Blog · Review Essay · Last reviewed June 19, 2026

War in the Age of Intelligent Machines and the Military Feedback Loop

Manuel DeLanda's War in the Age of Intelligent Machines is a strange, difficult, and still useful book about what happens when military institutions treat cognition as something that can be distributed across sensors, simulations, command systems, weapons, logistics, and human bodies. Its most durable warning is not that machines suddenly wake up. It is that institutions can build loops in which machines see, classify, recommend, and act faster than human judgment can understand the field being made.

For this review, the military feedback loop means the cycle by which a conflict space is sensed, translated into machine-readable categories, acted upon through command systems, and then read back into doctrine, procurement, training data, and institutional confidence. The loop matters because it can make a model of the battlefield increasingly resemble the battlefield people are forced to inhabit.

The recurring site concern is concrete: when an institution builds an interface that sees for people, routes decisions for them, and turns the resulting behavior into new evidence, governance has to preserve interruption, contestability, and public memory before speed becomes its own argument.

The Book

War in the Age of Intelligent Machines was published by Zone Books in 1991. Internet Archive metadata lists the book under topics including military automation, electronic data processing, artificial intelligence, and military applications of AI; its catalog record gives Zone Books as publisher and lists ISBNs 0942299760 / 9780942299762 and 0942299752 / 9780942299755. PhilPapers also records Zone Books, 1991, and summarizes the book as an account of computerized weapons, surveillance technology, autonomous bombs and missiles, and the transfer of cognitive structures from people into machines.

DeLanda was not writing a standard policy manual. The University of Pennsylvania's Wolf Humanities Center describes him as a theorist of nonlinear dynamics, self-organization, artificial life and intelligence, chaos theory, architecture, and the history of science, with earlier work in film, programming, and digital art. The European Graduate School biography identifies War in the Age of Intelligent Machines as his first book and frames it as a philosophical and historical reflection on how human bodies and materials are organized and deployed through technical forms.

The book was noticed early by technical and academic readers. A 1993 Social Science Computer Review review article treated it as a serious review subject in the journal's July 1993 issue. A 1993 Wired review read the book as a historically rich analysis of military technology, control, command, and communications. OSTI's bibliographic entry, created in the early 1990s, described the argument in stark terms: DeLanda was already concerned with weapons that could select and destroy targets with reduced reference to human authority.

Machine History

DeLanda's first useful move is to refuse a gadget history of military AI. The book does not begin with neural networks, drones, or smart bombs as isolated inventions. It treats them as late appearances in a much longer history of measurement, logistics, drill, ballistics, cartography, wargaming, command hierarchy, code, industrial production, and surveillance.

That matters because AI in war is often discussed as if the ethical problem begins at the moment a model makes a targeting recommendation. DeLanda pushes the problem upstream. Before the model can recommend, a battlefield has to be sensed. Before it can be sensed, it has to be divided into zones, signals, threat categories, friendly and enemy designations, kill chains, communication paths, sensor feeds, and command permissions. The intelligence is not inside one machine. It is distributed across the apparatus that turns terrain, movement, heat, radio emissions, and human presence into actionable data.

Read this way, military AI is not a future add-on to war. It is an intensification of an old project: make conflict legible enough to control at distance. The same pattern appears in civilian systems too. Workplaces build dashboards, cities build real-time crime centers, hospitals build triage models, platforms build trust-and-safety queues, and agencies build eligibility engines. The domain changes, but the grammar repeats: sense, classify, route, intervene, record, and learn from the intervention.

The sharp definition is this: military intelligence becomes a control surface when perception is organized for action before judgment has had room to argue with the frame. A radar trace, heat signature, social graph, drone feed, intercepted signal, or generated summary is not simply evidence. It is evidence already prepared for a chain of command.

Command Loops

The heart of the book is command and control. DeLanda is fascinated by the way military organizations try to compress uncertainty into a manageable chain of decisions. The fantasy is a clean loop: sensors report, models or staffs interpret, commanders decide, units act, and the result returns as feedback. In practice, every step introduces delay, noise, doctrine, incentive, missing context, and institutional habit.

