We’ll never have a model of an AI major-general: Artificial Intelligence, command decisions, and kitsch visions of war

ABSTRACT

Military AI optimists predict future AI assisting or making command decisions. We instead argue that, at a fundamental level, these predictions are dangerously wrong. The nature of war demands decisions based on abductive logic, whilst machine learning (or ‘narrow AI’) relies on inductive logic. The two forms of logic are not interchangeable, and therefore AI’s limited utility in command – both tactical and strategic – is not something that can be solved by more data or more computing power. Many defence and government leaders are therefore proceeding with a false view of the nature of AI and of war itself.

KEYWORDS:

• Strategy
• tactics
• Artificial Intelligence (AI)
• decision-making
• Clausewitz
• technological determinism
• problem of induction

Introduction

When I have learnt what progress has been made in modern gunnery

When I know more of tactics than a novice in a nunnery

In short, when I’ve a smattering of elemental strategy

You’ll say a better Major-General has never sat a gee

- The Pirates of Penzance, 1879, by Gilbert and Sullivan

Artificial Intelligence (AI) technologies as we know them cannot make, understand, nor explain command decisions in war. Rather than heralding an era of canny and rapid machine commanders or advisors for human leaders, AI technologies may at best amount to something resembling the satirical ‘model of a modern Major-General’ in Gilbert and Sullivan’s Pirates of Penzance comic operetta of 1879: a narrow intelligence that has encyclopaedic empirical knowledge yet is woefully ill-equipped in the relevant skills of command in war. We argue that the nature of war itself prevents ‘narrow’ AI from performing, understanding, advising, and explaining command decisions in strategy and tactics in any reliable fashion. This is because war requires abductive logical reasoning (which AI cannot do) rather than inductive logic (which it can), to comprehend and execute command decisions. We offer a novel, original, and epistemologically rigorous rebuttal to the optimistic claims of AI’s role in decision-making and advising in tactics and strategy in recent AI and strategy research. We use a philosophy of science perspective to show how induction-based AI is unsuited to war, and reinforce traditional Clausewitzian arguments against mechanical approaches to the conduct of tactics and strategy.

In his recent work, Lawrence Freedman explores command from many differing yet complementary interpretations, showing that ‘it is about much more than handing out orders and ensuring that they are enacted.’¹ Command refers to an ability to inspire subordinates and succeed with intuitive judgment as much as making the ‘correct’ decisions in directing small military units or entire bureaucracies in wartime, managing resources effectively, and connecting tactics with strategy and vice versa.² When it comes to defining command in Clausewitz’s theory of war (summarised throughout Book I Chapter 1 of On War), the differences in moral and human elements of tactics and strategy are matters of degree, and not absolute. Clausewitz argued that tactics and strategy are borne of the same phenomenon: organising for combat in one way or another, with tactics seeking a victorious battle and strategy being about the art of tying battles together, as well as all the other logistical and supporting factors, to meet the ultimate political object of the war.³ Command cannot be objectively, precisely, and consistently defined, as Clausewitz argued about his entire ontology. Instead, we refer to the central points of such a collection or ideas that correspond to ‘command’.⁴ Command decisions in tactics and strategy – and rounded advice in making those decisions – require multiple kinds of logical inference, and the good judgement to know when to use each one. Command decisions at their heart require judgement, which is something AI technologies cannot do – machine learning, ‘narrow’ AIs can only calculate.

Practicing strategy and tactics entails the same kind of command decisions with regard to abductive logic and the human, unpredictable elements of war. How far AI might be from being competent in this area depends heavily on one’s conceptualisation of strategy. One narrow, and insufficient, definition of strategic decision-making is the ‘means-ends rationality’ paradigm that was dominant among Cold War-era U.S. elites.⁵ This conceptualisation of strategic decision-making is overly materialist and ‘implied that an action from one’s adversary produces particular knowable calculable consequences that we can then respond to with known (and knowable) solutions.’⁶ As Yee-Kuan Heng observes, such a closed and finite model of strategy conveniently gave hope to dreams in the mid-20^th century of computer simulations that could predict outcomes based on inputs of ‘events and triggers.’⁷ As a form of decision-making that is knowable from end to end, ready-made for digital input, this seems to be the conceptualisation of decision-making that is adopted (often tacitly) by the optimistic accounts of AI technologies in warfare.⁸ The means-ends rationality account of decision-making comes with the additional baggage of a mistakenly battle-centric approach to strategy. Andreas Herberg-Rothe argues that battle-centrism interprets ‘the shortest way of connecting purpose and means’ to be battle and combat, a position taken but later rejected ‘fervently’ by Clausewitz (before writing On War) because it conceives of war as somehow independent of the political sphere.⁹ ‘Victory only has tactical significance … at the strategic and political level [where] both ends and means have to be considered’ Raymond Aron summarised of Clausewitz.¹⁰ That said, tactical results produce victories that then ideally deliver the ends of strategy – therefore AI is not unimportant if it can offer useful contributions to winning battles. Judging which tactical actions will deliver aggregate strategic results are best thought of as command decisions within operational art (noting that the operational level of war is not a useful analytical concept).¹¹ The degree of military AI’s usefulness will be highly restricted if it cannot engage with the concept and practices of command at tactical and strategic levels specifically, nor in the operational art that is the ‘planning, preparation, synchronization, and sustainment of tactics over sustained period of time or a large geographic expanse’ towards strategic goals.¹²

The nature of war is an insurmountable barrier to realising the optimistic predictions of narrow AI. It will remain a chimera for defence planners that cannot be resolved by acquiring more computing power, resulting in wasted resources at best or disaster in the field at worst. While we do not dispute that ‘narrow’ AI is capable of making decisions, existing literature on military AI has not done enough to distinguish between different kinds of decisions – assuming instead that command decisions in the tactical and strategic levels of war are similar enough to deciding on moves in a game with tightly governed and quantified rules. Criticism of military AI decision-making capability and roles have focused on technical limitations¹³ or the ethical failings of killer machines.¹⁴ Instead, we show that no technical advances or side-stepping or resolution of ethical issues will produce a competent AI commander that can reliably lead or advise commanders in tactics and strategy. This is because the AI that exists and dominates contemporary AI and strategy literature (‘narrow’, machine learning, or pattern-recognition AI) relies on inductive logic (deciding based on predictions drawn from prior observation) whilst command in both tactics and strategy requires abductive logic (deciding in the face of the unknown and unknowable). Abductive reasoning is fundamental to strategic theory and practice, yet it is missing in accounts of the potential of AI in warfare. Existing claims of competent AI advisors for command are only persuasive if one adopts a gamified, or ‘kitsch’ vision of war that the optimistic accounts of military AI tend to adopt. Mischaracterisations of war are compounded by the dubious empirical foundations that AI optimist accounts rely on to make their claims about narrow AI’s suitability for command duties. That narrow AI can play games like Chess and Go effectively, or fly a simulated aircraft, does not mean that narrow AI can be relied upon to perform command duties in war. That distinction between playing games and conducting war stems from the ability to win a game through inductive logic without the need for abductive logic, whereas the nature of war in the Clausewitzian sense is rooted in the need for abductive logic for effective command decisions. In other words, AI systems that can play games are using a different logic to the kind needed by commanders for tactics and strategy, and therefore the evidential basis for arguing AI can make reliable command decisions is deeply flawed. Narrow AI indeed has a role in tightly bounded pattern-recognition tasks, such as imagery filtering and analysis, and automated administrative tasks, but command in war requires different logics that AI have not been engineered to do, even only as ‘advisors’ to human decision makers, because advisors need to comprehend abductive logic in order to make qualified recommendations on command decisions.

Attempts to persuade others that AI could make or advise on command decisions rely on a kitsch vision of war. By kitsch, we mean visions that ‘purge … deeper and more profound (and contrary) expressions of the meaning’ of the original phenomenon.¹⁵ In other words, to view war as a game with clearly understood, knowable, quantified, and unbreakable rules and moves is kitsch. Kitsch tendencies appear in policy documents and over-emphasise the potential of AI in command decision-making and advising. The US Department of Defense’s (DoD) AI strategy states that ‘AI can generate and help commanders explore new options so that they can select courses of action that best achieve mission outcomes.’¹⁶ A variety of NATO-sponsored reports have taken the future strategic impact of AI decision-making as already settled.¹⁷ A former UK Prime Minister hailed AI as a technology ‘that will revolutionise warfare’, and his Defence Secretary conjured an image of ‘a soldier on the front line … guided by portable command and control devices analysing and recommending different courses of action.’¹⁸ Indeed, ‘veterans’ of the Revolution in Military Affairs (RMA) debates of the 1990s and 2000s, on the question of fog, friction, uncertainty, and human agency being eliminated through technology, will see parallels in the debate over the significance (or not) of AI in warfare, as noted by Goldfarb and Lindsay.¹⁹ If AI will never be able to produce reliably good command decisions in war, even only as recommendations to human commanders, then militaries have no reason to try implement them and the significant technical and ethical drawbacks of AI can be avoided.²⁰ The UK Defence Artificial Intelligence Strategy of 2022 claims that ‘Machines are good at doing things right (e.g. quickly processing large data sets). People are good at doing the right things (e.g. evaluating complex, incomplete, rapidly changing information guided by values … ’.²¹ Whilst the detail of the point is correct – that existing, machine-learning-based AI can process large datasets quickly – it is a nonsense to assume that AI can ‘do things right’ when they instead mean ‘quickly’. Even where narrow AI is useful, it is fraught with dangers that cannot be divorced from the nature of war. Rather than obfuscate these realities with flights of fancy into imaginary, general AI superintelligence, government planners need to build AI policy in relation to the unforgiving, human, and unpredictable nature of war. Optimist accounts of AI’s command potential do not survive contact with Clausewitzian conceptions of war.

