The Battlefield Nervous System

Military discussions of artificial intelligence often begin with drones because drones are visible. The deeper transformation is happening in the connective tissue.

Modern armed forces collect enormous amounts of information from satellites, aircraft, unmanned systems, radar, electronic sensors, human reporting, cyber systems and commercial sources. The problem is no longer simply whether information exists.

The problem is whether it can be converted into a useful decision before it becomes stale. The Department of Defense has spent years pursuing variants of this idea under concepts such as Combined Joint All-Domain Command and Control: connect sensors across services, make data interoperable, fuse information, and give commanders a common operating picture quickly enough to act. Artificial intelligence changes each stage.

The diagram matters because it separates the layers. An AI system that detects a vehicle in an image is not necessarily selecting a target.

A system that recommends additional collection is not necessarily authorizing surveillance. A command platform that presents possible courses of action is not necessarily making the command decision.

But when all of those systems are connected, machine outputs can accumulate influence long before a human reaches the final decision. The decisive question becomes less:

Did AI pull the trigger? and more: How much of the evidentiary and decision environment was already constructed by machines before the human arrived?

Project Maven Grows Up

Project Maven began in 2017 as an effort to use computer vision to help analysts process the flood of full-motion video produced by military drones. At the time, the concept seemed narrow: teach algorithms to identify objects in imagery so humans did not have to watch every frame.

The Pentagon’s own early explanation emphasized that the system would extract objects of interest from still and moving imagery and complement human operators rather than autonomously select targets. Maven did not remain narrow.

NGA says the GEOINT portion of the program moved to the agency in 2023. Today it describes Maven as an operational AI capability that automatically detects, identifies, characterizes, extracts and attributes features and objects in imagery and video. The agency says the system is producing large volumes of detections for warfighter requirements across multiple operational locations.

Those detections do not stay inside NGA. They flow into other military platforms.

NGA describes Maven as a thread connecting sensors from multiple branches of the armed forces into a unified AI network. That phrase reveals how far the program has traveled from its origin.

The important product is no longer an algorithm that recognizes vehicles. It is an architecture that moves machine-generated observations through military workflows.

Earlier NGA testimony reported more than 20,000 active Maven users across more than 35 service and combatant-command tools. NGA also reported that Maven reduced some targeting-workflow timelines by as much as 80 percent in an exercise, moving one targeting cell from hours to minutes from sensing to target engagement.

Those are agency-reported operational metrics, not independent combat measurements. But even treated cautiously, they illustrate the military value proposition:

latency is becoming a weapon.

The Pentagon has reinforced Maven’s institutional status with money. In 2024 it awarded Palantir a contract with a ceiling of $480 million for the Maven Smart System prototype. In 2025 the Army raised the ceiling through a $795 million modification. Reuters reported in March 2026 that Pentagon leadership intended to make Maven an official program of record, which would entrench it as a long-term military capability rather than an experiment. That evolution—from prototype, to operational software, to program of record—is the story of military AI adoption in miniature.

NGA: When the Map Starts Thinking

Few agencies illustrate the intelligence problem better than the National Geospatial-Intelligence Agency. NGA sits at the intersection of maps, satellite imagery, location data, intelligence and military operations. It is responsible for turning geospatial information into intelligence that commanders and policymakers can use.

The agency says the amount of available GEOINT data could roughly triple over the next five to ten years as government and commercial satellite programs mature. More collection does not automatically produce more knowledge.

A satellite image only matters if someone or something can inspect it, compare it with other information, recognize what changed, judge significance and deliver the result while it still matters. NGA’s response is to make AI part of the entire cycle.

Its public AI strategy identifies four broad objectives: improve computer vision and geolocation; integrate AI into the analyst workforce; use AI to inform which sensors collect what next; and build shared AI infrastructure for models, labels, detections and data. That third objective is especially important.

Traditional intelligence collection can be understood as a human asking a sensor to look somewhere. An AI-assisted system can continually evaluate standing intelligence needs, constellation constraints and dynamic events, then recommend which collection option should be used next.

That closes a feedback loop:

NGA’s ASPEN program adds another layer. The agency describes it as a modernization effort for intelligence production designed to monitor known behavior at known locations while also searching for unknown behavior in unknown locations.

This is pattern-of-life analysis at machine scale. The analytical promise is obvious.

The institutional risk is equally important. If an AI model establishes the baseline of what is “normal,” then the model’s assumptions help define what appears suspicious. That makes model training, provenance, false positives and feedback loops matters of intelligence tradecraft, not simply software quality.

The Intelligence Attention Problem

The classic intelligence problem was scarcity. There were too few satellites.

Too few intercepts.

Too few sources.

Too little information about denied places. Today some parts of intelligence face the opposite problem.

