AI, AGI, and Hyperwar

LESSON METADATA

SECTION 0 — Why This Lesson Exists

Artificial intelligence is not merely a new weapons platform; it is a transformation of the cognitive architecture of war. Where earlier revolutions in military affairs changed how force was delivered—gunpowder, mechanization, nuclear weapons—AI changes how decisions are made, how targets are identified, and how escalation unfolds. The puzzle this lesson addresses is whether the compression of decision cycles by algorithmic systems creates a novel and potentially uncontrollable pathway to war—what some analysts call “hyperwar”—and whether governance frameworks designed for human-paced conflict can survive machine-paced conflict. The student must learn to distinguish three analytically separate questions: whether AI changes the probability of war, whether it changes the character of war, and whether it changes who can wage war. Each question has a different theoretical answer and a different evidentiary base, and conflating them is the most common error in the policy literature.

The lesson draws on four theoretical traditions, each with a distinct mechanism. Realism supplies the security-dilemma logic: if AI confers first-mover advantages in targeting and command, states will race to deploy it, and the resulting instability may be structural rather than chosen. Liberal institutionalism supplies the governance question: can regimes like the Convention on Certain Conventional Weapons (CCW) or the EU AI Act constrain military AI, or will security externalities defeat regulation? Constructivism supplies the normative question: how do shared understandings of “meaningful human control” emerge, and can they be universalized across civilizational frames? Critical and post-colonial theory supplies the vantage that the other three often occlude: that AI is not a neutral technology but an infrastructure of power whose training data, compute, and standards are concentrated in a handful of states and corporations, reproducing global hierarchies. The canonical sources anchor these traditions: Scharre (2018) on the autonomy debate and the principal-agent problem in lethal systems; Horowitz (2018) on the diffusion of military innovations and the organizational barriers to adoption; Kissinger, Schmidt, and Huttenlocher (2021) on AI and the transformation of world order; Bostrom (2014) on the security dilemma of superintelligence and the alignment problem.

The African vantage is not a regional illustration of a Western theory; it is a distinct epistemic position. From this position, the central fact is the AI divide: the concentration of compute, data, talent, and platform control in a small number of states and firms, overwhelmingly outside Africa. African scholarship—from Rodney’s analysis of underdevelopment as a historical process, through Nkrumah’s neo-colonialism, to contemporary work by Aning, Danso, and others on African agency in global governance—frames AI not as a neutral tool but as an infrastructure that can deepen dependency. The demand is for equitable governance: African participation in standard-setting, data sovereignty, and the right to develop AI capacity rather than merely import it. The African perspective also notices what others miss: that AI-enabled warfare is already present on the continent (drone strikes in the Sahel, algorithmic surveillance in urban policing), but African states are rule-takers in the forums where the rules are written. This asymmetry is the lesson’s central puzzle viewed from its most revealing angle.

The lesson develops five mechanisms in sequence. First, narrow AI in military systems—ISR, targeting, logistics, and command and control—where the gains are incremental but the integration is deep. Second, Lethal Autonomous Weapons Systems (LAWS) and the CCW debate, where the principal-agent problem (who is responsible when a machine kills?) meets international humanitarian law. Third, decision compression and “hyperwar”: the risk that algorithmic speed collapses the deliberation window and produces flash escalation. Fourth, AGI/ASI and the security dilemma of superintelligence, where alignment becomes a global security problem rather than a technical one. Fifth, AI governance—the US executive order, the EU AI Act, Chinese regulations, and the UN framework—where the question is whether any of these can keep pace with capability.

After mastering the lesson, the student will be able to: (a) identify the mechanism by which a given AI system alters a military function; (b) reconstruct five perspectives on any AI-security case, steel-manning each; (c) adjudicate between competing explanations using evidence rather than assertion; (d) distinguish description, explanation, prediction, and prescription in policy memos; and (e) specify falsifiers for claims about AI and war. The worked example (Nagorno-Karabakh 2020 and Ukraine 2023-2024 drone deployments) and the problem pair (Israel’s Lavender and Gospel systems in Gaza) train these skills on real cases. The data lab trains quantitative and comparative reasoning on AI investment, autonomous systems inventories, and incident data.

The lesson is self-contained: all necessary concepts are defined inline, all sources are named, and no external reading is required to follow the argument, though the further-reading list points to depth. The student should read with a notebook, reconstructing arguments in their own words, because the assessment (the independent memo) requires original analysis rather than recall. The Bloom ceiling is Evaluate/Create: the student is expected not merely to describe but to judge and to design. The lesson’s central intellectual discipline is to hold mechanism and perspective in tension: every claim about what AI does must be paired with a claim about from whose vantage it is seen, and every governance proposal must be tested against the objection that it reflects the interests of its authors.

SECTION 1 — Pre-Class Priming

1.1 Retrieval Warm-Up

Answer each question from memory before reading the lesson. These draw on L24 (cyber and cognitive), L22 (conventional warfare), and L52 (semiconductors).

1. In L24, what is the difference between a cyber attack and a cyber operation, and why does the distinction matter for escalation theory?
2. L22 treated the offense-defense balance in conventional warfare. State one historical case where a technological innovation shifted that balance toward offense, and one where it shifted toward defense.
3. What is the “principal-agent problem” as discussed in L24’s treatment of cyber operations, and how does delegation to non-state proxies complicate attribution?
4. L52 examined semiconductor supply chains. Name the three chokepoints in the global chip supply chain and identify which state controls each.
5. In L22, how did the revolution in military affairs (RMA) literature of the 1990s define “information dominance,” and what critique did later scholars (including from the Global South) raise about its universal applicability?
6. L24 discussed cognitive warfare. Define “cognitive warfare” and give one example of a state employing it against an adversary’s decision-making process.

1.2 Orientation Questions

These questions frame the lesson’s central puzzle. Do not answer them yet; hold them in mind as you read.

1. Does AI change the probability of war, the character of war, or who can wage war—and can a single theory explain all three?
2. If algorithmic systems can compress the decision cycle from hours to seconds, what happens to the “deliberation window” that international humanitarian law and crisis management assume?
3. Is the demand for “meaningful human control” over lethal systems a universal norm or a culturally specific one, and who gets to define it?
4. Can governance frameworks written by AI-leading states be legitimate for AI-importing states, or do they reproduce the AI divide?

1.3 Vocabulary Pre-Teach

Provisional definitions; refine as you read.