To make this case, the article proceeds in three sections. Section 1 reviews optimistic arguments regarding the capabilities and potential of narrow AI technologies in command roles. Section 2 reviews the evidence base for AI optimism, which is heavily reliant on AIs that play games, and discusses the limitations of such evidence because games are bounded and finite ‘kitsch’ models of war and do not reflect fundamental aspects of Clausewitzian ‘real’ war. Section 3 explains this in more depth with a grounding in the classical problem of induction in the philosophy of science, showing how narrow AI cannot overcome it and cannot perform abductive logic (which is essential to command decisions).

AI optimism

AI optimism concludes that military AI will be competent beyond basic automation seen in administrative management and pattern recognition in large datasets. Simplistic distinctions between different kinds of decision making – perhaps better thought of more specifically as cognition – underpin AI optimists’ claims that demonstrations of AI fighter pilots and Chess players are valid indicators for heralding the appearance of competent artificial commanders in war. These claims are in turn based on overly rigid distinctions between levels of war that were intended by classic military theories as analytical and practical, rather than ontological, differences. Optimistic assessments of military AI capabilities are based on lowering the bar of what constitutes strategic and tactical skill and command decisions. What varies is how explicit these accounts are about their assumptions. Kenneth Payne exemplifies this when he states ‘that AI alters the nature of war by introducing non-human decision-making.’²² Payne argues that ‘centaur strategy’ – teaming humans with AI – is one solution to AI’s difficulties with creativity and common sense.²³ The AI commander in this instance would be improving the quality of human decision-making with potentially novel insights and suggestions²⁴ - effectively the role of a staff officer. Not all of those arguing for AI commanders and advisors are as transparent about their theoretical assumptions as Payne. However, the position is emblematic of mischaracterisations of AI and the nature of war.

There are internal contradictions within specific AI optimist accounts, making general summaries of arguments difficult. This is a symptom of a thin conceptualisation of war. We have attempted to argue against the positions of AI optimists, who are of course not a monolith and do contain nuances. However, within the same publications, there are inconsistencies that make this task very difficult. The conceptual muddiness and incoherence over specific AI technologies and the nature of war²⁵ and strategy leads to arguments such as ‘AI will tend to amplify the importance of human leadership and the moral aspects of war’.²⁶ This argument is unsatisfying, if not banal, given that these are essential if not foundational aspects of the Clausewitzian nature of war and a kernel of multiple strategic theorists. Arguments pointing both ways where AI is claimed to be transformative yet complementary,²⁷ or revolutionary and evolutionary at the same time do not help practitioners understand what AI is, what it does, and importantly what it cannot do regarding the conduct of war and the execution of strategy. The arguments suggesting something is transformative when it may be merely another complementary tool for military forces and bureaucracies that does not completely change war amounts to a ‘having your cake and eating it’ argument. Goldfarb and Lindsay argue that ‘AI is potentially transformative, we disagree about what that transformation might be.’²⁸ If something is argued to be transformative, surely the character of that transformation should be identified to make the case. These examples are why we have chosen to focus on the shared evidence base that underpins AI optimism. The discourse of optimism about military AI is uneven – but their arguments about AI’s role in tactics and strategy share the same empirical foundations and a ‘kitsch’ vision of war.

Military AI optimists do not distinguish sufficiently between different types of inferences and decisions. Yet, the type of decisions that the various paradigms of AI can make are at the centre of what roles they can competently take on. Technical cognition – such as keeping an aircraft on a flightpath – falls short of human-like consciousness that should nevertheless be considered a sophisticated kind of inference and decision-making.²⁹ The pressing question that AI optimists are reluctant to answer (or even acknowledge) is whether the unconscious cognition of the current paradigm of narrow AI can produce valid inferences and decisions in strategy and tactics. Whilst operating a weapon system lends itself to mechanical automation, the use of ever-more automated weapons systems is still a tactical affair requiring human decision-making using abductive logic engaging with the unknown and unknowable of the theatre of war, rather than induction based only on what has been observed in a quantified historical database. ‘Deep learning’ is the major development behind contemporary ‘narrow’ AIs, as opposed to 20^th century attempts.³⁰ Rather than making inferences solely based on rules programmed into the model, narrow AI today usually draws on deep learning to ‘discove[r] patterns and trends in observed data.’³¹ Deep learning-based narrow-AI is therefore statistical and probabilistic. As demonstrated throughout this article, this approach is rooted specifically in inductive logic that is inherently flawed for the purposes of command decisions in tactics and strategy. An Aegis Combat System prioritising targets and allocating, then firing interceptors in the immediate moment is different to deciding whether to engage or not in the first place, or hold ammunition in reserve based on other contextual factors beyond what can be observed in that one missile attack. Indeed, this is why the Aegis ‘Command and Decision’ computer operates under pre-delegated ‘doctrine’ written by the ship’s review board.³² Even this advanced automated decision making differs from a flotilla commander deciding to pursue an inferior force, stand and fight, or flee and give the slip to a superior force. As Julian Corbett argued, some of the most taxing decisions for naval commanders were not winning battles through superior gunnery and manoeuvring in themselves, but rather how to bring favourable battles about in the first place.³³

Optimistic accounts of AI in war are not waiting for ‘true’ machine intelligence or general AI – still a distant possibility – to emerge before pushing to assign it responsibility over command decisions and advisory roles on command.³⁴ The U.S. Defense Advanced Research Projects Agency’s (DARPA) Knowledge-directed Artificial Intelligence Reasoning Over Schemas (KAIROS) program attempts to create an AI that could make command decisions.³⁵ It is a narrow AI that is officially pitched as being able to assist humans with strategic decision-making, sifting through vast amounts of data on world events ‘in order to generate actionable understanding of these events and predict how they will unfold’.³⁶ The supposed promising outlook of KAIROS has been used as evidence by AI optimists that AI is already competent.³⁷ KAIROS can only make decisions based on probabilistic inferences, however. More impactful than KAIROS, the US military is pushing ahead with using AI to ‘make sense’ of conflict under the Joint All Domain Command and Control (JADC2) project.³⁸ Proponents of systems like KAIROS and JADC2 tacitly make the argument that command decisions can be based on probabilistic inference – inductive logic – alone. Consequently, while these decisions may be labelled as ‘strategic’ or ‘command’ decisions by AI optimists, we should be sceptical and scrutinise these claims further. ‘Tactical’ level examples of AI competence are the evidence supporting optimistic claims about military AI. There is a problematic distinction in believing that tactics are inherently more automated types of command decisions than strategy, such as when Payne concedes that strategic decision-making is ‘less tightly bounded’ and ‘less mechanistic’ than determining how to respond in air-to-air combat, for example.³⁹ DARPA’s Alpha Dogfight programme saw a human pilot defeated by an AI model five times in a row in a simulated air-to-air duel – and immediately became a touchstone piece of evidence for military AI optimists.⁴⁰ Similar, though less complex, examples of AI technical success at the ‘Tactical’ level are supporting target selection, or allegedly in one example in the Libyan Civil War, autonomous target selection and use of lethal force.⁴¹

A narrow AI presenting filtered or enhanced information is different to making or advising on decisions. The choice of this evidence demonstrates that AI optimists assume that the lion’s share of what makes decision-making in war are the practices around targeting and information processing. This is a convenient theoretical position to take to bolster their claims, because narrow AI has proven competent at targeting in its simplest sense.⁴² An AI ‘Top Gun’ certainly makes decisions in order to win, but there is little about deciding how to fly a plane against one opponent that relates to the trials of tactical or strategic command – the scenario means the AI does not choose whether to engage, for example. The result is that AI optimists are mistaking technical AI success at operating machinery at the ‘tactical level’ and conflating it with evidence that AI is ‘doing tactics’ or conducting tactical command. Narrow technical engagements between weapons platforms are hardly the entirety of tactics and command in battles, let alone strategy. Yet this is what AI companies appear to be selling to governments. With a caveat in the small print that its claims are only ‘forward-looking statements’ that ‘should not be read as a guarantee of future performance’, Palantir’s Artificial Intelligence Platform (AIP) for Defense claims to be able to suggest courses of action for tactical commanders.⁴³ Still, the courses of action it suggests are banal – merely various targeting options, not tactics. Tactics and strategy share much of the same qualities in terms of passion, reason, and chance and neither can be reduced to purely their technical and quantifiable elements.⁴⁴ AI advisors like Palantir’s AIP cross the line from merely presenting information to providing ersatz command judgements.