Commercial satellites revisit locations repeatedly. Sensors generate persistent streams. Open-source information arrives continuously. Communications, imagery and geospatial databases can contain more material than teams of analysts can inspect manually.

That changes the scarce resource.

The scarce resource becomes human attention. AI solves that problem by filtering.

But filtering is never neutral.

A model decides which frames are likely to contain an object. A ranking system decides which detections deserve attention.

A language model summarizes fifty reports into five bullets. An anomaly detector decides which activity differs enough from the baseline to flag.

An automated collection system decides which sensor option appears most useful. Each step can increase analyst productivity.

Each step can also hide evidence.

This leads to one of the central doctrines of machine-assisted intelligence: The system that chooses what the analyst sees can influence judgment without ever making the final judgment itself. That is why provenance becomes so important.

An analyst should be able to move backward from a model-generated claim to the underlying imagery, observation, source, time, model version and confidence information. Otherwise AI transforms intelligence from an evidentiary process into an answer machine.

The Pentagon Goes AI-First

The institutional language surrounding military AI has changed. Years ago, Pentagon officials spoke about pilots, experiments and narrow use cases.

Today CDAO describes its mission as putting AI tools into the hands of warfighters at speed and scale. The Department’s current AI portfolio includes the Maven Smart System, an Agent Network for AI-enabled battle management and decision support, simulation programs, intelligence-to-capability pipelines, GenAI.mil and enterprise-agent development.

GenAI.mil illustrates the change from specialist tool to general infrastructure. CDAO says the platform provides Department-wide access to frontier generative-AI models at protected impact levels. By June 2026, CDAO said more than 1.6 million personnel had used the platform, generating tens of millions of prompts and deploying hundreds of thousands of AI agents.

Those numbers are Department claims and tell us more about institutional adoption than about mission effectiveness. But that is enough to make the larger point.

Military AI is no longer confined to data scientists. It is moving onto ordinary desks.

The June 2026 national-security AI memorandum accelerates that transition. It directs military and intelligence agencies to review procurement, onboard advanced models from multiple vendors, build secure compute infrastructure, establish cross-agency model and data exchanges, expand AI training and develop standardized testing and assurance. That is the administrative infrastructure of an AI-native national-security system.

Frontier Models Enter National Security

The first generation of military AI was dominated by specialized models. A computer-vision model detected a particular object.

A predictive-maintenance system estimated component failure. A logistics model optimized routing.

Frontier foundation models change that pattern because one model can perform many cognitive tasks through language, images, code and tools. CDAO moved aggressively in that direction in 2025 when it announced contracts with Anthropic, Google, OpenAI and xAI, each with a ceiling of $200 million, to develop agentic AI workflows for national-security missions.

The stated use cases include command and control, operational planning, logistics, weapons development and testing, uncrewed systems, intelligence, information operations and cyber operations. This does not mean a commercial chatbot is commanding weapons.

It means the Pentagon sees frontier models as general-purpose cognitive infrastructure. The appeal is understandable.

One model can summarize reports, generate code, search policy, translate foreign-language material, query databases, reason across documents and orchestrate tools. Agentic systems add another step: the model can call software, retrieve information and execute bounded workflows.

That makes the same technology useful in a headquarters, intelligence cell, logistics office and cyber operations center. It also makes the security problem much larger.

CIA and the Digital Intelligence Service

Public information about CIA operational AI is necessarily limited. That limitation should be respected rather than filled with speculation.

What the Agency does say publicly is revealing enough. CIA’s Directorate of Digital Innovation brings together information technology, data, artificial intelligence, cyber collection, cyber defense and open-source intelligence. The directorate was created in 2015 because the Agency concluded that digital technology had changed both the intelligence target and intelligence tradecraft.

That integrated structure matters.

Artificial intelligence in an intelligence agency is not merely an analytic product. It sits beside cyber capabilities, collection, data engineering and operational security.

CIA’s own professional journal, Studies in Intelligence, has increasingly examined AI’s effects on the profession. A 2026 essay on human intelligence argues that frontier AI is already changing the environment in which clandestine officers work and raises the possibility that AI-assisted surveillance, identity analysis and digital traces could make traditional tradecraft harder.

Another 2026 issue devoted multiple articles to generative AI and intelligence cognition. The important conclusion is modest but substantial:

AI changes both sides of espionage. It helps intelligence services search and synthesize information. It also helps counterintelligence services find patterns in the digital exhaust people leave behind.

NSA: AI as Capability and Attack Surface

The National Security Agency sees AI through two lenses simultaneously. One is opportunity.

The other is attack surface.

The NSA Artificial Intelligence Security Center says its mission is to promote secure AI development and adoption across National Security Systems and the Defense Industrial Base. Its definition of AI security reaches across training data, frameworks, models, model capabilities and the machine-learning lifecycle.