• Narrow AI (weak AI): Machine-learning systems trained for specific tasks (image recognition, target classification, logistics optimization) that do not possess general reasoning; the dominant form of military AI today.
• AGI (Artificial General Intelligence): A hypothetical system capable of performing any intellectual task a human can, across domains; not yet achieved; treated in security studies as a future capability with present-day strategic effects.
• ASI (Artificial Superintelligence): A hypothetical system exceeding human cognitive performance across virtually all domains; the subject of Bostrom’s (2014) security-dilemma analysis.
• LAWS (Lethal Autonomous Weapons Systems): Weapons that can select and engage targets without further human intervention after activation; the focus of the CCW debate.
• Meaningful human control: The principle that a human must retain sufficient control over the use of force to be morally and legally responsible; contested in definition and scope.
• Hyperwar: A hypothesized form of conflict in which algorithmic speed collapses decision cycles, producing escalation faster than human deliberation can manage.
• Decision compression: The reduction of the time available for human judgment in a crisis due to the speed of AI-enabled sensing, targeting, and response.
• Principal-agent problem (in autonomous weapons): The difficulty of assigning responsibility when an autonomous system, acting as agent, produces an outcome the human principal did not intend or foresee.
• Alignment problem: The challenge of ensuring that an AI system’s objectives match human (or specified) values; in security terms, a global problem because misaligned superintelligence is a shared threat.
• AI divide: The asymmetry between states and firms that produce AI (compute, data, talent, platforms) and those that consume or are governed by it.
• Compute: The processing power (chips, data centers) required to train and run AI models; a strategic resource analogous to energy or fissile material.
• CCW (Convention on Certain Conventional Weapons): The UN framework under which states discuss LAWS; includes the Group of Governmental Experts (GGE) on LAWS.

SECTION 2 — LESSON CORE

2.1 Learning objectives

By the end of this lesson, the student will be able to:

1. Define and classify narrow artificial intelligence applications across the military kill chain—intelligence, surveillance, and reconnaissance (ISR), targeting, logistics, and command and control (C2)—and distinguish them rigorously from artificial general intelligence (AGI) and artificial superintelligence (ASI), specifying the technical and epistemic boundaries that separate pattern-recognition systems from systems capable of general reasoning (Bloom: knowledge/comprehension).

2. Explain the causal mechanism by which narrow AI compresses the observe-orient-decide-act (OODA) loop and redistributes cognitive labor between human operators and machines, identifying how training data, optimization objectives, and interface design jointly determine where lethal authority actually resides in a given system (Bloom: comprehension/analysis).

3. Evaluate scope conditions under which AI-enabled military systems increase operational effectiveness versus escalation risk, including data abundance, electromagnetic contestation, adversarial machine learning, distribution shift, and the feasibility of meaningful human control, and specify the falsifying evidence that would challenge each claim (Bloom: analysis/evaluation).

4. Reconstruct and adjudicate among the African (primary), Chinese, Russian, Western, and UN perspectives on narrow AI in military systems, identifying each tradition’s core contribution, its characteristic occlusion, and the normative commitments that drive its policy prescriptions (Bloom: synthesis/evaluation).

5. Apply the analytical framework to the 2020 Nagorno-Karabakh war and 2023–2024 Ukraine drone deployments, producing a memo that distinguishes description, explanation, prediction, and prescription, and that identifies at least two falsifiers for any causal claim advanced (Bloom: application/synthesis).

These objectives are assessed through the Problem Pair, Independent Memo, Hinge Question, and Data Lab components of this lesson. Mastery requires not merely recalling definitions but tracing mechanisms, specifying scope conditions, and holding multiple perspectives in productive tension without collapsing them into a single “correct” view.

2.2 KEY IDEA 1 — Narrow AI in military systems: ISR, targeting, logistics, and C2

Bolded Causal Claim: Narrow AI does not autonomously wage war; it compresses the OODA loop and redistributes cognitive labor across the kill chain, shifting the locus of lethal decision from the human operator to the system’s training data and optimization objective, thereby reconstituting the principal-agent relationship between commanders and machines in ways that are invisible to traditional accountability frameworks.

Mechanism: The mechanism operates through four interlocking channels. First, in ISR, narrow AI—principally computer vision, multi-sensor fusion, and natural-language processing—transforms raw sensor feeds into structured intelligence at speeds no human analyst can match. Systems such as the U.S. Project Maven pipeline apply convolutional neural networks to full-motion video and satellite imagery to detect, classify, and track objects of interest, converting an ocean of pixels into a ranked list of candidate targets. The mechanism here is not understanding but statistical pattern recognition: the system learns a mapping from input features to labels from training data, then generalizes (or fails to generalize) to new instances. Second, in targeting, narrow AI fuses ISR outputs with geospatial, signals, and open-source data to propose aimpoints, estimate collateral damage, and prioritize strikes according to commander-defined objectives. The optimization objective—minimize time-to-target, maximize probability of kill, minimize civilian harm as estimated by the model—becomes the de facto decision rule, and because the objective is encoded in code and weights rather than in an order that a human can read, the “decision” is distributed across the pipeline. Third, in logistics, predictive maintenance, demand forecasting, and route optimization reduce the friction of supply, which in turn raises the sustainable tempo of operations and lowers the political cost of prolonged campaigns; the mechanism is that AI-enabled logistics converts attrition warfare into a more continuous, lower-friction form of violence. Fourth, in C2, decision-support systems aggregate multi-domain information, propose courses of action, and simulate outcomes, effectively pre-structuring the commander’s choice set. The mechanism across all four channels is the same: narrow AI does not replace human judgment wholesale but re-locates it, compressing the time available for deliberation and embedding assumptions—about what counts as a target, what counts as harm, what counts as success—into technical artifacts that are difficult to inspect, contest, or revise under operational pressure. Scharre (2018) emphasizes that the critical variable is not autonomy per se but the degree of human control that remains meaningful given the speed and complexity of the system; Horowitz (2018) adds that the diffusion of these technologies will be shaped by organizational capacity and strategic culture, not merely by technical availability. Kissinger et al. (2021) frame the deeper risk: as machines assume cognitive tasks, human leaders may lose the ability to comprehend the strategic environment their systems are acting within, producing a form of strategic opacity that is itself a source of instability.