AI success at isolated tactical interactions is taken to be evidence of greater things to come because of how AI optimists imagine tactical successes contributing to strategic goals. Here, the questionable intellectual legacies of John Boyd, John Warden, and Arthur Cebrowski hang over the arguments of AI optimists.⁴⁵ Notions of Observe, Orient, Decide, Act (OODA) loops, hierarchies of centres of gravity, and the war-winning effects of precision strike complexes are the intellectual primordial soup of AI optimists’ visions of war and strategy. The AI optimist literature not only resembles the problems of the RMA debates and OODA loop concepts, but also what Michael Handel called the ‘tacticisation of strategy’⁴⁶, where lower-level operational considerations start driving strategic objectives, where the military means excessively influence the political ends of war.⁴⁷ One of the expected ‘benefits’ of AI systems in war stems from taking decisions more quickly than humans.⁴⁸ Johnson argues that ‘at its current development stage, AI in isolation has few genuinely strategic effects;’⁴⁹ yet also argues that

militaries that use AI will doubtless gain significant advantages on the battlefield (e.g. remote-sensing, situational-awareness, battlefield manoeuvre, and a compressed decision-making loop), compared to those who depend on human judgment alone.⁵⁰

Johnson here explicitly involves AI in decision-making structures, not only at the tactical level but also effectively at the strategic level, by taking a targeting-oriented vision of strategy and tactics. Even if only used as advisors, if human commanders come to depend on AI prompts and options, AI are in effect making command decisions as part of a collective command bureaucracy. AI optimists’ insistence that war is mostly a matter of optimising decision speed and quantities of data is a central plank of their kitsch vision of war.

AI optimists envisage the existence of objectively ‘correct’ decisions in a given situation. Ayoub and Payne imply this ontology when discussing how one of AI’s supposed benefits is escaping some sources of ‘bias,’ a point that implies that there is an unbiased truth to be attained.⁵¹ In fairness, Goldfarb and Lindsay argue that humans are needed to identify when data is incomplete or biased, requiring ‘close supervision’ of AI decision suggestions, but again it implies that there can be such a thing as unbiased data, or ‘quality data’ as they call it. This replaces the efficiency gains of AI imagery and targeting analysis with the inefficiencies of needing more people to assess the AI’s own command suggestions.⁵² Rather than requiring humans to prevent ‘biased data’ corrupting an AI’s work, Francis Hoffman more problematically pushes in the other direction, arguing that AI will reduce human biases.⁵³ Keith Dear argues that ‘before one can integrate AI into decision-making, one is going to need far more rigour in one’s decision-making at all levels. Before one can have explainable AI, one needs explainable humans.’ ⁵⁴ This is an example of an explicit call to change existing human practices to allow the integration of AI.

Optimism around military AI is heavily reliant on machine success in games and simulations as supporting evidence. AI optimists’ promises of ‘unprecedented insight,’⁵⁵ a ‘prediction revolution’,⁵⁶ ‘an entirely new approach to military affairs’,⁵⁷ and that ‘human ability might be easily superseded’⁵⁸ tend to use AI success in games as evidence.⁵⁹ By claiming that narrow AI’s proven abilities to win games means that AI will help win wars, AI optimists commit themselves to claiming that these games and war share key ontological similarities. For example, Keith Dear argues that ‘Google’s Deepmind’s remarkable effort in developing an AI that could beat the world’s leading player of the Chinese game Go is suggestive of where decision-making in defence and security is headed in the long term.’⁶⁰ Ayoub and Payne over-extend Clausewitz’ brief, pedagogical comparison between war and a game of cards, using it to justify why recent AI victories against humans in Chess, Go, and video games are evidence of AI’s applicability to strategic decision-making.⁶¹ Francis Hoffman similarly comments that AI defeated humans in the ‘intellectual’ pursuits of playing such classic board games.⁶² Where some optimists have accepted that war and games have ‘profound’ differences,⁶³ it is important to identify what AI optimists think these differences are. These optimistic accounts do not go far enough to distinguish between games and real war, lacking scrutiny of their evidence for making such claims for AI’s command potential.

The Clausewitz Engine

When AI bested humans at Chess or Go, it did so with superior calculation capabilities. The famous 1997 success of Deep Blue against Gary Kasparov was heavily dependent on computational speed. In 1977, a computer defeated a grandmaster for the first time in a special ‘blitz game’ that lasted only five minutes in its entirety – the shorter game gave the machine an additional advantage.⁶⁴ Deep Blue belongs more to the last generation of ‘AI’ – sometimes called ‘expert systems’ – that precedes the machine learning-based systems of the 21^st century (though both are kinds of ‘narrow’ AI). In effect, the machine was programmed with the rules of chess and then ‘searched 2 to 2.5 million chess positions per second per chip,’ looking multiple moves ahead in a branching fashion.⁶⁵ Targets for attack had preassigned values that were used to decide between options, and do cost-benefit analysis in anticipation of retaliation from the opponent.⁶⁶ AlphaGo, the first machine victor against a master human player of Go, is significantly more complex than Deep Blue. Where Deep Blue simply used brute computational power, this would have been insufficient to win against an excellent human player of Go.⁶⁷ Rather than being programmed with knowledge from books, AlphaGo used deep learning to generate principles to follow by playing against itself.⁶⁸ However, it did not dynamically use deep learning during a game. Instead, AlphaGo used the Monte Carlo tree search method. This meant that rather than attempt to calculate every possible move, it randomly sampled from a ‘tree’ of possible games and used probabilities to decide which move to take next.⁶⁹ This kind of decision-making is still computationally intensive, and far beyond the power of an unassisted human mind. In other words, AlphaGo and humans use different methods to play Go, and the computational power of AlphaGo seems to enable the machine to utilise a superior process to achieve victory in the game. The differences in approach – and the brittleness of AlphaGo’s technique – was later demonstrated by human victories over similar Go-playing AIs, KataGo and Leela Zero.⁷⁰ The human player benefitted from a pre-game computer analysis of the adversary machine learning, identifying a fatal flaw derived from an induction-based AI to ‘understand’ the board.⁷¹ As the AI players were surrounded, they showed little sign of responding and adapting to their predicament and were defeated in short order.

A kitsch vision of war emerges if we unpick the claims of the AI optimists about machine success in games. Decision-making in war under this implied vision is within a closed, rule-based system. Dear explicitly responds to criticism that AI success in games was only because they are ‘deterministic’ by arguing that ‘our mental models of the world are limited, heavily simplified facsimiles of the real world too.’⁷² It is a de facto argument that theories of victory for real wars are therefore directly comparable to definitions of victory in Chess and Go. This is despite Clausewitz’s explicit tutelage that war in the real world is different to that imagined and scripted by humans on paper (or silicon). Theory can only go so far in educating people and simulating war as seen below. Command decision in war, according to AI optimists, are about quickly and efficiently moving towards a discrete threshold to victory, divining a way forward by comparing alternative paths or targets to that threshold. If AI is to be a useful assistant to future commanders, the theory of victory and the competing alternatives, probabilities and targets must be digitizable and quantifiable. In other words, the best action to take next can be mathematically calculated with patterns of past action in a closed rule-based system. They conclude, then, that because AI is very good at calculation it will soon be very good at fighting wars as well. Instead, we argue that using games as evidence of the suitability of AI commanders inherently means adopting a false, kitsch vision of war because it does not account for the immaterial, unquantifiable, and unknowable aspects of command decision making in wars that cannot be programmed or observed in datasets.

Conceiving of war as a kind of game or closed system allows AI optimists to envisage a future in which AI will be able to make or advise on command decisions. Viewing wars as mere games reduces strategy and politics to a version of war that is battle-centric, (if not entirely centred on killing), a-geographical, a-historical, frictionless, quantifiable, repeatable, non-paradoxical and computable.⁷³ Clausewitz warned that war in the ‘real world’ made up of humans and chaos cannot be reduced to ‘a sort of algebra of action’. If all variables and outcomes could be known, and if war was a purely rational affair, there would be ‘no need of the physical existence of armies, but only of the theoretical relations between them’.⁷⁴ Accidents and good luck play a great part in every level of war and it is therefore perplexing that Chess- and Go-playing AIs are championed as harbingers of ‘warbots’ to come rather than the AIs designed to play war, strategy, and political games that intentionally, but imperfectly, simulate passion, reason, and chance in conflict.⁷⁵ It is no coincidence that the Swedish video game developers Paradox Interactive named the software engine that runs their grand strategy titles such as Hearts of Iron IV and Victoria III the ‘Clausewitz Engine’. While fun, they are just that: games that simulate political and social abstractions that have to be distorted into mechanics for a computer to manage and compute so as to produce entertainment. At a very high level of abstraction, war and games both have victory as a goal. However, paths to victory in games are restricted by finite and knowable rules and measures.

