AI for Defense and Intelligence

Patrick T. Biltgen wrote AI for Defense and Intelligence in 2024 after a career spent inside the defence-analytics industry — the sort of work where machine-learning models are built to sift signals intelligence, fuse satellite feeds, and score targets against a moving picture of the battlespace. He is the co-author of an earlier reference volume on activity-based intelligence and has spent years at contractors serving the National Geospatial-Intelligence Agency and the broader intelligence community. That biography matters, because the book is written in the register of a working engineer rather than an academic or a commentator. It joins a small shelf of practitioner-authored guides to AI for the national-security workforce, and it reads like a briefing deck expanded into paragraphs — dense, unadorned, aimed at readers who have to make decisions about the technology on Monday morning.

The context is a defence and intelligence community that has spent nearly a decade being told AI will change everything, without a shared vocabulary for what “AI” actually names. Programme officers read vendor pitches promising autonomous target recognition. Analysts are handed dashboards whose recommendations they cannot audit. Acquisitions staff are asked to write requirements for capabilities they have never seen tested. The academic literature moves faster than doctrine can catch up. Biltgen positions the book against that gap. He does not try to argue that AI is transformative or dangerous — that debate is elsewhere. He argues, instead, that the community needs fluency with the toolbox before it can meaningfully specify, buy, or deploy anything from it, and that fluency is teachable.

The argument running through the pages is that artificial intelligence is not one technology but a shelf of techniques, each with a shape, a cost, and a characteristic failure mode. Machine learning is not a black box that produces answers; it is a family of methods with measurable performance curves. Neural networks are not oracles; they are function approximators whose behaviour on inputs unlike their training data is often bad and always unpredictable. Biltgen treats these systems the way an old field manual treats a weapon: he explains what the thing does, what conditions it needs to work in, what it breaks on, and how a professional evaluates it. The refusal to romanticise the technology is a stance in itself. In a policy conversation dominated by claims of imminent artificial general intelligence at one pole and existential-risk warnings at the other, a book that says here is what a support-vector machine does and here is why it is unsuitable for this problem is doing quiet but useful work.

The organisation moves from fundamentals to applications. The opening chapters walk through the categories of machine learning that any newcomer to the field encounters: supervised learning on labelled data, unsupervised learning that clusters or reduces dimensions without labels, reinforcement learning in which an agent optimises against rewards in an environment. Each is introduced through problems an intelligence analyst would recognise rather than through the standard academic examples. Classifying vessels in synthetic-aperture radar imagery. Grouping intercepted messages by unstated topic. Learning a route through contested airspace under fuel and time constraints. From there Biltgen moves through neural networks — the perceptron and its descendants, convolutional networks for imagery, recurrent networks and then transformers for sequences, generative models — and treats each as a set of engineering choices rather than a magical ingredient. Natural language processing gets its own extended treatment, since so much of intelligence work is text: named-entity recognition, coreference, translation, sentiment on chat traffic, retrieval-augmented question answering over document corpora. Computer vision follows the same pattern, with attention to the specific difficulties of overhead imagery — small objects, oblique angles, seasonal change, adversarial camouflage.

The second half of the book is the operational half. Biltgen maps the techniques onto the phases of the intelligence cycle — collection, processing, exploitation, dissemination — and onto the warfighting enterprise beyond it. Sensor fusion across multiple intelligence disciplines is a central case, because it is where the promise and the difficulty of machine learning are both most visible: a model that can combine electro-optical, infrared, radar and signals cues into a single track has enormous operational value, and gets almost every hard problem in the discipline at once. Automated target recognition receives careful treatment, including the honest observation that performance on a benchmark dataset is almost never what one gets against a real adversary who has read the same papers. Signals analysis in cluttered electromagnetic environments — the everyday reality of modern conflict, from Ukraine to the Red Sea — makes its appearance, as does the growing body of work on adversarial machine learning: what happens when labels are missing, when training data has been poisoned, when an opponent is actively spoofing the model with crafted inputs.