Scope Conditions: The claim that narrow AI compresses the OODA loop and redistributes cognitive labor holds under identifiable conditions and fails under others. It holds when (a) data is abundant, labeled, and representative of the operational environment; (b) the adversary’s behavior is sufficiently patterned to be learned; (c) the electromagnetic spectrum is permissive enough for sensors and communications to function; (d) rules of engagement are codifiable into optimization objectives; and (e) the organization has the technical capacity to maintain, update, and audit the system. It weakens or reverses when (a) data is scarce, biased, or drawn from a different distribution than the target environment—producing confident misclassification, as when a system trained on one theater is deployed in another; (b) the adversary employs adversarial machine learning, camouflage, or deception specifically designed to exploit the system’s learned features; (c) electromagnetic contestation degrades sensors and communications, forcing reversion to human judgment at precisely the moment tempo is highest; (d) the operational environment is novel in ways the training data did not capture, producing distribution shift and silent failure; and (e) organizational incentives favor speed over scrutiny, so that “human in the loop” becomes a rubber stamp rather than a genuine control. The claim also has a temporal scope: it describes the current generation of narrow AI, which excels at perception and optimization within bounded domains but lacks causal reasoning, common sense, and the ability to recognize its own ignorance. Should systems approach general intelligence, the mechanism would change qualitatively, which is the subject of Key Idea 4. A further scope condition concerns the level of analysis: at the tactical level, compression is most evident; at the strategic level, the effect is mediated by doctrine, alliance politics, and domestic politics, so that the same technical system can produce different strategic outcomes in different states.

Observable Implications: If the causal claim is correct, we should observe (i) measurable reductions in sensor-to-shooter timelines in conflicts where AI-enabled systems are deployed, as reported in Ukraine’s use of the Delta situational-awareness platform and in Israeli systems; (ii) procurement and doctrinal shifts toward AI-enabled platforms and away from platforms whose value depends on human-paced decision-making; (iii) new failure modes—automation bias, fratricide from misclassification, and “automation surprise” when the system behaves in ways operators did not anticipate; (iv) an arms-race dynamic in AI-relevant inputs, particularly advanced semiconductors and training data, linking this lesson to Lesson 52; (v) growing divergence between states that can field and sustain such systems and those that cannot, with implications for the African perspective developed below; and (vi) increasing opacity in accountability chains, as the number of human decision points shrinks and the technical complexity of each grows. Falsifying evidence would include: documented cases where AI-enabled systems slowed decision-making without improving accuracy; cases where human operators consistently overrode system recommendations with better outcomes; or evidence that the principal bottleneck in targeting remains human political authorization rather than technical processing speed.

Perspective Note (five traditions in brief):

African (primary). The African vantage begins from the technology gap and the AI divide, not from the battlefield. From within African scholarship, the relevant question is not how narrow AI compresses the OODA loop for those who possess it, but how the global distribution of AI capability reproduces and deepens structural inequality. Grovogui’s critique of the international order as a hierarchy of sovereignty applies directly: the states that build and control military AI are the same states whose sovereignty has historically been treated as full, while African sovereignty has been treated as conditional and tutelary. Mbembe’s concept of necropolitics—the power to dictate who may live and who must die—illuminates how algorithmic targeting systems trained on Western data and Western threat perceptions may encode a hierarchy of human value in which African lives are more readily classified as threats. Bayart’s analysis of extraversion suggests that African states’ engagement with military AI will be shaped by their position in global networks of dependency, importing systems they cannot audit or modify. Rodney’s underdevelopment thesis and Nkrumah’s neo-colonialism frame the material dimension: the AI divide is not a natural gap but a produced one, rooted in the extraction of data, talent, and minerals (including cobalt and coltan from the Democratic Republic of Congo) that feed the very systems that exclude African states. Mamdani’s work on the colonial state warns that imported technologies of control may reinforce bifurcated governance rather than strengthen accountable institutions. Aning and Danso’s security scholarship emphasizes that African security architectures—such as the African Standby Force—lack the technical infrastructure to field or govern AI-enabled systems, making the continent a consumer, not a producer, of military AI. Tamale’s intersectional analysis and Oyěwùmí’s epistemological critique raise the further point that the categories embedded in training data—what counts as a “combatant,” a “threat,” a “civilian”—carry gendered and racialized assumptions that may not map onto African social realities. The African perspective therefore reframes narrow AI in military systems as a problem of global equity and epistemic justice: the demand is not merely for African states to acquire AI but for equitable governance, technology transfer, capacity building, and a voice in setting the standards by which these systems are evaluated. What this perspective reveals that the others miss is that the “AI arms race” framing presupposes a symmetry of actors that does not exist; for much of the world, the relevant dynamic is not racing but being raced past.

Chinese. The Chinese perspective treats narrow AI in military systems through the lens of “intelligentized warfare” (智能化战争), a doctrinal concept that elevates AI to a core component of future combat power, while simultaneously framing AI governance as a sovereignty issue. Chinese official discourse emphasizes that AI development must serve national security and that external attempts to restrict China’s access to AI technologies—such as U.S. semiconductor export controls—constitute a threat to sovereign development. On the governance side, China has participated in UN CCW discussions on lethal autonomous weapons while resisting binding prohibitions that it argues would disadvantage developing states and lock in Western technological advantages. The Chinese perspective contributes an insistence that AI governance must not become a vehicle for technological hegemony, and it occludes questions about domestic surveillance applications of the same technologies and about the transparency of PLA AI programs.

Russian. The Russian perspective frames narrow AI within military modernization and strategic stability. Russian doctrine treats AI as a force multiplier that can offset conventional inferiority, with emphasis on autonomous systems, decision support, and information warfare. Moscow’s strategic-stability framing warns that AI-enabled compression of decision times—especially in nuclear C2—could undermine crisis stability by reducing the window for human deliberation. Russia has been skeptical of Western-led governance initiatives, viewing them as attempts to constrain competitors while preserving Western advantages, and has emphasized national sovereignty over AI development. The Russian perspective contributes a focus on escalation dynamics and nuclear risk, and it occludes the gap between aspirational doctrine and demonstrated capability, as well as questions about the use of AI in information operations.

Western. The Western perspective, particularly the U.S. and allied view, treats AI as a strategic priority and a source of military advantage, while simultaneously hosting the most developed internal debate about lethal autonomous weapons and algorithmic accountability. The U.S. Department of Defense has adopted ethical principles for AI and has invested heavily in programs such as Project Maven and the Joint All-Domain Command and Control (JADC2) concept. The Western perspective contributes technical sophistication, an institutionalized ethics discourse, and a self-critical tradition that examines civilian harm, algorithmic bias, and accountability gaps. It occludes the global distributional consequences of its own leadership and the tendency to treat Western threat perceptions as universal.

UN. The UN perspective frames narrow AI in military systems through international humanitarian law, human dignity, and multilateral governance. The Secretary-General has called for a ban on lethal autonomous weapons systems that function without human control, and the CCW Group of Governmental Experts has debated definitions, characteristics, and possible regulatory frameworks for years without reaching consensus. The Pact for the Future includes commitments on AI governance that emphasize human rights, capacity building, and inclusive participation. The UN perspective contributes a normative anchor in international law and a forum for inclusive deliberation, and it occludes the power asymmetries that make consensus difficult and the gap between normative commitments and enforcement capacity.