Victory is a deceptively simple goal in war in both tactics and strategy. Both the objective as well as the means of achieving it are as variable as each other and are impacted by one another as war is still politics by other means.⁷⁶ There are no universal rules as to what is feasible and possible beyond material reality and the flexible constructed ‘rules’ of socio-political constraints, whether forced or voluntary. The Japanese military theorist Miyamoto Musashi wrote that when ‘it comes to winning victory … the heart of the matter is to use the power of the knowledge of martial arts to gain victory any way you can.’⁷⁷ No constructed ‘game’ can match this creative space for strategic and tactical planning and command as real war and politics can. This difficulty of command in the real world is a result of the political, uncertain and chaotic nature of war. In a passage describing the ability of friction to frustrate all efforts in war through things going wrong or being harder than anticipated, Clausewitz pre-emptively critiqued the AI literature’s scientific approaches almost 200 years before modern technologists committed them to paper:

As long as we have no personal knowledge of war we cannot conceive where the difficulties of the matter lie, nor what genius and the extraordinary mental and moral qualities required in a general really have to do. Everything seems so simple, all the kinds of knowledge required seem so plain, all the combinations so insignificant, that in comparison with them the simplest problem in higher mathematics impresses us with a certain scientific dignity. But if we have seen war, all becomes intelligible. Yet it is extremely difficult to describe what brings about this change and to put a name to this invisible and universally operative factor. Everything is very simple in war, but the simplest thing is difficult. These difficulties accumulate and produce a friction of which no one can form a correct idea who has not seen war.⁷⁸

Clausewitz’s solution on how to learn about war from books is to theorise the role of passion, reason, and chance in war – not to gamify them via quantification or become entirely subservient to observed precedents in datasets (or military history).⁷⁹ Indeed, Clausewitz’s constant deployment of counterfactuals in dissecting command decisions shows the crucial value of abductive logic and the importance of the un-observed phenomenon in learning from military experience. This is the challenge of preparing for war without actually doing it, with general friction being a crucial aspect of the distinction between war in reality and war in a theory, simulation, or game. Machine learning today requires repeatedly training AI in a simulated environment which can never replicate the real thing, and with inductive logic alone an AI can never consider anything conceptually ‘outside’ of its datasets. The Clausewitz Engine of the video gaming world cannot supplant the ‘Clausewitz Engine’ of classical theories of war and the canonical foundations of Strategic Studies and global military history. Though useful for military and political education, theory is never a full replacement for experience and practice in the real world.

Unlike war, few if any games have an infinite number of ways to win because games have a set number of rules and mechanics, which lend themselves to machine learning’s inductive logic. Checkers was formally, mathematically solved in 2007, with a total number of around 5 x 10²⁰ possible legal moves,⁸⁰ while Chess has 10¹²⁰ and Go around 10¹⁷⁰.⁸¹ War is not a bounded system with a finite number of moves, and as a result the military AI debate should move to debating the notion of decidability. AI optimist accounts in Strategic Studies are yet to engage with the concept of decidability – an otherwise openly debated notion in technical AI communities.⁸² This goes beyond simply incomplete information, although this is a part of the phenomenon. ‘Decidability’ is a logical test that determines whether brute force calculation can ‘solve’ a given phenomenon. Indeed, a form of Kriegsspiel was one game assessed for potential undecidability.⁸³ The collectable card game Magic: The Gathering may well be the hurdle that AI fails to cross, a game more mathematically complex than Go but undoubtedly less complex than the practice of war. In 2019, a team of computer scientists proved that there is at least one real-world game of Magic ‘for which determining the winning strategy is non-computable.’⁸⁴ While the authors of the paper predict that an AI could be written to beat a human at Magic most of the time,⁸⁵ this should not give hope to the AI optimists hoping for evidence of a useful AI commander. Instead, the nature of Magic’s complexity, in a mathematical sense, that most impacts applications for AI in war, is that there are already known games where ‘it’s impossible in the general case for an algorithm to look at a board state and see whether it’s possible for the game to end at all.’ As a more complex phenomenon than Magic: The Gathering, an AI commander in war will encounter undecidable calculations and collapse into a never-ending analytical loop, or analysis paralysis, no matter the computing power allocated. Or, the AI commander could churn out arbitrary suggestions for command in the same way Paul the Octopus ‘predicted’ the winners of football World Cup games by arbitrarily choosing one of two boxes containing food identified with the relevant national flags.⁸⁶ Paul the Octopus indeed came to a decision and made some correct predictions, but Paul was not making a judgement about football.

This is the spectre of the limits of inductive logic in another form: war is non-computable by its nature and requires a sort of logic that ‘expert system’ AIs like Deep Blue or pattern-recognising AIs like Palantir’s AIP for Defense cannot perform. The key selling point of AI commanders or advisors – rapid computation – is therefore moot because war is logically ‘undecidable’ and cannot be resolved by computing power and datasets alone. The notion that there are objectively ‘correct’ choices in strategy that are enacted on a battlefield looks particularly kitsch in light of a historical record filled with examples of unexpected or non-battle centric routes to defeat and victory. An exchange between historian and Vietnam war veteran, Colonel Harry Summers, and a former North Vietnamese Army officer illustrates a classic strategic failure:

Summers:

‘You never defeated us in the field.’

NVA Officer:

‘That is true. It is also irrelevant.’⁸⁷

Even after a decisive military result, the result of war is not absolute, especially in political terms. Peace may just be a transitory evil until the next opportunity to use political violence appears at a later time.⁸⁸ These points demonstrate why AI capabilities in speedy targeting and decision-making mean very little beyond very narrow use-cases in operating specific digital information-intensive machines, weapons systems, sensors, or software. What actions need to be changed now to prevent a greater, bloodier resumption of hostilities later? These are the thorny questions that are borne of the simplest, most quoted, yet least dwelled-upon utterance from Clausewitz: that war is merely the continuation of politics by other means.⁸⁹ As war is that continuation of politics by other means and the instrument of policy, the means of war can never be separated from the ends of policy.⁹⁰ An AI taught that rapidly destroying enemy forces and occupying urban areas should lead to victory may leave it as unprepared for continued resistance as Napoleon was in Russia in 1813. The routes to victory are infinite and often unclear, and Clausewitz did not engage in listing them as ‘it would be pedantry to attempt to classify them.’⁹¹ An AI relying on inductive logic or highly prescriptive programming would have to be quite the pedant and could only rely on such an exhaustive list of precedents and knowable conditions that could never be complete, because war is fraught with the unknown and unforeseeable.

The limitations of game-winning AIs show that an AI commander cannot use brute force computation or elaborate statistical methods to find the ‘correct’ command decision. There is no prospect that AI technologies that exist will develop the best logical underpinnings of human judgement – abductive inferences – without falling back on the wishful thinking of the ‘inevitability thesis’ that someday someone will invent a kind of AI that can do it.

Induction, abduction, and genius

Whether AI optimism uses evidence of success in existing military programs or gaming, the technology in question is induction-based.⁹² Because of the problem of induction, the entire evidence base that supports an optimistic assessment of military AI for command decisions should be considered limited at best, invalid at worst. Even where AI optimists are more moderate, they still point to the existing record of narrow AI’s successes as evidence that it will inevitably continue to improve – what Erik Larson has called the ‘inevitability thesis.’⁹³ This is the specific technological determinist argument made by AI optimists, including those beyond the national security field, to the effect that successes in narrow AI prove that superior, or even a ‘general AI’, is an inevitable and even imminent development. The narrative postulates leaping smoothly from proven narrow AI technologies to the promise of systems that can better mimic human thought. If KAIROS, Alpha Dogfight, or AIP for Defense are not up to the task today, goes their reasoning, then it is not long anyway until ‘decision-makers are going to have to accept strategic and tactical recommendations [from AI] that they cannot understand.’⁹⁴ Understanding the insurmountable challenge of the problem of induction shows the invalidity of the game-based evidence for arguing so, as the AI optimists do.

The reason we can be sure that faith in the maturation of narrow AI into something more capable is misplaced is the problem of induction. As it is an epistemological problem, and because war is not logically ‘decidable’, it cannot be ameliorated simply by gathering more data, more powerful computers, or turning more things into quantifiable metrics.⁹⁵ At its simplest, induction ‘means acquiring knowledge from experience,’ and specifically turning those experiences (that we might also term observations) into generalised claims about how the world works.⁹⁶ These generalisations are derived probabilistically. For example, if we see an explosion every time gunpowder is exposed to fire, then it is a good generalisation to assume the same thing will occur next time gunpowder is burned.⁹⁷ Yet we know that is not categorically true. The originator of the ‘problem of induction’ within Western philosophy is David Hume. Hume claims that ‘not only our reason fails us in the discovery of the ultimate connexion of causes and effects, but even after experience has inform’d us of their constant conjunction,’tis impossible for us to satisfy ourselves by our reason, why we shou’d extend that experience beyond those particular instances.’⁹⁸ The popular ‘Black Swan’ analogy⁹⁹ is a version of the problem of induction, but one that must be caveated because it misleadingly implies that errors in AI inductive inference will be due to supposedly rare gaps in data – implying induction ‘rarely’ fails. What Hume is driving at is that all inductive inferences are suspect because a set of observations that appeared at one time to indicate correlation – or even cause and effect – are not a ‘chain of reasoning’ at all, but merely an assumption that the observations are alike. Fatally, this initial assumption is not supported by data or reasoning at all. The conclusion is therefore never certain, it always remains a qualified claim that the generalisation based on observation is probably correct. ‘Black Swan’ events are catastrophic examples of proof that induction cannot predict everything or produce infallible generalisations. The larger problem is that inductive inferences can never lead to deeper knowledge about the phenomena we observe. The subsequent socio-political problem is that induction is used to address problems that it is inherently unsuited for, but nevertheless is misleadingly convincing at first glance. Command in strategy and tactics requires abductive logic – an ability to think and make decisions based on the constant presence of unknowns and unknowable things that may never appear in a historical dataset or past experience.

