AI for Digital Warfare

Niklas Hageback wrote AI for Digital Warfare in the aftermath of a decade in which the vocabulary of conflict had quietly been rewritten. Published by CRC Press in 2022, the book arrives on a shelf already crowded with first attempts to describe what had happened to war when generative models, recommender algorithms and automated influence networks became cheap and ubiquitous. Hageback is not a soldier or an intelligence officer. His career has been spent at the intersection of technology and risk — stints at Deutsche Bank and KPMG, and a long parallel interest in behavioural finance that produced earlier volumes on market psychology and the irrational engines beneath apparently rational systems. That background matters, because he reads digital warfare the way he reads markets: as a contest between actors trying to move the beliefs of a crowd, with asymmetric information and automated tooling as the levers of advantage. The book was written for a reader who wants the mechanics but does not come from computer science — a defence official, a corporate security lead, a journalist on the beat, a policy adviser trying to understand what their analysts keep flagging. It takes the technical ground seriously enough to describe the architectures involved, but refuses to retreat into jargon.

The central argument is that artificial intelligence has already moved the centre of gravity in modern conflict away from kinetic operations toward the manipulation of information, perception and trust. Conventional militaries, in Hageback’s framing, remain organised for a fight that is no longer the decisive one. The battlespace that matters is cognitive — a contest over what populations believe, what they fear, whom they trust, and what they choose to do in response. The weapons of that contest are generative models that synthesise faces and voices, recommender systems that decide what a billion phones display each morning, automated accounts that manufacture apparent consensus, and increasingly capable offensive software that explores networks faster than any human defender can respond. The thesis is not that tanks and missiles have become irrelevant. It is that the state or non-state actor willing to industrialise cognitive and cyber operations can achieve strategic effects — the collapse of a government’s legitimacy, the paralysis of an electorate, the disablement of critical infrastructure — without ever crossing the threshold at which the laws and institutions of conventional war begin to apply. The book treats this as a condition that democracies are badly configured to meet, both institutionally and intellectually.

The structure of the book moves from the technical substrate to the operational pattern and then to the doctrine and governance that frame both. Early chapters introduce the artificial intelligence techniques that underpin the arsenal, with particular attention to generative adversarial networks, large language models, and the recommender architectures that shape attention at scale. Hageback is careful here to explain what these systems actually do, how they are trained, and what their failure modes look like, because the operational chapters that follow depend on the reader being able to tell a credible claim from a marketing one. From that technical base the book moves into deepfakes and synthetic media, then into AI-driven social-media manipulation, then into automated and augmented cyber operations. A pivot chapter steps back to describe the doctrines inside which these capabilities are deployed, dwelling in particular on Russian reflexive control and the Chinese concept of unrestricted warfare, before returning to Western institutional responses. The closing chapters treat cognitive bias — the terrain the whole enterprise exploits — and the legal and governance grey zones in which almost all of this activity sits.

The examples give the book its texture. On synthetic media Hageback walks the reader through the generative-adversarial-network architecture that produces modern deepfakes, pairing a generator that fabricates faces or voices against a discriminator trained to catch them, with both improving by mutual pressure until the output passes human inspection. He surveys the forensic techniques developed in response, from inconsistencies in eye blink and specular reflection to the subtler statistical fingerprints that early detectors learned to flag, and he is honest about the arms-race shape of the problem — each detector trains the next generator. He describes early-stage incidents in which doctored audio was used to impersonate executives in fraudulent wire transfers, and the more consequential concern that similar tooling will reach the political arena, where a well-timed fabricated clip of a head of state can move markets or trigger a cross-border response before verification catches up. He is attentive to the asymmetry: fabrication is cheap and instantaneous while debunking is slow and rarely reaches the same audience.

On social-media manipulation the book maps the industrial pattern. It describes bot farms that use large language models to generate posts indistinguishable from those of a motivated human, the use of machine learning to segment target populations and identify the psychological levers that move each segment, and the astroturfing operations that manufacture the appearance of grassroots consensus around a given narrative. Hageback draws explicit lines back to the Internet Research Agency and the operations attributed to it against the 2016 United States election and the Brexit referendum, and he traces the evolution of these techniques through subsequent elections across Europe. He catalogues the specific manoeuvres that have become doctrine — the amplification of existing grievances rather than the invention of new ones, the use of in-group infiltration to steer rather than argue, the strategic release of hacked material to synchronise with media cycles — and places them in the longer Soviet tradition of active measures, which they extend rather than replace.

The chapter on automated cyber operations is where the book turns technical in a way that most general-audience treatments avoid. Hageback discusses machine-learning-augmented penetration testing, adaptive malware that modifies its behaviour based on the defender it encounters, and the prospect of self-directed offensive agents that reconnoitre, pivot and exfiltrate at machine speed. He is sober about where these capabilities actually are at the time of writing — most are closer to force-multipliers for existing operator skill than to autonomous attackers — but he argues that the trajectory is unambiguous, and that the defensive side has been slower to industrialise its own machine-learning tooling. He discusses specific incidents: the NotPetya attack and its cascading effects on global shipping and logistics, the SolarWinds supply-chain compromise, the Colonial Pipeline ransomware event, and a series of operations attributed to Russian and Chinese services against energy and research targets across Europe and North America. Each is used to illustrate a point about automation, speed, deniability, or the mismatch between the pace of attack and the pace of institutional response.

