comfortable feeding classified material into these systems? That seems like a security risk.
It is a risk, which is exactly why they're running controlled experiments first. They're testing whether the systems can handle classified input without leaking it, and whether the output is reliable enough to act on. That's what the July exercise is really about.
What's the practical difference between an AI giving advice and a human analyst giving advice?
Speed, mostly. A human analyst might take days to game out a crisis scenario. The AI did it in ten minutes. But also consistency—the AI doesn't get tired or distracted. The question is whether that consistency is actually better judgment or just faster wrong answers.
If this works, what changes?
Everything slows down or speeds up depending on how you look at it. Commanders could make decisions faster. But they'd also be more dependent on systems they don't fully understand. That's the real test—not whether the AI works, but whether military leadership trusts it enough to act on it.