2.3 KEY IDEA 2 — Lethal Autonomous Weapons Systems: The LAWS Debate, CCW, and the Principal-Agent Problem

Bolded Causal Claim: The central danger of lethal autonomous weapons systems is not simply that machines may kill without a human pressing a trigger, but that autonomy in the critical functions of target selection and engagement dissolves the principal-agent relationship on which legal accountability, moral responsibility, and strategic restraint depend.

Mechanism: A lethal autonomous weapons system is generally understood, in the UN Convention on Certain Conventional Weapons discussions, as a system that can select and engage targets without further human intervention after activation. The mechanism of concern is not automation in the narrow sense of a human choosing a target and a machine executing that choice; it is the transfer of the lethal decision itself to an algorithmic agent. In conventional armed conflict, the state is the principal and the human combatant is the agent: the human can be trained, disciplined, court-martialed, and held responsible under international humanitarian law and domestic law. A machine cannot be punished, cannot possess intent, and cannot be deterred by moral or legal sanction. The principal-agent chain therefore breaks at the point where the machine becomes the effective decision-maker. Responsibility then diffuses backward across designers, programmers, data providers, commanders, operators, and procurement officials, none of whom may have had sufficient knowledge, control, or intent to satisfy ordinary standards of accountability. This is the accountability gap: lethal force is exercised, but no agent is clearly answerable for its wrongful use.

The mechanism is intensified by the epistemic limits of machine perception in war. International humanitarian law requires distinction between combatants and civilians, proportionality between military advantage and civilian harm, and precautions in attack. A system may classify patterns in sensor data with high statistical accuracy in training conditions, but war is an adversarial, degraded, and novel environment. Civilian objects can resemble military ones; surrender, hors de combat status, and the presence of human shields require contextual judgment; proportionality is not a classification task but a normative balancing act. The claim that autonomy can improve compliance with international humanitarian law by reducing human fatigue, fear, or bias is therefore contestable: it assumes that the relevant failures are primarily human cognitive failures rather than structural uncertainties about intent, context, and value. The counterargument—that autonomous systems may be more precise and consistent than stressed humans—has real force in highly structured environments, but it does not resolve the problem of accountability when the system errs in an unstructured one.

The CCW debate has turned on whether the appropriate response is a legally binding prohibition, a regulatory framework, or non-binding principles. Some states and civil society actors have called for a ban on development, production, and use of fully autonomous lethal weapons systems, often framed around the requirement of meaningful human control. Others argue that existing international humanitarian law is sufficient, that autonomy exists on a spectrum rather than as a binary category, and that a ban would be unverifiable or would disadvantage states that comply while non-compliant actors proliferate. The concept of meaningful human control has become the focal normative standard, but it remains contested: it can mean a human in the loop, on the loop, or out of the loop with a veto; it can require comprehension, intention, and accountability, or merely the technical possibility of intervention. The principal-agent problem persists because nominal human control is not the same as meaningful control. A human who cannot understand the system’s reasoning, cannot intervene in time, or is structurally incentivized to defer to the machine is not exercising control in the legally and morally relevant sense.

Scope Conditions: The accountability gap is most acute where autonomy operates in the critical functions of target selection and engagement, where decision times are short, where environments are complex and adversarial, and where human oversight is nominal rather than substantive. It is also acute where kill chains are long, where software and data are proprietary, and where training data do not represent the operational environment. The argument is weaker where autonomy is confined to navigation, logistics, or non-lethal functions; where human operators retain genuine comprehension and intervention capacity; and where systems are tested, audited, and constrained by robust legal review. Even under those conditions, however, the structural problem remains: the machine cannot be the bearer of responsibility, and human responsibility must be deliberately constructed rather than assumed.

Observable Implications: If the principal-agent mechanism is operating, one should observe national policies that distinguish between automation and autonomy while often leaving the boundary ambiguous; legal reviews of new weapons under Article 36 of Additional Protocol I; procurement of systems with autonomous terminal guidance, swarming capabilities, or AI-assisted targeting; and diplomatic positions in the CCW that cluster around prohibition, regulation, or status-quo arguments. One should also observe a gap between declared human control and operational practice, especially in high-tempo domains where operators are likely to defer to machine recommendations. The absence of a binding international instrument despite years of discussion is itself an observable implication: the principal-agent problem is recognized, but states disagree about whether the solution is prohibition, regulation, or improved human-machine teaming.

Perspective Note (five traditions in brief):

African (primary). From an African vantage point, the LAWS debate is not merely a technical question about whether machines can comply with international humanitarian law; it is a question about who bears the costs of experimentation with lethal autonomy and who is excluded from the design of the rules. African scholarship on power and technology provides the interpretive frame. Rodney’s analysis of underdevelopment, Nkrumah’s critique of neo-colonialism, and Fanon’s account of technocratic violence suggest that military technologies are rarely neutral instruments; they are embedded in structures of domination that externalize risk onto populations with little voice in their design. Mbembe’s concept of necropolitics illuminates how certain lives are rendered expendable in the name of security, while Mamdani’s work on the bifurcated state highlights how legal and political orders can separate rulers from ruled in ways that make accountability formal rather than substantive. Grovogui’s critique of international law as a civilizing discourse and Bayart’s analysis of extraversion suggest that global governance forums may reproduce hierarchies even when they invoke universal norms. Oyěwùmí’s epistemic critique reminds us that categories such as “human control” may carry culturally specific assumptions about agency and responsibility. African states participating in CCW discussions have often stressed meaningful human control, capacity building, and the dangers of an unregulated market in autonomous systems. The African perspective reveals that LAWS governance is not only about machine compliance but about distributional justice: African populations may face autonomous weapons in asymmetric conflicts, peace operations, border control, or internal security contexts without having shaped the systems or the rules. It also reveals a technology gap: many African militaries lack the industrial base, data infrastructure, and legal-technical capacity to develop or audit LAWS, which can produce dependence on imported systems and external standards. The occlusion in this perspective is that it can understate the genuine security pressures that drive states toward autonomy, including force protection and the desire to reduce soldier casualties, and it may not fully specify how capacity building would be funded or governed.

Chinese. Chinese discourse frames lethal autonomous weapons within a broader AI governance agenda that combines sovereignty, development rights, and strategic stability. The Chinese position has generally supported international discussion while resisting constraints that would lock in Western military advantage or restrict legitimate applications of AI. It emphasizes that any framework must preserve space for peaceful and defensive uses, respect state sovereignty, and avoid politicized or discriminatory standards. Chinese scholarship and official statements often treat the “AI arms race” as a risk to be managed through governance rather than as an inevitable competition, while also insisting that China’s AI development is a sovereign right. The perspective contributes a demand that governance not become a vehicle for technological containment, and it occludes the possibility that Chinese military modernization itself contributes to competitive dynamics.