This framing is important.

When a military organization deploys an AI model, it has not merely installed software. It has created a new chain of dependencies:

NSA’s 2026 guidance on agentic AI warns that agents inherit the risks of language models while adding complexity and a larger attack surface because they can interact with tools and services. Its separate guidance on the Model Context Protocol focuses on exactly this problem: once AI systems can call databases, services and automation tools, security must govern not only what the model says but what the model can do.

This directly echoes the problem identified earlier in the Critical Technology Stack’s Article 003: identity, authorization, delegation and accountability become more important as autonomy increases. In a national-security environment the stakes are much higher.

A prompt injection that produces a bad summary is inconvenient. A manipulated agent with access to a mission system is a security event.

Machine-Speed Cyber Operations

Cyber conflict may be the domain where AI speed matters first. Code is already machine-readable.

Networks already operate at machine speed. Vulnerability discovery, malware analysis, log inspection, phishing detection and incident response all involve information volumes that exceed human attention.

In June, the Five Eyes cybersecurity agencies issued an unusually urgent statement saying frontier AI is changing both offensive and defensive cyber capability and warning that the relevant timeline is measured in months rather than years. That does not prove an autonomous cyberwar is imminent.

It does indicate that agencies responsible for defending some of the world’s most sensitive networks believe assumptions can become obsolete quickly. The contest is symmetrical.

Defenders can use AI to examine logs, discover suspicious behavior, generate detection rules, analyze malicious code and help operators investigate incidents. Attackers can use AI to scale reconnaissance, adapt social engineering, analyze code, search for weaknesses and automate portions of intrusion workflows.

The result is likely to be a cyber battlefield where humans increasingly supervise machines fighting other machines. The control problem is familiar:

How much authority should a defensive agent possess? May it isolate a host?

Revoke credentials?

Block network traffic?

Alter a firewall?

Patch a system?

Launch an active response? Every step upward in authority creates faster defense and a larger blast radius if the agent is wrong or compromised.

From Decision Support to Command Architecture

The military value of AI is often described as “decision advantage.” The phrase sounds abstract. In practice it means seeing something sooner, understanding it faster, generating options more quickly and acting before the opponent completes the same cycle.

This turns AI into part of command architecture. The Maven Smart System already fuses sensor and intelligence data into workflows used for target identification and strike approval. CDAO’s current portfolio describes an Agent Network aimed at AI-enabled battle management and decision support from campaign planning through kill-chain execution.

The Pentagon’s earlier CJADC2 work pursued the same objective without requiring generative AI: connect sensors and decision makers across services and domains. Frontier models and agents make that architecture more conversational and more automated.

A commander may eventually interact with a complex operational data environment by asking: What changed? Which observations are corroborated?

Which units are at risk?

What collection gaps remain?

What logistics constraint will bind first? Which course of action conflicts with current rules?

The danger is obvious.

A fluent answer can feel authoritative even when its evidentiary foundation is weak. Military AI therefore needs something ordinary consumer AI often lacks:

traceable reasoning inputs. The user must be able to move from answer to evidence.

Autonomous Weapons: The Language Matters

Public debate often collapses several different technologies into the phrase autonomous weapon. That creates confusion.

A drone can autonomously hold altitude without autonomously selecting a target. A missile can autonomously navigate toward designated coordinates without autonomously deciding whom to attack.

A defensive system can automatically respond to incoming threats inside tightly defined parameters. A reconnaissance platform can autonomously choose a route while leaving every use-of-force decision to humans. These are different allocations of authority.

DoD Directive 3000.09 remains the foundational public policy for autonomy in weapon systems. It requires autonomous and semi-autonomous systems to be designed so commanders and operators can exercise appropriate levels of human judgment over the use of force. It also requires realistic testing, understandable human-machine interfaces and compliance with the law of war, rules of engagement and other applicable requirements.

The phrase appropriate levels is deliberate. Congressional Research Service analysis notes that the policy does not require the same form of manual control in every case. Human judgment can include decisions about when, where, why and under what constraints a weapon operates.

That nuance is essential.

The meaningful policy question is not whether a human touches every control. It is whether humans retain genuine authority over the conditions under which lethal force is used. The June 2026 national-security AI memorandum directs the Department to update Directive 3000.09 and review it annually, reflecting how quickly the underlying technology is changing.

What Does Human Judgment Actually Mean?

There is a difference between a human being present and a human being meaningfully in control. Imagine a system that produces a target recommendation in three seconds.

The operator has ten seconds to approve it. The evidence consists of several machine-generated detections, an automated confidence score and a summarized intelligence report.

Technically, the human decides.