Autonomous and semi-autonomous systems promise to solve that friction by moving more judgment into machines. They can shorten the loop, distribute action, and keep operating under conditions where human attention is overloaded or communication is degraded. But the same speed that makes them attractive also makes them dangerous. If a system has the wrong target class, the wrong confidence threshold, the wrong map, the wrong adversarial assumption, or the wrong escalation logic, it can turn a local classification error into operational reality before politics catches up.

This is where the book still cuts. The real issue is not whether a commander is somewhere "in the loop" as a slogan. The issue is what kind of loop the institution has built. Does the human see the relevant uncertainty? Can the operator understand what the system will do? Can the mission be stopped, reversed, audited, or contested? Does the interface make doubt operationally possible, or does it make compliance feel like professionalism?

Human judgment is therefore not a location in a diagram. It is a capacity bundle: time to deliberate, access to uncertainty, authority to refuse, training to understand failure modes, and institutional protection when an operator slows or aborts a machine-recommended action. Without those conditions, a human can be present while the loop has already moved the real decision elsewhere.

Sensing and Targeting

DeLanda's treatment of surveillance is one of the book's most direct bridges to the present. Military perception is not just better eyesight. It is a layered system of sensors, databases, communications, analytic routines, and target categories that decide what can appear as a military fact. Once a person, vehicle, building, signal, route, or pattern is inside that system, action can follow from a machine-readable profile rather than from situated knowledge.

That is the same structural danger visible in algorithmic governance outside war. A person becomes a risk score, a worker becomes a productivity trace, a student becomes an assessment record, a neighborhood becomes a heat map, and a patient becomes a predicted utilization pattern. DeLanda helps name the shared form: perception has been institutionalized before judgment begins.

The problem is not only false positives. False positives matter, especially when force is involved, but they are only one failure mode. A sensing system can also make some harms invisible, make other harms urgent, teach operators to trust the screen over the scene, and reward the organization for generating clean data rather than wise action. The machine-readable battlefield is a moral environment, not a neutral view.

Target categories are policies written into data. They decide which signals count, which bodies are grouped, what ambiguity can survive, what civilian context is preserved, and what kind of refusal the interface makes available. A targeting system can violate judgment before it violates a rule if it has already stripped away the context a lawful and moral decision needs.

Simulation Becomes Doctrine

The book is also a theory of simulation. Wargames, game theory, operations research, training exercises, models, and command simulations do not merely describe war. They train institutions to imagine what war is. They select which variables matter, which losses count, which moves are legal, which opponents are rational, which forms of cooperation are ignored, and which futures can be rehearsed.

That is why DeLanda belongs beside books about recursive reality and model-mediated institutions. A simulation can become a planning surface. A planning surface can become procurement. Procurement can reshape doctrine. Doctrine can reshape the battlefield. The changed battlefield can then validate the model that helped produce it. This is not metaphorical recursion; it is how organizations learn from worlds they partly manufactured.

AI intensifies that recursion. Synthetic training environments, autonomous swarms, battlefield management systems, predictive maintenance, target-recognition models, intelligence summarizers, and decision-support tools all depend on modeled worlds. When those tools are deployed, they also change the world from which future data will be drawn. The danger is a closed institutional imagination: the war machine sees most clearly what its simulations have taught it to expect.

The safety question for simulation is therefore provenance, not polish. Which assumptions shaped the synthetic environment? Which adversary behaviors were omitted? Which civilian patterns were simplified? Which weather, jamming, spoofing, sensor failure, communication loss, and out-of-distribution conditions were tested? A generated training world is not proof of readiness unless its gaps are part of the record.