The doctrinal chapter is where the book’s framing is sharpest. Hageback devotes considerable space to reflexive control — the Soviet-era theory, elaborated inside the Russian general staff, that an adversary can be led to make a decision in one’s favour by shaping the information on which that decision rests. He traces its application through hybrid operations in Georgia, Ukraine and the information environment surrounding both, and he reads contemporary Russian influence operations as an industrialisation of that older theory with the aid of modern tooling. He treats the Chinese concept of unrestricted warfare, as laid out by the two People’s Liberation Army colonels in the 1999 volume of that name, as a parallel and in some respects more encompassing framework: a willingness to use any instrument — financial, legal, informational, technological — against an adversary, and a refusal to accept the Western distinction between war and peace as operationally meaningful. He examines Chinese operations across cyberspace, academic infiltration, standards bodies and critical infrastructure procurement, and reads them as components of a long strategic competition rather than discrete incidents. Western responses receive a more sceptical treatment: NATO’s Strategic Communications Centre of Excellence in Riga, the various national cyber commands, the European efforts at platform regulation. Hageback credits the diagnostic work these institutions have produced but argues that they are institutionally cautious, legally constrained, and slow to industrialise the offensive side of the same techniques — a posture he reads as a liability rather than a virtue, given the pace at which adversaries are proceeding.

The behavioural chapter is in some ways the most distinctive, and it reflects Hageback’s own background. He argues that the effectiveness of AI-driven influence rests not on sophistication of production but on the systematic exploitation of cognitive biases that have been catalogued for decades — confirmation bias, availability heuristics, in-group identification, authority deference, the asymmetric weight of loss aversion. What machine learning adds is the capacity to identify which bias moves which person at which moment, and to serve the stimulus that triggers it at scale. He draws on behavioural finance literature to describe how herd dynamics in markets are mirrored by information cascades in social media, and how both can be nudged by actors who understand the architecture better than their targets do. This chapter is the one that gives the book its interpretive key: the technical capabilities earlier described are only useful because human cognition is predictable, and the predictability is itself the attack surface.

The closing chapters turn to governance, attribution and the legal grey zones. Hageback catalogues the obstacles to attribution in cyber and influence operations — proxies, false flags, the plausible deniability built into decentralised botnets, the use of commercial infrastructure that makes the distinction between state and criminal activity analytically useless. He surveys the thin edifice of international law that nominally applies to these operations, from the Tallinn Manual’s attempts to extend the laws of armed conflict into cyberspace to the ad hoc sanctions regimes under which specific operations have been punished, and he is realistic about how little any of it has deterred. He closes with recommendations that are deliberately institutional rather than technical: the industrialisation of defensive machine-learning tooling, the integration of cognitive-domain operations into mainstream military doctrine, the hardening of democratic information environments without surrendering them to censorship, and a candid willingness among Western states to conduct offensive operations in the same arena. He does not describe this as desirable; he describes it as unavoidable if the asymmetry between adversaries and defenders is to be closed.

Reception of the book has been mixed in the predictable ways. Specialists in cyber operations have found the AI chapters serviceable but not original, while specialists in AI have made the symmetrical complaint about the cyber chapters, and both camps have agreed that the integrative framing is the book’s contribution. The behavioural material has been better received, in part because the author’s grounding is real and in part because it supplies a vocabulary the field had been reaching for. Reviewers in the policy community have treated it as a competent primer, more useful for executives and generalists than for operators. Pushback has come on two fronts. One is the familiar objection to any book that frames conflict in cognitive terms: that it elevates influence operations beyond their actual strategic effect, and that the empirical record of persuasion research does not support the premise that bot networks swing elections. The second is a methodological objection — that attribution of specific operations to specific actors is harder than the narrative needs it to be, and that the book sometimes treats contested attributions as settled. Both objections have merit, and both are partially answered by the book’s own hedging where it hedges.

In the broader literature on AI and conflict, the book sits closer to Peter Singer and Emerson Brooking’s LikeWar than to Ben Buchanan’s The Hacker and the State, but with more technical detail on the AI side than either and less narrative texture than LikeWar’s long-form reporting supplies. It pairs well with Thomas Rid’s Active Measures as a historical companion, with Kai-Fu Lee’s AI Superpowers for the industrial-strategic context, and with Christopher Wylie’s Mindf*ck for a case-level view of the influence operation its chapters describe in general terms. Readers who want original archival research or named case studies of specific intelligence operations will find it thin; readers who want a structured map of how generative models, influence operations, cyber tooling and doctrine fit together will find it useful. It is also valuable as a vocabulary reference — a place where terms that still move between disciplines are defined in one place and in a consistent register.

What is likely to age well is the framing. The specific tools Hageback describes have already evolved, and some of the detector-versus-generator battles he catalogues have moved several rounds since publication, but the diagnosis — that the decisive contest is cognitive, that the tooling that wages it is automated and cheap, that democracies are configured for the wrong fight — has held up. What is already dated is the technical surface: large language models in particular have advanced beyond what the book could anticipate, and the operational doctrine around their deployment is being written faster than any single volume can track. Hageback’s instinct to look at behaviour rather than hardware is what gives the book its longer shelf life. Markets, he has argued throughout his career, are patterns of belief under pressure. He reads this new battlespace as the same thing, with the pressure applied deliberately and at machine speed, and the argument lands.