Practically, the human may be validating a machine conclusion he has no time to reconstruct. This is automation bias: the tendency to over-trust automated recommendations, especially when the system is normally correct and the human is overloaded.

Military environments amplify the problem because time pressure is real. Incoming threats move.

Targets disappear.

Adversaries jam networks.

Communications degrade.

Operators are tired.

Information is incomplete. The appropriate human-control question therefore has at least four dimensions:

If any of those disappear, “human in the loop” can become ceremonial. This point matters because the technology should be evaluated as part of the surrounding system rather than as an isolated claim or capability.

The Danger of Compressed Time

Military organizations have always sought to shorten decision cycles. AI changes the scale.

Computer vision can inspect imagery faster than teams of analysts. Automated correlation can combine sensor tracks instantly.

Language models can summarize reports in seconds. Agents can call tools without waiting for a person to manually move data between systems.

This can improve precision.

More information can be checked before acting. But speed can also produce escalation pressure.

If one military believes the other can detect and strike targets rapidly, it may feel pressure to delegate more authority to machines simply to remain competitive. That creates what might be called an automation tempo trap:

The danger is greatest when uncertainty and speed increase together. A defensive system responding to a known incoming missile may operate under very different assumptions from an AI model classifying ambiguous human behavior in a crowded environment.

The more ambiguous the target and the larger the consequences, the more valuable deliberate human judgment becomes. The strategic problem is that deliberate judgment consumes the very resource AI promises to save:

time.

The Private-Company Dependency

Military AI is being built during an unusual period in technological history. Many of the world’s most capable general-purpose models are not government systems.

They are commercial products developed by private companies. The Pentagon wants access to that innovation rather than trying to reproduce the entire frontier-model industry internally. Its multi-vendor contracting strategy reflects that logic.

But 2026 exposed the strategic downside. A dispute between the Pentagon and Anthropic over safety guardrails and acceptable military uses led the government to move away from the company. Reuters subsequently reported that parts of Palantir’s Maven workflows had been built around Anthropic’s Claude, forcing replacement work inside a strategically important military system.

The details of the dispute are political and legal as well as technical. The larger architectural lesson is durable.

A mission system is not sovereign if a private model provider can become an irreplaceable dependency. The White House’s June memorandum addresses this directly. It directs the national-security enterprise to use diverse suppliers and states that commercial entities should not possess the ability to prevent use of, disable, degrade or materially modify mission-dependent AI systems without government knowledge and approval.

This is not an argument for government-built models only. It is an argument for replaceability.

Decision Sovereignty

Article 007 introduced semiconductor sovereignty: the ability to preserve enough of the chip supply chain that no external chokepoint can switch off a nation’s technological future. Military AI creates an analogous concept.

Decision sovereignty is the ability of a state to retain control over the systems that shape high-consequence decisions even when underlying analytical components come from commercial providers. A sovereign military AI architecture should be able to replace one foundation model with another without rebuilding the entire command system.

It should keep identity, permissions and action authorization outside the model. It should preserve government-controlled logs.

It should retain evidence provenance. It should define rules and escalation pathways independently of vendor policy.

It should survive loss of network access to one provider. And the final authority for force should remain in the lawful chain of command.

The SURVXCOM Machine Warfare Authority Test

The most useful way to evaluate a military AI system is not to ask whether it contains “AI.” Ask where authority actually moves. This point matters because the technology should be evaluated as part of the surrounding system rather than as an isolated claim or capability.

What to Watch Next

The Machine-Speed Battlefield

The popular image of artificial intelligence at war is a robot deciding whom to kill. That image is dramatic.

It may also distract from the more immediate transformation. The battlefield is becoming a network in which machines increasingly determine what humans can see, what they notice first, how quickly evidence is fused and how rapidly options reach command.

The satellite takes the image.

The model finds the object.

The system compares behavior.

The network fuses the track.

The agent retrieves supporting reports. The command platform displays the option.

The human decides.

That final line remains essential.

But every line above it is changing. And if those systems reduce hours to minutes and minutes to seconds, then the human decision itself changes because it takes place inside a machine-shaped information environment moving at machine-shaped speed.

This is why military AI cannot be governed only at the trigger. Governance has to begin at the sensor.

It has to include data provenance, model validation, identity, permissions, auditability, cybersecurity, collection policy, vendor dependence and command authority. Artificial intelligence may make military operations faster.

It may make some operations more precise. It may make intelligence systems capable of processing information no human workforce could absorb.

It may also make errors travel farther and faster. The central challenge is therefore not choosing between artificial intelligence and human judgment.

It is designing systems in which artificial intelligence expands what humans can know without quietly replacing what humans are responsible to decide. Machine speed can create military advantage. Human accountability must remain faster than the temptation to surrender it.