The Stanford Encyclopaedia of Philosophy describes Hume’s argument as ‘one of the most famous in philosophy,’ yet there is little evidence of an understanding of it within the AI optimist literature despite discussions of the problem of induction in machine learning literature.¹⁰⁰ Where induction is discussed, the definitions can be muddled or incorrect. ¹⁰¹ For example, Andrew Ilachinski (an expert computer modeller at the Center of Naval Analysis) claims IBM’s Watson used abductive – not inductive – inferences to defeat humans at the gameshow, Jeopardy.¹⁰² However, Watson actually used probabilistic reasoning to choose between possible answers – inductive inferences.¹⁰³ Muddled definitions contribute to an inflated sense of AI’s abilities, both now and in the future. Guglielmo Tamburrini argues that philosophers and AI researchers mean different things when using the term ‘induction’, and that AI researchers are ‘often less demanding’ because they are interested in finding procedures that ‘find parsimonious hypotheses that are consistent with available data.’ ¹⁰⁴ The problem of induction, in the form that philosophers are concerned with, that of ‘justifying the plausibility of such hypotheses’, is ‘either deferred … or simply dismissed.’¹⁰⁵ While there may be good reasons to dismiss the problem of induction for other applications of AI, this is unsustainable for those seeking to create an AI capable of decision-making, command, prediction, and prescription in war.

To argue that the induction-based AI of Alpha Dogfight or Watson is compatible with decision-making in war is inevitably an argument that many of those decisions are (or can be) based on inductive inferences too. This requires a different logic to comprehend, conduct, and communicate (i.e. explain and justify) military-political behaviour: abductive logic. The example of Alpha Dogfight shows that induction alone seems to be able to go some of the way to competently performing some wartime decisions, in particular here in operating an aircraft in a simulated lone combat engagement. The sleight of hand that military AI optimism plays is to extrapolate this very specific exercise to non-comparable types of command decisions. By focusing on command decisions (rather than ‘tactical’ versus ‘strategic’ levels of war), AI’s inbuilt limitations are easier to identify. AI currently cannot make judgements, but rather makes probabilistic inferences. Nor can it make useful decisions in the absence of comprehensive data in a closed system. The notion that despite this AI could advise commanders, or command forces itself, is only sustainable if the core aspect of command decision-making is ignored. Revisiting the concepts of tactics and strategy provides a corrective by drawing out the qualities of command decisions beyond simple questions of targeting and optimisation.

Categories of tactics and strategy are best used as analytical distinctions that are useful under some circumstances, while remembering that war is not purely rational, and is a phenomenon where passions, courage, will, and human character shape behaviour and possibilities. Raymond Aron commented that ‘the analysis of available means and of the end sought does not produce the decision alone; it never reduces it to a strict calculation of probability.’¹⁰⁶ The nature of war prevents command being merely a matter of probabilistic calculation at both tactical and strategic command levels, as both remain human activities that engender reason and passion within a chaotic universe. Regarding Napoleon’s supposed genius and indefatigable drive, David Chandler observed that ‘every quality has its perversion, and the dividing line between genius and madness is notoriously slender.’¹⁰⁷ Trying to reproduce human passions and genius in machines, or making such subjective characteristics comprehensible as digital datasets, poses insurmountable logical challenges to the inductive rationalism of narrow AI algorithms and machine learning. It is at odds with the AI optimists’ kitsch conception of war which treats wars like games with neatly categorised phenomena, domains, and attributes.

The genius of decision-making applies in tactics as well as strategy, providing a strong reason to see command decisions across tactical and strategic levels of war rather than being only the preserve of the top echelon of strategists and political leaders. A general AI’s capabilities are needed to lead in battle in order to begin to truly transform the role of human reason and passion in war. Narrow AI could not command (or support the command decisions) of a whole battle (tactics) or a series of them (operational art) because these are arts reliant on building material and immaterial situational awareness and making judgements with incomplete and contradictory information, and often with little guidance ‘in the moment’. Genius is desirable in a commander because rules of thumb are only as good as they are applied in an intelligent and critical fashion at the commander’s discretion. ‘Precepts are vain’, the Welsh mercenary-General and philosophe Henry Lloyd maintained, and genius alone knows how and in what shape to apply any rules of thumb on when to attack, defend, counter, retreat, disperse, concentrate, or admit defeat.¹⁰⁸ AI has not yet been proven to be able to see what is not there with no prompts and act accordingly in a reasonable fashion; it cannot perform abductive logic. It can act only on what data it has been provided with. This quality of genius and the need for abductive thought in command is something that cannot be taught – and therefore cannot be programmed or clearly defined in a historical dataset. This recalls the failed, previous generation of Expert Systems AI research that attempted to catalogue rules and facts for a machine to then reason from,¹⁰⁹ but the machine-learning based narrow AI paradigm continues to have the same drawback. Programming AI with objectified datasets of military history risk guiding AI down the road of mechanical thinking according to dogmatic adherence to rigid rules, where taxonomic empirical knowledge is valued more than a grasp of tactics, strategy, and politics, like the ‘Modern Major-General’ in The Pirates of Penzance.

Genius in command is not only learning about when and how to apply principles, but when to do the opposite of what may be generally good practice. As Musashi wrote,

having attained a principle, one detaches from the principle; thus one has spontaneous independence from the science of martial arts and naturally attains marvels… The way to win in a battle according to military science is to know the rhythms of the specific opponents, and use rhythms that your opponents do not expect, producing formless rhythms from rhythms of wisdom.¹¹⁰

Grappling at the same concept of genius or coup d’oeil, this well-recognised intuitive aspect of strategy, art, decision-making, paradox, surprise, cunning, and luck, is something that is beyond AI. Musashi admitted to this pedagogical difficulty, confessing that ‘it is by no means possible for me to write down this science precisely as I understand it in my heart. Yet, even if the words are not forthcoming, the principles should be self-evident.’¹¹¹ Such intuitive and experience-laden teaching is alien to the inductive logics of narrow AI or Gilbert and Sullivan’s Major-General. Critics of AI in military affairs have not gone far enough in applying this understanding to assessing the potential of AI. Goldfarb and Lindsay warn that linking autonomous weapons systems to AI predictions ‘can quickly make tragic mistakes’ and correctly argue that asking AI to be a ‘strategic corporal’ is too much.¹¹² Going further, Hoffman cautions that as AI ‘matures’, it will make stupid decisions, but then forecasts an era of ‘cyber d’oeil’ once the AI has ‘matured’.¹¹³ However, we argue that the inability of an AI to be a strategic corporal, or fully expecting stupid military and political decisions, is part of the inherent nature of narrow AI. The AI ‘strategic corporal’ remains an unattained standard because the burdens of command are never totally concentrated at the top of the chain of command and rely on abductive logic that only ‘genius’ and human decision-making thought can perform.

A commander and their advisors facing the challenges posed by passion, reason, and chance must have ‘courage’ to make prudent calculations of all kinds in the face of danger.¹¹⁴ Sometimes, correct decisions are made but luck is simply against the commander and undoes what in other situations could be perfectly feasible plans. An AI capable of command would have to learn to distinguish between bad ideas that succeed with luck or enemy incompetence, or good ideas done badly or which suffer from external events that none could have foreseen. That is, while probabilistic inference is useful in war, it is surrounded by subjective judgements, interpretations, and calculations before, during and after the event.¹¹⁵ This is often due to prudent qualitative calculations providing no clearly better course of action or one without high levels of risk. Such courage and self-confidence to fling oneself into uncertainty is required by a good commander or leader. A genius embraces the unknown for the possibilities it affords, and does not attempt to eliminate the uncertainty of war or ‘gamify’ warfare by attempting to turn war into an exercise of pure physics and mathematics. A good commander knows uncertainty will always be there, whether in tactics or in strategy.¹¹⁶ An AI commander (or advisor) cannot help but constantly operate in unhelpful incomprehension of this fact. Narrow AI cannot hope to perform the judgement and empathic qualities needed to make and understand all the requirements and implications of command and leadership decisions in warfare, at any level.