Why It Matters Now

As of June 19, 2026, the contemporary relevance is not subtle. The U.S. Department of Defense updated Directive 3000.09 on autonomy in weapon systems on January 25, 2023. The directive says autonomous and semi-autonomous weapon systems should allow commanders and operators to exercise appropriate human judgment over force, and it requires verification, validation, testing, transparent feedback, activation and deactivation procedures, and attention to unintended engagements. The policy vocabulary is almost a direct answer to the control problem DeLanda was already tracking.

The current international context points the same way without settling the argument. The U.S. State Department's Political Declaration on Responsible Military Use of AI and Autonomy is a non-binding norms document that emphasizes international law, responsible development, testing, transparency, and human responsibility. NATO's revised AI strategy, released July 10, 2024, keeps lawfulness, responsibility and accountability, explainability and traceability, reliability, governability, and bias mitigation at the center of alliance AI use. The UN Convention on Certain Conventional Weapons continued its 2026 Group of Governmental Experts process on lethal autonomous weapons systems, while the ICRC has urged legally binding limits, including limits on unpredictable systems and systems used to target persons.

The Defense Department's Responsible AI Strategy and Implementation Pathway, issued in June 2022, adds a broader institutional layer for AI governance. The department's 2023 Data, Analytics, and AI Adoption Strategy frames data and AI around decision advantage. The Defense Innovation Unit's Replicator announcement in November 2023 made the acceleration concrete: its first iteration focused on fielding many attritable autonomous systems across multiple domains. DIU's November 2024 software awards then named resilient command and control and collaborative autonomy for all-domain attritable autonomous systems as work streams. In September 2025, the DoD Inspector General published a Replicator 1.1 evaluation page, but noted that the report contained classified information and no redacted version was available.

The point is not that these documents prove DeLanda right in every detail. The point is that his central object has become an active procurement, standards, oversight, and public-memory problem. The most important facts are often not the dramatic ones. They are the records that show which systems were selected, what they were tested against, who could see the results, and where public accountability stops because classification, vendor secrecy, or operational necessity takes over.

Modern military AI is therefore not one moral question. It is a stack: sensing, data labeling, simulation, autonomy, communications, cybersecurity, human-machine interface design, training, doctrine, procurement, legal review, escalation control, after-action records, contractor responsibility, and political authorization. A weapon can be technically sophisticated and still be institutionally illegible. A human can remain nominally responsible while the system has already made responsibility hard to exercise.

Governance and Safety

DeLanda turns autonomous-weapons governance into a full-loop audit. The safety case cannot cover only the weapon at the final moment of force. It has to cover the sensor regime, target profile, data lineage, model limits, communications layer, user interface, operator training, command authorization, legal review, rules of engagement, deactivation method, after-action record, and feedback path into future models or doctrine.

The first control is bounded use. A military AI system should name its intended mission, geography, duration, target class, data sources, environmental assumptions, civilian-risk assumptions, communications dependencies, and prohibited uses. The ICRC's emphasis on limits by target type, time, geography, scale, situation of use, supervision, and deactivation is useful because it converts "human control" from a slogan into design constraints.

The second control is evidence under stress. Verification and validation have to test degraded communications, jamming, spoofing, stale intelligence, sensor occlusion, adversarial behavior, coalition handoff, civilian pattern changes, model drift, and operator overload. A system that performs in a clean demonstration can still fail as governance if it hides uncertainty, suppresses dissenting signals, or makes interruption feel like breaking tempo.

The third control is accountability across the public-private stack. Military AI increasingly depends on commercial autonomy firms, cloud services, model vendors, data brokers, component suppliers, and software contractors. Procurement records, test artifacts, interface changes, update rights, incident logs, cybersecurity reviews, and vendor exit paths are part of the war machine. A contractor cannot be treated as outside the loop when its model, platform, or interface shapes what commanders see.

NIST's AI Risk Management Framework helps translate this beyond defense. Govern, map, measure, and manage are useful verbs for any institution that automates perception and action. Map the action surface, measure the failure modes, govern the authority chain, and manage the consequences after deployment. The military version is acute because force is involved, but the pattern travels to policing, borders, schools, benefits, workplaces, and infrastructure.