Russian. Russian discourse treats autonomous weapons as part of military modernization and strategic stability management. It has tended to argue that existing international humanitarian law is sufficient, that a legally binding ban would be difficult to verify, and that states should retain flexibility in developing military AI. Russian official rhetoric often emphasizes human control in declaratory policy while investing in automation and autonomy for force protection and battlefield advantage. The perspective contributes a realist emphasis on verification and strategic stability, and it occludes the accountability and civilian-harm concerns that motivate prohibition advocates.

Western. Western, especially US and allied, discourse centers on AI leadership, the LAWS debate, and an emerging governance self-critique. The United States has maintained that autonomous weapons systems should be designed to allow commanders and operators to exercise appropriate levels of human judgment over the use of force, while opposing a blanket ban. Allies have varied, but many emphasize human control, IHL compliance, and the potential benefits of precision. The governance self-critique acknowledges risks of accountability gaps, algorithmic bias, and civilian harm, and has produced policy reviews, ethical principles, and legal standards. The perspective contributes operational realism and a focus on human-machine teaming, and it occludes the global distributional consequences of Western-led AI military development.

UN. The UN perspective frames LAWS through international humanitarian law, human dignity, and multilateral governance. The Secretary-General has called for a ban on lethal autonomous weapons systems that function without human control, and the CCW Group of Governmental Experts has debated definitions, characteristics, and possible regulatory frameworks for years without reaching consensus. The Pact for the Future includes commitments on AI governance that emphasize human rights, capacity building, and inclusive participation. The UN perspective contributes a normative anchor in international law and a forum for inclusive deliberation, and it occludes the power asymmetries that make consensus difficult and the gap between normative commitments and enforcement capacity.

2.4 KEY IDEA 3 — Decision Compression and Hyperwar: Algorithmic Escalation and Flash Wars

Bolded Causal Claim: As AI compresses the observe-orient-decide-act cycle below the threshold of human deliberation, military competition shifts from attrition to tempo, creating structural incentives for pre-delegation, automation bias, and preemptive action that can convert localized incidents into rapid, irreversible escalation.

Mechanism: The mechanism begins with the OODA loop—observe, orient, decide, act—which has long been a heuristic for military decision-making. AI-enabled systems compress each stage: sensors and data fusion accelerate observation; machine learning and automated analysis accelerate orientation; decision-support systems and autonomous agents accelerate decision; and networked weapons accelerate action. The result is not merely faster warfare but a change in the kind of warfare: when the tempo of interaction exceeds human cognitive and institutional capacity, the advantage shifts to the side that can delegate more authority to machines. This creates a structural incentive to pre-delegate lethal or escalatory decisions, to accept higher automation, and to reduce human veto points. The mechanism is not that machines are malevolent; it is that organizations optimize for speed under competitive pressure, and speed rewards delegation.

Decision compression interacts with automation bias, the tendency of human operators to defer to machine recommendations, especially under time pressure and uncertainty. A human in the loop who cannot independently verify a machine’s recommendation within the available time is not a meaningful check; they are a liability shield. The mechanism also involves tight coupling: as military systems become networked and interdependent, a local error can propagate across domains before human institutions can intervene. The analogy to financial flash crashes is instructive: algorithmic trading can produce rapid, nonlinear price movements that no individual trader intended, and circuit breakers are required to halt cascades. In military affairs, there are no agreed circuit breakers, and the stakes include escalation across conventional, cyber, space, and nuclear domains.

Hyperwar, as the term is used in discussions of AI-enabled conflict, denotes warfare in which the speed and complexity of machine interaction outpace human comprehension and control. The risk is not only accidental escalation but also strategic instability: if adversaries believe that the other side can act faster, they may adopt launch-on-warning postures, pre-delegate authority, or strike first to avoid being struck. Crisis stability erodes because the window for signaling, bargaining, and de-escalation shrinks. Kissinger and colleagues have argued that AI challenges traditional concepts of deterrence and strategic stability by introducing non-human agency into the most consequential decisions. Scharre’s analysis of autonomy emphasizes that the danger lies less in rogue machines than in the interaction between human organizations and autonomous systems under pressure. Horowitz’s work on military innovation diffusion suggests that the adoption of AI will be shaped by organizational culture, resource constraints, and strategic competition, not by technology alone.

The counterargument is that AI can also improve situational awareness, reduce human error, and create off-ramps by clarifying ambiguous situations faster than humans can. Better early warning, more accurate target identification, and automated deconfliction could in principle reduce accidental escalation. This is plausible in some domains, but it depends on calibration, data quality, and adversarial robustness. In a competitive environment, the side that uses AI for speed may force the other side to do the same, producing a race to the bottom in human control. The mechanism therefore predicts not universal catastrophe but a structural bias toward tempo over deliberation, with the severity of risk depending on domain, organizational incentives, and the presence or absence of shared norms.

Scope Conditions: The risk of flash wars is highest in domains where decision times are short and consequences are large: air and missile defense, cyber operations, space systems, and any process adjacent to nuclear command and control. It is also high where systems are tightly coupled, where adversaries interact repeatedly under uncertainty, and where organizational incentives reward speed over caution. The argument is weaker where AI is used primarily for logistics, intelligence analysis with human review, or long-cycle planning; where human-machine interfaces are well designed; and where there are robust testing, red-teaming, and escalation-management procedures. Even in those cases, however, the competitive pressure to compress decision cycles may erode safeguards over time.

Observable Implications: If decision compression is driving military change, one should observe doctrine and procurement emphasizing speed, AI-enabled command-and-control systems, and human-machine teaming in high-tempo domains. One should observe exercises and simulations that test machine-speed escalation, as well as proposals for “speed limits,” human control requirements, or crisis-communication mechanisms. Historical analogues of automation bias—such as the 2003 Patriot fratricide incidents, in which automated systems misidentified friendly aircraft—illustrate how human operators can defer to machine classifications under pressure. The 2010 flash crash in financial markets illustrates how algorithmic interaction can produce rapid cascades without malicious intent. In the military domain, observable implications include near-misses, false positives in air defense, and debates over whether AI should be permitted in nuclear command and control. The absence of agreed international norms for machine-speed escalation is itself an implication: states are building faster systems before they have built shared brakes.