Classical military theory does not talk in the formal, technical terms of induction and inferences. However, this section has gone some way to showing that command decisions and genius require more than induction alone and require an abductive logic. Igor Douven explains that

both [induction and abduction] are ampliative, meaning that the conclusion goes beyond what is (logically) contained in the premises … but in abduction there is an … appeal to explanatory considerations, whereas in induction there is not; in induction, there is only an appeal to observed frequencies or statistics.¹¹⁷

Abductive reasoning reaches out to propositions that were not contained in or elaborated from the existing dataset, and requires explanatory argumentation to ‘make sense’. The Duke of Wellington supposedly said that ‘all the business of war … is to endeavour to find out what you don’t know by what you do; that’s what I called “guessing what was at the other side of the hill.”’ ¹¹⁸ Command decisions rely on ‘going beyond’ what is known already, but via an educated guess rather than quantitative analysis. The key aspect pertaining to command decisions is how to determine the conclusion that best explains a situation in the absence of data rather than a statistical calculation and prediction based on only the data available. War relies on the art of command decisions in the face of unknowns and the unknowable, not only what can be observed.

Clausewitz did not highly rate induction’s value in how decisions are made in war. He wrote that:

It is at all times only conjecture or guesses at truth that we have to act upon. This is why differences of opinion are nowhere so great as in war, and the stream of impressions acting counter to one’s own convictions never ceases to flow.¹¹⁹

The emotional, chaotic, and political nature of war is why decisions must be justifiable rather than probabilistically derived. Military action must ‘make sense’ in political and cultural terms, else the emotional nature of war will manifest in a lack of trust in commanders. This means that AI’s emotionless and occasionally puzzling calculation is not an advantage but rather a liability. The problem of induction and AI’s inability to engage in abductive reasoning leads us to consider what knowledge about war is possible at all. Specifically, what data exists, or could exist, for an AI to be trained upon? Grappling with the problem of induction and the absence of abductive inference in existing AI models is a productive way to provide a stronger analytical foundation for AI’s current and future decision-making value. Payne explains these differences with psychology and neuroscience, but does not take advantage of established insights from philosophy of science and formal logic, gesturing instead to AI’s inability to fear mortality or feel emotions.¹²⁰ This is definitely a part of the impediment to useful military AI as Payne correctly observes, but ultimately it is the lack of ability to reason abductively that means there is no prospect AI can understand or explain existing data and command decisions in war.

The why (i.e. the causality) of the outcomes of fighting even at a tactical level is subjective, contestable knowledge, and at worst unknowable. Abductive logic meets multi-causal thought in investigating such military histories. Due to the nature of war and the problem of induction, however, gaining greater and greater levels of detail about specific battles has not led to new insights beyond the purely mechanical. DARPA attempted to model a battle from the 1991 Gulf War – 73 Easting – in the tiniest detail that they could within a computer model called SIMNET.¹²¹ A Wired journalist breathlessly claimed ‘The Battle of 73 Easting has become the single most accurately recorded combat engagement in human history.’¹²² It serves as a test case for the problem of induction in Strategic Studies. SIMNET’s methodology was inductive – it was not known exactly what insights would flow from gathering all of the various data points together, but the researchers hypothesised that they might learn something useful.¹²³ General Larry Welch argued

This opportunity provided by 73 Easting … comes at a time when sensors, data collection, information processing, communications and display technologies … can be brought together … to add to our understanding of strategies, concepts and forces that will win in the future.¹²⁴

In the end, 73 Easting’s in-depth analysis could only add one additional point of debate – or at worst a footnote – to discussing the causes of the Coalition’s successes. The more the analyst deviates into particulars and away from generalities, the less useful specific examples become as analogies for the present or future. Taken together, Clausewitzian insights about the nature of war combined with the problem of induction, mean that the project of SIMNET could never succeed in discovering strategies that ‘will win in the future’. The hopes and dreams that propelled SIMNET still live on in a state’s defence programs today. The UK Ministry of Defence has Project Improbable, a full ‘synthetic environment’ simulating a world with cities and people – and an appealing place, as an AI optimist has argued, to train an AI in the art of war.¹²⁵ There are already warning signs that training AI on synthetic data results in a ‘self-consuming loop’ that results in deterioration of quality or diversity of output.¹²⁶ Endeavours like Project Improbable are doomed to be expensive failures if the hope is to develop a model of an AI Major-General capable of command decisions or advising on command choices.

Conclusion

Payne argues that strategy is the ‘thinking part of warfare’ and describes AI as a ‘master tactician’ and ‘a moron strategist’, implying that command decisions in strategy is different to tactics, with narrow AI being successful at the latter based on its success in playing games.¹²⁷ We believe that narrow AI will in fact remain a halfwit tactician as well as a ‘moron strategist’, because tactics is also the thinking part of warfare and requires the same kind of logic as strategy and politics. Being good at Chess does not make an AI good at devising a plan to storm a redoubt. Narrow AI technologies are incapable of being reliable decision-makers or command advisors in war, despite the successes of narrow AI in playing Chess or Go. Success in games is not sufficient evidence for being good at warfare, both tactical and strategic.

We are able to make this bold claim by revisiting the core concepts of strategic theory that shapes our most fundamental understandings of the conduct of war. In addition, we have provided a novel and rigorous application of a philosophy of science approach to AI that grapples with the core inability of narrow AI in performing abductive logic which is fundamental to command decisions, and the flawed empirical foundations over the claims of AI in military command roles based on games and simulations. The criticism we make is less about debating the merits of human control and AI decisions and advice, but more that AI as we know it simply cannot make appropriate command decision because it is engaged in a different kind of logic that can only make sense within a kitsch, gamified vision of war. Narrow AI or machine learning can enhance autopiloting, ISR, administration, and raw data analysis, yes, but conceiving of a city’s siege, seeing it through, and routing a force following a break in the lines are different kinds of activities that need the kind of logic and intellect that narrow AIs have not been engineered to perform. Narrow AI should be viewed simply as a software-defined machine that can learn how to operate specific hardware or analyse quantifiable digitised datasets via pattern recognition, not an intelligence and not with personable qualities. It is inductive inference-making at high speed and mass scale, with all the inescapable shortcomings that this entails when war requires abductive logic and empathy.

We hope that our argument and our approach can be a reference point to improve the empirical and epistemological foundations upon which academic research and policy discussions on the use of AI in military decision-making rest. AI optimist arguments have captured the hearts of many defence bureaucracies, yet they have crucial empirical and epistemological shortcomings in their current form. The evidence supporting AI optimism does not heed the warnings of the likes of Clausewitz, Musashi and Lloyd. For AI optimist arguments about the utility of AI commanders and advisors to carry weight, the Clausewitzian universe that underpins modern military thought must be rejected, making them anti-Clausewitzian. The Prussian himself argued that ‘positive approaches to a theory of war were omitting the very real difficulties of its practice and that its conduct has no fixed limits in any direction’.¹²⁸ There is no such thing as the optimal solution or an objectively best course of action in war based on only what has come before. Rather, there is only what the commander thinks is best in the moment in the face of unknowns and unknowables, and competing interpretations in interminable debates of future military history. Existing machine success on the (often simulated) battlefield does not relate to command decision-making capability, only the most basic mechanical-technical processes of combat platform operations and networking.

While narrow AI has indeed proven itself adept at playing clearly defined games, that is no basis for developing trust in AI algorithms for making reliable command decisions or advisories. The impact for government AI policy is clear: pouring money into narrow AI with the hope of producing genius machine commanders is folly, ending up with something more like the ‘modern Major-General’ of The Pirates of Penzance. Perhaps it is for the best that AIs restrict themselves to a nice game of Chess.

Acknowledgements

The authors would like to thank the reviewers for their helpful comments, aiding us in clarifying our argument. Our thanks also to the group of scholars who kindly attended our paper workshop at the University of Leicester and provided suggestions at an early stage, and to Dr Clare Stevens for thoughtful feedback on a later draft. Finally, thanks are due to Dr Brian Weeden for introducing us to Dr Erik Larson’s book and thereby unwittingly providing the initial spark to write this paper.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The work was supported by the European Research Council [866155].

Notes on contributors

Cameron Hunter

Cameron Hunter is a Research Associate in Nuclear Politics at the School of History, Politics, and International Relations, University of Leicester.

Bleddyn E. Bowen

Bleddyn E. Bowen FHEA is Associate Professor of International Relations at the School of History, Politics, and International Relations, University of Leicester, specialising in strategic theory, space warfare, and astropolitics. He has published widely on space policy, doctrine, and strategy, including two monographs: Original Sin: Power, Technology and War in Space (Hurst/Oxford University Press, 2022) and War in Space: Strategy, Spacepower, Geopolitics (Edinburgh University Press, 2020). He has advised on space policy and strategy to military and government institutions across the transatlantic community and beyond. He founded and co-convenes the British International Studies Association’s Astropolitics Working Group, and is an Associate Fellow of the Royal United Services Institute.

Notes

1 Lawrence Freedman, Command: The Politics of Military Operations from Korea to Ukraine (Allen Lane, 2022), 2.

2 Freedman, Command, 1-9.

3 Carl von Clausewitz, On War, O.J. Matthijs Jolles, trans. in: Caleb Carr, ed. The Book of War (New York: Modern Library, 2000), 328-354.