Woven through the applications half is a strand on the programme-management realities that determine whether any of this technology actually reaches the field. Data pipelines get more attention than the algorithms in some places, because in defence work the pipeline is where most projects die. How data is labelled, by whom, under what quality control. How model performance is measured, on what test set, against what baseline. How a model is retrained as the world drifts underneath it. The chapters on test and evaluation are among the most useful, because they describe a regime — the developmental and operational testing that defence acquisition demands — that has no equivalent in the commercial AI world and that vendors arriving from that world often misunderstand. Biltgen also examines the ecosystem that has grown up to serve the national-security market: how a data-integration platform like Palantir fits alongside a hardware-plus-software prime like Anduril and a European entrant like Helsing, and what the shape of that ecosystem implies about who ends up owning the analytical core of a mission system.

The examples are the concrete heart of the book, and where the value for a reader unfamiliar with defence AI lands hardest. Biltgen returns repeatedly to the difference between a demonstration and a fielded capability. A model that identifies a class of vehicle in a curated evaluation set may collapse when moved to a new sensor, a new geography, a new season. A dashboard that flags anomalies in a stream of transactions is only as good as its analyst’s willingness to trust it under pressure, and that trust is a function of explainability that most modern architectures do not provide. He is careful with the recent history: the reader is walked through the shape of programmes like the Pentagon’s early efforts at automated video analysis of drone feeds, and the acquisitions vehicles the Defense Department has built to move faster than the traditional procurement cycle. Named organisations appear throughout — the Chief Digital and AI Office, the constellation of service AI cells, the combatant-command experimentation shops — not to endorse them but to give the reader something to hold onto. The joint all-domain command-and-control ambition, which promises networked decision-making across services and coalition partners, sits behind much of the operational discussion, since it will succeed or fail on exactly the fusion problems the book describes.

Reception has followed the pattern of practitioner books more than trade press. The military and intelligence education community has picked it up — it has appeared on reading lists inside professional education institutions and acquisition offices — while the academic AI literature has largely passed it by, because it is not making a research claim. Its readers are people who need to work, not people who need to publish. Where discussion has surfaced, the pushback tends to concern breadth rather than depth: a book that tries to cover every branch of machine learning in a single volume necessarily makes each treatment shorter than a specialist would want, and readers coming from a computer-science background will find sections that gloss over subtleties they would prefer explored. Biltgen’s counter, implicit in the design, is that the audience is not the specialist but the operator, the officer, and the programme manager who needs enough grounding to ask sharp questions. On that criterion the coverage is well-judged. The critical reception from the applied-ethics side has been muted, largely because the book does not stake out a strong position on autonomy or lethal decision-making; it prefers to describe the technical shape of the problem and leave the doctrinal question to doctrine.

Where it sits in the field is closer to a reference than to a thesis. There is no shortage of works on the ethics of autonomous weapons — Paul Scharre’s Army of None and Kenneth Payne’s I, Warbot are the reference points there — and no shortage of general-audience introductions to machine learning. What has been missing, and what Biltgen supplies, is the translation layer between the two: enough machine-learning content to make a defence reader literate, and enough defence content to make a technologist understand what the operational context actually demands. It pairs naturally with those books; it does not replace them, and it does not try to. It also complements the more sceptical practitioner literature — the writing of engineers who have watched machine-learning projects fail in the field, and the growing memoir literature from analysts who have spent a career watching promised systems arrive late and underperforming. Someone assembling a reading list on the subject would put Biltgen early, for grounding, and read outward from there.

The book is not a work of foresight and does not pretend to be. It does not attempt to predict where AI in warfare will go, which capabilities will mature by which decade, or how the balance between manned and unmanned systems will shift. Its horizon is short: what does the technology do now, and what should the defence and intelligence workforce know about it now. That focus is what makes it useful and what will date it. The specific techniques it describes will be superseded — some already have been, since the pace of change in large language models alone has been faster than any textbook can track. The vendor landscape will consolidate or fracture. The programmes named in the operational chapters will be renamed, cancelled, or absorbed into others.

What is likely to age well is the underlying disposition — the insistence on treating AI as engineered artefact rather than mythic force, and the discipline of asking what data, what test, what failure mode, what alternative, before what promise. That habit outlasts any particular architecture. The passages on test and evaluation, on the differences between commercial and defence data pipelines, on the human factors that determine whether a model is trusted in the loop — these are describing problems that have not been solved and will not be solved by the next generation of models. In that sense Biltgen has written the book he set out to write: the one that ends up annotated in the desk drawers of acquisition officers, intelligence-systems engineers, and the growing number of officers who have realised that the AI conversation is not going to slow down long enough for them to catch up unless they sit down and read.