Perspective Note (five traditions in brief):

African (primary). From an African vantage point, hyperwar is not an abstract future risk but an extension of existing asymmetries in who controls the tempo of violence. African scholarship on state capacity, extraversion, and the political economy of technology suggests that African states are more likely to be recipients, testing grounds, or collateral arenas of AI-enabled warfare than architects of it. Rodney’s analysis of technological dependence, Bayart’s concept of extraversion, and Mamdani’s account of the bifurcated state imply that imported military technologies can strengthen coercive capacity without corresponding accountability. Mbembe’s necropolitics highlights how populations in weakly governed spaces may be subjected to lethal violence with little recourse, while Fanon’s critique of colonial violence reminds us that speed and technological superiority have historically been used to overwhelm political resistance. The African perspective reveals that decision compression is not only a problem among great powers; it can accelerate violence in contexts where oversight is weak, data are poor, and accountability mechanisms are fragile. It also reveals a technology gap: African militaries and regulators may lack the capacity to audit, test, or govern AI-enabled systems, making them dependent on external vendors and standards. African diplomatic and scholarly voices have tended to demand equitable governance, transparency, capacity building, and meaningful human control, arguing that the benefits of AI in security should not accrue only to technologically advanced states. The occlusion in this perspective is that it can understate the genuine security dilemmas faced by African states confronting insurgencies, terrorism, and cross-border threats, and it may not fully resolve how human control can be operationalized in high-tempo defensive systems.

Chinese. Chinese discourse frames decision compression within strategic stability and AI governance debates. It expresses concern about an AI arms race and about the destabilizing effects of machine-speed warfare, while also pursuing military modernization and “intelligentized” capabilities. Chinese official and scholarly voices often emphasize sovereignty, the right to develop AI, and the need for international rules that do not constrain legitimate defense. The perspective contributes a demand for stability-oriented governance and an acknowledgment that tempo competition can be dangerous, and it occludes the possibility that Chinese modernization contributes to the very compression it warns against.

Russian. Russian discourse treats AI-enabled speed as part of military modernization and strategic stability management. It has emphasized the risks of destabilization while also investing in automation, electronic warfare, and AI-assisted command systems. Russian strategic stability framing often focuses on the interaction between conventional precision, cyber, and nuclear domains, warning that compressed decision times can undermine deterrence. The perspective contributes a realist focus on escalation management, and it occludes the accountability and civilian-harm dimensions that dominate Western and UN debates.

Western. Western discourse centers on AI leadership, the LAWS debate, and a governance self-critique that increasingly includes decision compression. US and allied militaries emphasize human-machine teaming, speed as a warfighting advantage, and the need to maintain decision superiority. At the same time, Western scholars and policymakers have warned about automation bias, flash escalation, and the erosion of human judgment. The perspective contributes operational and technical sophistication, and it occludes the global distributional consequences of tempo competition and the possibility that Western acceleration drives others to automate faster.

UN. The UN perspective frames decision compression through international humanitarian law, human dignity, and multilateral governance. The Secretary-General’s calls for human control over lethal force and the CCW’s debates over LAWS are relevant because compressed decision times make human control harder to realize. The Pact for the Future’s AI commitments emphasize human rights, capacity building, and inclusive participation, which imply that speed must not outrun accountability. The UN perspective contributes a normative anchor and a forum for deliberation, and it occludes the power asymmetries and enforcement gaps that make shared brakes difficult to establish.

2.5 KEY IDEA 4 — AGI/ASI: The Security Dilemma of Superintelligence and Alignment as a Global Security Problem

Bolded Causal Claim: The prospect of artificial general intelligence (AGI) and artificial superintelligence (ASI) transforms AI from a military-technology competition into a security dilemma over existential risk, because states and firms racing to develop systems whose capabilities and alignment properties are uncertain may each treat the other’s progress as a threat, while the technical problem of ensuring that a vastly more capable system pursues human-intended goals becomes a collective-security problem that no single state can solve alone.

Mechanism: The mechanism begins with the security-dilemma logic familiar from international-relations theory: in an anarchic system, investments made for defensive or competitive purposes are difficult to distinguish from preparations for dominance, so rivals respond with counter-investments, secrecy, and worst-case planning. Applied to AI, this means that frontier compute, data, talent, and model weights become strategic assets whose accumulation by one actor is read by others as a potential shift in the balance of power. The race dynamic is intensified by the possibility of large first-mover advantages. If AGI or ASI can automate research, production, or military planning, the first actor to achieve a decisive capability advantage may gain compounding benefits, creating incentives to move quickly, to conceal progress, and to underinvest in safety. This is the “race to the bottom” concern in AI safety: the alignment tax, or the cost of making a system safe, robust, and controllable, may be treated as a competitive disadvantage when rivals are moving fast. The alignment problem itself is not merely a software bug. It includes specification problems (defining the right objective), robustness problems (performance under distribution shift), interpretability problems (understanding what the system is doing), and control problems (ensuring that a more capable system remains corrigible). As capability scales, small misspecifications can produce large harms, and the gap between what a system optimizes and what humans value may widen. The global-security dimension arises because ASI externalities do not respect borders. A misaligned or misused superintelligent system could affect critical infrastructure, financial systems, biological research, or military command networks across states. No single sovereign can contain such a system through territorial control alone. Therefore, alignment becomes a global public good, while the incentives to produce it remain national or corporate. Bostrom (2014) frames superintelligence as a strategic challenge in which the transition to machine intelligence may be abrupt enough that institutions have little time to adapt, and in which the default outcome may not be benign. Kissinger et al. (2021) extend this concern to world order, arguing that AI can alter strategic stability by changing the speed, opacity, and scale of decision-making. The mechanism is therefore a three-layer interaction: competitive racing produces secrecy and acceleration; acceleration increases the risk of deploying misaligned systems; and the transnational nature of the risk makes unilateral safety insufficient.

Scope Conditions: This key idea applies most strongly under a set of demanding assumptions: that AGI or ASI is technically feasible within a strategically relevant horizon; that capability growth may be rapid or discontinuous; that alignment remains unsolved at scale; that actors face strong competitive incentives; and that international institutions lack the authority or capacity to monitor frontier development. It applies less strongly if AGI proves impossible, if progress plateaus, if alignment turns out to be comparatively easy, or if compute governance becomes effective enough to slow racing. It also applies unevenly across actors: frontier labs and great powers are the central cases, while most states are consumers rather than producers of frontier models. The idea is about tail risk and structural incentives, not a claim that superintelligence is imminent. Much of the policy relevance comes from the possibility of capability overhang, in which an algorithmic or hardware breakthrough suddenly makes a much more capable system deployable before governance has adapted. Under those conditions, ex post regulation is likely to be too slow.

Observable Implications: If this mechanism is operating, one should observe frontier laboratories publishing “responsible scaling” policies, dangerous-capability evaluations, and red-teaming reports; governments creating AI Safety Institutes and funding alignment research; export controls on advanced chips and cloud compute; debates over open-weight models; calls for pauses or moratoria on giant experiments; talent restrictions; and summit diplomacy on frontier AI safety. One should also observe military doctrines emphasizing human control, automation bias, and the risks of over-reliance on AI in command and control. The absence of these measures would suggest either that actors do not take the risk seriously or that competitive pressures are suppressing safety investment. Conversely, the presence of safety measures without binding enforcement would suggest that governance is performative rather than constraining.