4 Clausewitz, On War (Jolles), 786-787.

5 Yee-Kuan Heng, ‘Reflexive Rationality and the Implications for Decision-Making’ in Heidi Kurkinen (ed.) Strategic Decision-Making in Crisis and War, (Helsinki, National Defence University 2010), 21-22.

6 Yee-Kuan Heng, ‘Reflexive Rationality’, 22.

7 Ibid, 22-23.

8 See section 1. See also Hoffman, ‘Will War’s Nature’, 22, 27-28; Goldfarb and Lindsay, ‘Prediction and Judgment’, 39.

9 Andreas Herberg-Rothe (2014) ‘Clausewitz’s Concept of Strategy – Balancing Purpose, Aims and Means’, Journal of Strategic Studies, 37/6-7, 904.

10 Raymond Aron, Clausewitz: Philosopher of War, Christine Booker and Norman Stone (trans.), (Englewood Cliffs: Prentice-Hall, 1985), 328.

11 Brett A. Friedman, On Operations: Operational Art and Military Disciplines (Annapolis, M.D.: Naval Institute Press, 2021).

12 Ibid, 5.

13 Maaike Verbruggen, ‘AI & Military Procurement: What Computers Still Can’t Do’, War on the Rocks (blog), 5 May 2020, https://warontherocks.com/2020/05/ai-military-procurement-what-computers-still-cant-do/. See also Sam Tangredi and George Galdorisi, ‘Introduction’, in Sam Tangredi and George Galdorisi (eds.) AI at War: How Big Data, Artificial Intelligence and Machine Learning are Challenging Naval Warfare (Annapolis, M.D.: Naval Institute Press, 2021), 3.

14 Elke Schwarz, ‘Autonomous Weapons Systems, Artificial Intelligence, and the Problem of Meaningful Human Control’, The Philosophical Journal of Conflict and Violence 5/1, 53-72; John Emery, ‘Algorithms, AI, and Ethics of War’, Peace Review 33/2 (2021), 205-212; Neil Renic, ‘A Gardener’s Vision: UAVs and the Dehumanisation of Violence’, Survival 60/6, 57-72; Heather Roff, ‘The Strategic Robot Problem: Lethal Autonomous Weapons in War’, Journal of Military Ethics 13/3 (2014), 211-227; Lucy Suchman, ‘Algorithmic warfare and the reinvention of accuracy’, Critical Studies on Security 8/2 (2020), 182.

15 Erik Larson, The Myth of AI: Why Computers Can’t Think the Way We Do (London: Harvard UP), 60-61.

16 DoD, ‘Summary of the 2018 Department of Defense Artificial Intelligence Strategy’, (Washington DC: GPO 2019), 11, https://media.defense.gov/2019/Feb/12/2002088963/-1/-1/1/SUMMARY-OF-DOD-AI-STRATEGY.PDF.

17 See for example: Kathleen McKendrick, ‘The Application of Artificial Intelligence in Operations Planning’, NATO STO (2017); Michael Rüegsegger et al., ‘Deep Self-optimizing Artificial Intelligence for Tactical Analysis, Training and Optimization’, NATO STO (2021); Alex Wilner, ‘Artificial Intelligence and Deterrence: Science, Theory and Practice’, NATO STO (2019), 14-11.

18 MoD, ‘Defence Artificial Intelligence Strategy’, (London: HMG 2022), 1. https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/1082416/Defence_Artificial_Intelligence_Strategy.pdf; see also Helen Warrell, ‘UK military planners deploy AI to gain edge over adversaries’, Financial Times, 12 March 2021. https://www.ft.com/content/94d59a36-099a-4add-80d3-475127b231c7.

19 Avi Goldfarb and Jon R. Lindsay, ‘Prediction and Judgment: Why Artificial Intelligence Increases the Importance of Humans in War’, International Security 46/3 (2022), 9-10.

20 See Yuna Wong et al., ‘Deterrence in the Age of Thinking Machines’ (Santa Monica: RAND 2020) for a discussion of AI’s escalatory tendencies. See footnote 2 for literature providing ethical critiques.

21 MoD, ‘Defence AI Strategy’, 15.

22 Kenneth Payne, I, Warbot: The Dawn of Artificially Intelligent Conflict, (London: Hurst 2021), 83.

23 Ibid, 186-188.

24 Ibid, 192.

25 For example Payne argues some aspects of war are ‘less bounded’ but then later argues that war is unbounded, see Payne, I Warbot, 2, 76, 174.

26 Goldfarb and Lindsay, ‘Prediction and Judgment’, 50.

27 Ibid, 48.

28 Ibid.

29 N. Katherine Hayles, Unthought: The Power of the Cognitive Nonconscious, (Chicago: Chicago UP 2017), 10-11, 24.

30 Eda Kavlakoglu ‘AI vs. Machine Learning vs. Deep Learning vs. Neural Networks: What’s the Difference?’ IBM Cloud Blog, 27 May 2020. https://www.ibm.com/cloud/blog/ai-vs-machine-learning-vs-deep-learning-vs-neural-networks. ‘Narrow’ AI can be contrasted against ‘general’ or ‘strong’ AI. A ‘general’ AI only exists in science fiction, and there is no known program of research that could lead to it, according to a former insider. See Larson, Myth of AI.

31 Wong et al., ‘Deterrence in the Age of Thinking Machines’, 20.

32 Paul Scharre, Army of None: Autonomous Weapons and the Future of War (New York: W.W. Norton 2018), 163-164.

33 Julian S. Corbett, Principles of Maritime Strategy (Mineola: Dover 2004) 165-166.

34 See Matthew Price, Stephen Walker, and Will Wiley. ‘The Machine Beneath: Implications of Artificial Intelligence in Strategic Decision Making’ PRISM 7/4 (2018), 92-105; Payne, I Warbot, 28.

35 DARPA, ‘Generating Actionable Understanding of Real-World Phenomena with AI’, 4 Jan. 2019. https://www.darpa.mil/news-events/2019-01-04.

36 Ibid.

37 James Johnson, ‘Delegating strategic decision-making to machines: Dr. Strangelove Redux?’, Journal of Strategic Studies 45/3 (2022), 9; Zhimin Zhang et al., ‘Artificial intelligence in cyber security: research advances, challenges, and opportunities’, Artificial Intelligence Review 55 (2022), 1045.

38 Department of Defence, ‘Summary of the Joint All-Domain Command & Control (JADC2) Strategy’, March 2022, 3. https://media.defense.gov/2022/Mar/17/2002958406/-1/-1/1/SUMMARY-OF-THE-JOINT-ALL-DOMAIN-COMMAND-AND-CONTROL-STRATEGY.PDF. See also Lucy Suchman, ‘Imaginaries of omniscience: Automating intelligence in the US Department of Defense’, Social Studies of Science (2022), 1-26.

39 Payne, I Warbot, 2, 76.

40 Patrick Tucker, ‘An AI Just Beat a Human F-16 Pilot In a Dogfight – Again’, DefenseOne, 20 August 2020, https://www.defenseone.com/technology/2020/08/ai-just-beat-human-f-16-pilot-dogfight-again/167872/; James Johnson, ‘Automating the OODA loop in the age of intelligent machines: reaffirming the role of humans in command-and-control decision-making in the digital age’, Defence Studies, (2022), 10; Amir Husain, ‘AI is Shaping the Future of War’, PRISM 9/3, 54; Payne, I Warbot, 92; Jazmin Furtado and Chris Dylewski, ‘AlphaDogfight should scare the Air Force straight … into scaling AI efforts’, C4ISRNet, 21 January 2021, https://www.c4isrnet.com/thought-leadership/2021/01/21/alphadogfight-should-scare-the-air-force-straight-into-scaling-ai-efforts/.

41 Richard Spencer, ‘Killer drones used AI to hunt down enemy fighters in Libya’s civil war’, The Times, 3 June 2021, https://www.thetimes.co.uk/article/killer-drones-used-ai-to-hunt-down-enemy-fighters-in-libyas-civil-war-2whlckdbm.

42 Goldfarb and Lindsay usefully doubt the validity of Alpha Dogfight as evidence because it was fooled by a simpler threat after becoming overly accustomed to complexity. Goldfarb and Lindsay, ‘Prediction and Judgement’, 35

43 Palantir, ‘AIP for Defense’, Palantir.com, 2023, https://www.palantir.com/platforms/aip/.

44 Clausewitz, On War, 328-329.

45 Examples of direct favourable references include Payne, I, Warbot, 96-98; James Johnson, AI and the Future of Warfare, 30, 115; see Suchman, ‘Imaginaries of Omniscience’ for a comprehensive analysis.

46 Handel, Masters of War, 353-360.

47 Ibid., 355.

48 See Wong et al. 6 for their link between this ‘benefit’ and Boyd.

49 James Johnson, ‘Artificial intelligence & future warfare: implications for international security’, Defense & Security Analysis 35/2 (2019), 148.