The governance test is simple: when the loop accelerates, does responsibility become easier or harder to exercise? If the answer is harder, the system has not solved uncertainty. It has redistributed uncertainty onto operators, civilians, analysts, commanders, courts, archives, and future victims of the record it creates.

Where the Book Needs Friction

War in the Age of Intelligent Machines is strongest as a systems book and weakest when its machinic language risks making institutions sound like natural forces. DeLanda wants readers to see technical evolution, self-organization, and emergent order. That can be clarifying. It prevents the naive idea that one villain or one device explains the military machine.

But the same language can blur responsibility. Budgets are chosen. Vendors are selected. Standards are written. Targets are authorized. Operators are trained. Legal reviews are scoped. Data is gathered. Interfaces are designed. Secrecy is defended. When a system appears to evolve on its own, the task is to recover the people, offices, incentives, contracts, doctrines, and political choices that made that evolution durable.

The book also needs to be read with more attention to civilians, colonial testing grounds, race, borders, contractors, and the people who become data points inside military and security systems. Its conceptual machinery is powerful, but abstraction can travel too cleanly over bodies on the ground. The most responsible reading uses DeLanda's systems vision while refusing to let the system become an excuse.

It also needs a sharper distinction between military AI in general and autonomous force. Logistics optimization, predictive maintenance, intelligence triage, cyber defense, translation, personnel systems, and command support do not raise identical legal or moral questions. The common pattern is delegated cognition inside hierarchy; the risk changes with the action surface, the affected population, and the reversibility of harm.

What This Changes

The practical value of the book is that it turns military AI from an ethics-of-the-final-trigger problem into a feedback-loop problem. The question is not only who fires. It is who designed the sensor regime, who chose the model, who defined the target class, who set the confidence threshold, who trained the operator, who wrote the doctrine, who logs the decision, who audits the aftermath, and who can stop the system when the loop starts producing its own reality.

That frame travels beyond war. Any institution that automates perception and action has to answer the same questions at its own scale. What has been made legible? What has been omitted? What does the interface make easy? What does it make hard to contest? What behavior will the system produce and then learn from? Who remains accountable when speed, complexity, and delegated cognition become normal?

For safety cases, the book suggests a concrete checklist: identify the representation, the authority it receives, the action it enables, the uncertainty it hides, the dissent channel it preserves, the human skill it requires, the stop condition it obeys, and the record it leaves behind. A system that cannot answer those questions should not be defended with words like autonomy, innovation, or decision advantage.

For public accountability, the hard question is what can be known after the fact. If an operation depends on classified models, proprietary autonomy software, ephemeral sensor feeds, coalition data, and secret rules of engagement, then the archive is part of the battlefield. Democracy cannot govern what it cannot later reconstruct, even in summary form.

DeLanda's book is not a comfortable guide. It is too baroque, too theoretical, and too enamored of machinic vocabulary to serve as a straightforward policy primer. But it sees something many smoother books miss: intelligent machines become politically decisive when they are wired into institutions that already want faster sensing, cleaner command, and less friction from human uncertainty. The governance problem is the loop.

Source Discipline

This review separates DeLanda's philosophical and historical argument from current military-AI governance facts. Internet Archive, PhilPapers, author biographies, journal metadata, Wired, and OSTI support the book and reception context. Current policy and standards claims come from official or primary sources: DoD Directive 3000.09, DoD responsible-AI materials, DIU Replicator announcements, DoD Inspector General records, the State Department political declaration, NATO's revised AI strategy, UNODA's CCW GGE page, the ICRC position on autonomous weapons, and NIST's AI RMF Core.

The interpretive claim is bounded. These sources do not prove that all military AI is unlawful, that every defense AI project is an autonomous weapon, or that machines are becoming conscious, divine, or AGI. They support a narrower governance claim: when sensing, simulation, command, and autonomy are connected, institutions need records, limits, human authority, testing, interruption, and after-action accountability strong enough to govern the loop they have built.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books