Perspective Note (five traditions in brief): African (primary). The African perspective begins from the technology gap and the AI divide, not from the frontier-lab race. African scholars such as Grovogui and Mamdani have long argued that international orders encode hierarchical premises and that institutions often speak in universal language while distributing power unevenly. From this vantage point, AGI discourse can appear as a conversation among a small set of powerful actors about risks that are distant for many African societies, while more immediate harms—surveillance, biased systems, data extraction, labor displacement, and infrastructural dependence—are already present. The African perspective therefore demands equitable governance, data sovereignty, capacity building, and meaningful participation in standard-setting. It warns that governance frameworks negotiated now may lock in asymmetries before African states have the technical capacity to shape them. At the same time, it does not dismiss existential risk; rather, it insists that superintelligence debates should not crowd out present development and rights, and that global security must include the security of those who are usually absent from the table. Chinese. The Chinese perspective frames AI as a strategic technology central to national rejuvenation and sovereignty. It emphasizes development rights, state oversight, and opposition to hegemonic control, particularly export controls that it regards as containment. Alignment is treated as a technical and safety issue, but not at the expense of development or sovereignty. Governance is preferred through the United Nations and inclusive multilateralism, while domestic regulation emphasizes social stability, content control, and security. The “AI arms race” is a concern, but so is the risk that safety discourse becomes a tool for preserving Western technological leadership. Russian. The Russian perspective emphasizes strategic stability, military modernization, and sovereignty. ASI is a distant prospect, but current AI is already relevant to strategic parity, nuclear command-and-control resilience, and great-power competition. Russia tends to be skeptical of Western AI safety framing and wary of governance mechanisms that could constrain its military or technological development. Governance is acceptable when it preserves strategic stability and sovereign equality, but not when it functions as a Western-led constraint. Western. The Western perspective combines AI leadership, safety research, and governance self-critique. It is shaped by Bostrom’s existential-risk framing, by frontier-lab safety culture, and by strategic competition with China. It supports alignment research, export controls, safety institutes, and multilateral declarations, while also debating industry capture, enforcement gaps, and the risks of over-regulation. The Western view is not unified: it contains both accelerationist and safety-first strands, and its governance proposals often reflect the interests of large technology firms. UN. The UN perspective frames AGI/ASI through human rights, human dignity, and inclusive global governance. The Secretary-General’s advisory processes and the Pact for the Future emphasize capacity building, participation, and the need for a global framework. The UN contributes a normative anchor and a forum for deliberation, but it faces the familiar gaps of enforcement, funding, and power asymmetry. It occludes less than the great-power perspectives, but it cannot by itself resolve the competitive incentives that drive racing.

2.6 KEY IDEA 5 — AI Governance: Fragmented Regimes and Techno-Geopolitical Competition

Bolded Causal Claim: AI governance is emerging as a fragmented, competitive regime in which states use regulation, industrial policy, and security controls to shape AI development according to divergent values and power interests, producing partial convergence on risk management but persistent divergence on sovereignty, innovation, and security, with the result that governance itself becomes an arena of techno-geopolitical competition rather than a neutral solution to it.

Mechanism: The mechanism is regulatory competition under anarchy. The European Union’s AI Act, adopted in 2024, uses a risk-based approach: it prohibits certain practices, imposes obligations on high-risk systems, and creates rules for general-purpose AI models, with enforcement through a centralized AI Office and national authorities. Its logic is rights-based and market-shaping, and it relies on the “Brussels effect,” in which extraterritorial market size induces global compliance. The United States has taken a more sectoral and security-oriented approach. Executive Order 14110, signed in October 2023, used existing authorities to require reporting on powerful models, promote red-teaming, and direct agencies to develop standards, while separate export-control measures targeted advanced semiconductors and certain AI capabilities. The US approach combines voluntary commitments, procurement power, and security controls rather than a comprehensive statute. China’s approach is state-led and content- and security-focused. Its 2022 provisions on algorithmic recommendation, 2023 rules on deep synthesis, and 2023 interim measures for generative AI services require security assessments, content compliance, data governance, and alignment with socialist core values, while also promoting industrial development. These three models differ in their premises: the EU emphasizes rights and risk; the US emphasizes innovation and security; China emphasizes sovereignty, social stability, and state capacity. The mechanism of fragmentation is that each model reflects domestic political economy, legal tradition, and strategic interest. Firms respond through compliance programs, forum shopping, and regulatory arbitrage. Standards bodies become battlegrounds. Military and intelligence applications are often exempted or governed separately, creating a large gap between civilian AI governance and military AI use. At the same time, there is partial convergence: red-teaming, model evaluations, transparency, watermarking, incident reporting, and safety institutes are becoming common reference points. The United Nations provides a multilateral layer through the Secretary-General’s AI advisory processes, the Global Digital Compact and Pact for the Future adopted in 2024, and the Convention on Certain Conventional Weapons discussions on lethal autonomous weapons systems. But UN processes move slowly and depend on member-state consensus, while the most capable actors prefer flexible or bilateral arrangements. Governance is therefore also industrial policy: subsidies, semiconductor controls, talent visas, public research, and “friend-shoring” shape who can build AI, not merely how it is used.

Scope Conditions: This key idea applies where AI has cross-border effects, strategic value, and concentrated industrial capacity. It applies less to low-capability systems or purely domestic applications with limited externalities. Its explanatory power depends on enforcement capacity, technical expertise, and geopolitical trust. Where enforcement is weak, rules may be symbolic. Where technical capacity is low, states may adopt rules they cannot implement. Where trust is high, convergence is more likely; where trust is low, governance becomes another arena of competition. The idea also assumes that AI is not merely a normal technology but a general-purpose capability with military and economic implications, which makes states reluctant to delegate authority to neutral regulators.

Observable Implications: If this mechanism is operating, one should observe compliance programs, model cards, watermarking, incident databases, national AI strategies, procurement rules, litigation over bias or copyright, standards such as ISO/IEC 42001, compute tracking, export-control lists, licensing debates, and the creation of AI Safety Institutes. One should also observe regulatory arbitrage, such as firms locating development or deployment in jurisdictions with lighter rules; tensions over open-source models; and disputes over whether military AI should be governed by civilian AI law. The coexistence of convergence on safety vocabulary with divergence on enforcement and scope would be a signature of fragmented governance.