50 Ibid, 150.

51 Ayoub and Payne, ‘Strategy in the Age of AI’, 799.

52 Goldfarb and Lindsay, ‘Prediction and Judgment’, 20, 35, 42.

53 Francis G. Hoffman, ‘Will War’s Nature Change in the Seventh Military Revolution?’, Parameters, 47/4 (2017), 22, 27.

54 Keith Dear, ‘AI and Decision-Making’, RUSI Journal 164/5-6 (2019), 25.

55 Ibid, 20, see also 22-23.

56 Johnson, ‘Delegating strategic decision-making to machines’, 8.

57 John Arquilla, Bitskrieg: The New Challenge of Cyberwarfare (Polity, 2021), 78.

58 McKendrick, ‘The Application of AI in Operations Planning’, 2.1-6

59 See also Payne, I, Warbot, 4, 68, 170-171.

60 Dear, ‘AI and Decision-Making’, 23.

61 Ayoub and Payne, 794.

62 Hoffman, ‘Will War’s Nature’, 22.

63 Payne, I, Warbot, 68, 152, Arquilla, Bitskrieg, 83-84. Aycock and Glenney, conversely, correctly argue that ‘AlphaGo ain’t warfare, and it ain’t strategy’ but seem to accept some ‘tactical’ nous of AI. As our subsequent section shows, this argument is not pessimistic enough because it does not assess the logical basis of AI decision making. See Adam Aycock and William Glenney, ‘Trying to Put Mahan in a Box’, in Sam Tangredi and George Galdorisi (eds.) AI at War: How Big Data, Artificial Intelligence and Machine Learning are Challenging Naval Warfare (Annapolis, M.D.: Naval Institute Press, 2021), 265-285.

64 Feng-hsiung Hsu, ‘IBM’s Deep Blue Chess Grand Master Chips’, IEEE MICRO, 19/2, (1999), 70-71.

65 Ibid, 71-72.

66 Ibid, 76.

67 Fei-Yue Wang et al., ‘Where does AlphaGo go: from church-turing thesis to AlphaGo thesis and beyond’, IEEE/CAA Journal of Automatica Sinica, 3/2 (2016), 115.

68 Ibid, 116; Larson, Myth of AI, 125.

69 Larson, Myth of AI, 162; Diego Perez et al., ‘Multiobjective Monte Carlo Tree Search for Real-Time Games’, IEEE Transactions on Computational Intelligence and AI in Games 7/4, (2015), 348.

70 Richard Waters, ‘Man beats machine at Go in human victory over AI’, Ars Technica, 19 Febraury 2023, https://arstechnica.com/information-technology/2023/02/man-beats-machine-at-go-in-human-victory-over-ai/

71 Ibid.

72 Dear, ‘AI and Decision-Making’, 23-34.

73 A product of AI optimists’ selectivity – many manual wargames avoid battle-centrism and make pains to simulate friction.

74 Clausewitz, On War, (Jolles) 266.

75 Ibid, 277.

76 Aron, Clausewitz, 57-59.

77 Musashi, Book of Five Rings, 77.

78 Clausewitz, On War, 321.

79 See also Michael Howard, ‘The Use and Abuse of Military History’, Parameters, 11/1 (1981), 13.

80 Alexis Madrigal, ‘How Checkers Was Solved’, The Atlantic, 19 July 2017, https://www.theatlantic.com/technology/archive/2017/07/marion-tinsley-checkers/534111/.

81 Feng-Hsiung Hsu, ‘Cracking Go’, IEEE Spectrum, October 2007, 51-55.

82 Payne (see I Warbot, 40-42) mentions Gödel and Turing’s proofs of mathematical incompleteness, but does not explain how AI will overcome it or engage with the biggest hurdle we identify – undecidability.

83 Alex Churchill et al., ‘Magic: The Gathering is Turing Complete’, arXiv.org, https://arxiv.org/abs/1904.09828 (2019), 2.

84 Ibid.

85 Churchill quoted in Jennifer Ouellette, ‘It’s possible to build a Turing machine within Magic: The Gathering’, ArsTechnica, 23 June 2019. https://arstechnica.com/science/2019/06/its-possible-to-build-a-turing-machine-within-magic-the-gathering/.

86 See Hannah Jane Parkinson, ‘Paul the octopus, Taiyo the otter and the World Cup’s other psychic animals’, The Guardian, 12 December 2022, https://www.theguardian.com/sport/2022/dec/12/paul-the-octopus-taiyo-the-otter-world-cup-psychic-animals.

87 Herberg-Rothe, ‘Clausewitz’s Concept of Strategy’, 905.

88 Clausewitz, On War¸270.

89 Ibid, 280.

90 Ibid, and see also 355-357.

91 Ibid, 289-290.

92 With the exception of Deep Blue.

93 Larson, Myth of AI, 1, 41.

94 Dear, ‘AI and Decision-Making’, 18.

95 Job de Grefte, ‘Epistemic benefits of the material theory of induction’, Studies in History and Philosophy of Science 84, (2020), 101.

96 Larson, Myth of AI, 115.

97 Leah Henderson, ‘The Problem of Induction’, The Stanford Encyclopedia of Philosophy, Edward N. Zalta (ed.), (2020) https://plato.stanford.edu/archives/spr2020/entries/induction-problem/.

98 As quoted in Henderson ‘Problem of Induction’, n.p.

99 Larson, Myth of AI, 124, Jochen Runde, ‘Dissecting the Black Swan’, Critical Review 21/4, (2009), 491-505.

100 See for example Davor Lauc, ‘Machine Learning and the Philosophical Problems of Induction’ in Sandro Skansi (ed.), Guide to Deep Learning Basics (Springer 2020), 93–106.

101 See Mary Cummings, ‘Artificial Intelligence and the Future of Warfare’, Chatham House (2017), 7. https://www.chathamhouse.org/sites/default/files/publications/research/2017-01-26-artificial-intelligence-future-warfare-cummings-final.pdf.

102 Andrew Ilachinski, ‘AI, Robots, and Swarms Issues, Questions, and Recommended Studies’, CNA (2017), 65 https://www.cna.org/archive/CNA_Files/pdf/drm-2017-u-014796-final.pdf.

103 Watson was also programmed with human-defined rules to guide its inductive inference-making. See Larson, Myth of AI, 222-224.

104 Guglielmo Tamburrini ‘Artificial Intelligence and Popper’s Solution to the Problem of Induction’, in Ian Jarvie et al. (eds.), Karl Popper A Centenary Assessment Volume II: Metaphysics and Epistemology, (Aldershot: Ashgate 2006), 265-284

105 Tamburrini, ‘AI and Popper’, 267.

106 Aron, Clausewitz, 113.

107 David Chandler, The Campaigns of Napoleon (London: Weidenfeld and Nicholson, 1993), xl.

108 Henry Lloyd, The History of the Late War in Germany, Vol. I (1766) in: Patrick J. Speelman, ed. War, Society and Enlightenment: The Works of General Lloyd (Leiden: Brill, 2005) 114.

109 Wong et al., ‘Deterrence’, 19-20, Larson, Myth of AI, 54.

110 Miyamoto Musashi, The Book of Five Rings, Thomas Cleary trans., ed. (London: Shambala, 2003), 14, 22.

111 Ibid, 24.

112 Goldfarb and Lindsay, ‘Prediction and Judgment’, 34, 44.

113 Hoffman, ‘Will War’s Nature’, 28.

114 Clausewitz, On War, 277-278.

115 See Jon T. Sumida, Inventing Grand Strategy and Teaching Command: The Classic Works of Alfred Thayer Mahan Reconsidered (Washington, D.C.: Woodrow Wilson Center Press, 1997), 106.

116 Clausewitz, On War, 278-279.

117 Igor Douven, ‘Abduction’, in Edward Zalta (ed.), The Stanford Encyclopedia of Philosophy, (Summer 2017). https://plato.stanford.edu/archives/sum2017/entries/abduction/.

118 as quoted in The Croker Papers, Louis Jennings (ed.), Vol. II, (New York: Scribner 1884), 463.

119 Clausewitz, On War, 308.

120 See for example Payne, I Warbot, 25-26, 54-55, Kenneth Payne, Strategy, Evolution and War, (Washington DC: Georgetown UP 2018), passim.

121 DARPA (1992) ‘73 EASTING: Lessons from Desert Storm via Advanced Distributed Simulation Technology’, (Alexandria: IDA 1992). https://apps.dtic.mil/sti/pdfs/ADA253991.pdf.

122 Bruce Sterling (1993) ‘War is Virtual Hell’ Wired, 1 Jan. 1993. https://www.wired.com/1993/01/virthell/

123 Sharon Weinberger (2017) Imagineers of War, 288-290.

124 DARPA (1992) I-9.

125 Payne, I Warbot, 16.

126 Sina Alemohammad et al., ‘Self-Consuming Generative Models Go MAD’, arXiv.org (2023), https://arxiv.org/pdf/2307.01850.pdf.

127 Payne, I, Warbot, 1, 74.

128 Clausewitz, On War, (Jolles) 338, 340