Perspective Note (five traditions in brief): African (primary). The African perspective treats AI governance as a question of equitable participation and structural inequality. African scholars such as Grovogui, Mbembe, and Mamdani have emphasized that global orders often reproduce hierarchies through ostensibly neutral institutions, and that the language of universalism can mask unequal power. From this vantage point, the risk is not only unsafe AI but rule-taking: African states may be expected to adopt standards designed elsewhere, while lacking the capacity to enforce them or to shape them. The demand is for equitable governance, data sovereignty, capacity building, and meaningful voice in standards and institutions. The African perspective also insists that governance address present harms—bias, surveillance, labor exploitation, and infrastructural dependence—not only frontier risks. It sees the technology gap as a governance problem, because rules written by and for advanced economies may not fit African contexts. Chinese. The Chinese perspective emphasizes sovereignty, development, and opposition to extraterritorial control. It supports governance through the United Nations and inclusive multilateralism, while insisting that states retain the right to regulate AI according to their own social and security needs. Domestic regulation focuses on content, data, and social stability. China is concerned about the “AI arms race” and US-led export controls, but also seeks to shape global standards and promote its own governance model. It views governance as compatible with industrial policy and state capacity. Russian. The Russian perspective emphasizes strategic stability, sovereignty, and military modernization. It is skeptical of Western-led governance frameworks and wary of rules that could constrain Russian development or military applications. Governance is acceptable when it preserves sovereign equality and strategic stability, but not when it functions as a constraint imposed by technologically dominant states. Russia’s interest lies in preventing governance from becoming a tool of containment. Western. The Western perspective combines AI leadership, risk management, and governance self-critique. The EU emphasizes rights and risk; the US emphasizes innovation, security, and voluntary standards. Western actors support safety institutes, export controls, and multilateral declarations, while debating industry capture, enforcement gaps, open-source tensions, and the risk of over-regulation. The Western view is internally divided between acceleration and caution, and its governance proposals often reflect the influence of large technology firms. UN. The UN perspective frames AI governance through human rights, inclusive participation, and multilateralism. The Secretary-General’s advisory body, the Global Digital Compact, and the Pact for the Future emphasize capacity building, human rights, and a global framework, while the CCW process addresses lethal autonomous weapons. The UN contributes legitimacy and a forum for deliberation, but its effectiveness is limited by consensus requirements, funding gaps, and the reluctance of major powers to submit to binding constraints. It occludes less than great-power perspectives, but it cannot by itself overcome competitive incentives.

2.7 FIVE PERSPECTIVES IN DEPTH

2.7.1 African Perspective (PRIMARY)

The African vantage point on AI, autonomous weapons, and hyperwar begins from a different question than the great-power debate. Where Western, Chinese, and Russian discussions often ask how AI alters the balance of power among leading states, much African scholarship asks how technological systems reproduce or transform structures of domination that predate the current AI boom. The relevant lineage is not primarily the security dilemma or arms-control tradition but the political economy of underdevelopment, coloniality, and epistemic subordination. Rodney’s analysis of how Europe underdeveloped Africa, Nkrumah’s account of neo-colonialism, and Fanon’s critique of colonial violence provide the historical grammar: contemporary AI infrastructures can be read as new instruments through which value, data, and decision-making capacity are extracted from African societies while strategic control remains elsewhere. Mbembe’s work on necropolitics and the postcolony sharpens the point: technologies of classification and targeting do not merely optimize force; they can sort populations into categories of expendability, especially where accountability is weak and sovereignty is fragmented. Bayart’s notion of extraversion helps explain why African elites may embrace external AI systems for regime security or donor alignment rather than continental autonomy. Mamdani’s analysis of the bifurcated state suggests that AI governance in Africa must contend with legal and political orders split between civic and customary domains, so that “human control” over weapons or algorithms cannot be assumed to mean accountable public control. Grovogui’s argument that Africans have often been treated as quasi-sovereigns in international order is directly relevant: in AI governance, African states may be invited to participate without possessing the compute, data, or institutional capacity to shape outcomes. Cabral’s insistence that liberation requires cultural and epistemic self-determination, Nyerere’s emphasis on collective self-reliance, Oyěwùmí’s critique of imported categories of knowledge, and Tamale’s decolonial feminist scholarship all push the analysis beyond “access” toward epistemic justice: whose concepts of risk, harm, and security define the global AI agenda?

On narrow AI in military systems, this perspective highlights asymmetry and dependency. African militaries and security services increasingly encounter AI-enabled ISR, targeting, logistics, and command systems through foreign suppliers, donor programs, and commercial platforms. The African contribution is to show that the diffusion of narrow AI is not a neutral modernization process but a reconfiguration of dependency. Imported systems may improve situational awareness, but they also create opaque supply chains, vendor lock-in, and gaps between technical capability and legal accountability. African security scholars such as Aning and Danso have long emphasized that security in Africa is inseparable from governance, human security, and the legitimacy of the state; applied to AI, this means that a targeting algorithm or surveillance platform is not merely a capability but a political intervention into already fragile civil-military relations. The perspective also notes that African states may use AI for border management, counterterrorism, or resource protection, and that these uses can be coercive. It therefore refuses a simple narrative of African victimhood while insisting that procurement, oversight, and data governance are the decisive variables.

On lethal autonomous weapons, the African vantage point foregrounds the principal-agent problem in its colonial and postcolonial form. The question is not only whether a machine can distinguish combatants from civilians under international humanitarian law, but who designs the categories, who owns the data, who profits from deployment, and who is answerable when harm occurs. Autonomous weapons lower the political cost of force for technologically advanced actors, which matters acutely for regions that have historically experienced external intervention. African delegations in multilateral forums have often stressed humanitarian protection, accountability, and the asymmetry between those who build autonomous systems and those who live with their consequences. The perspective contributes a demand that “meaningful human control” be understood not as a procedural checkbox but as democratic, locally accountable control over the entire life cycle of a weapons system, including data collection, model training, and post-strike review. It occludes less than great-power perspectives on the human cost of remote warfare, but it may understate the genuine security dilemmas faced by African states confronting insurgencies or cross-border threats, and it can sometimes treat “Africa” as more unified than it is.

On hyperwar and decision compression, African analysis emphasizes collateralization. Flash wars driven by algorithmic escalation among major powers would not be fought in Africa, but their effects—disrupted shipping, food and energy price shocks, refugee flows, and the diversion of diplomatic attention—would fall heavily on African societies. The mechanism is not direct battlefield exposure but systemic vulnerability: African economies are often price-takers in global commodity and food markets, and African security architectures depend on external funding and attention. Bayart’s extraversion is relevant here: African elites may be pulled into great-power technology blocs, accepting surveillance or military AI partnerships in exchange for security assistance, thereby narrowing continental strategic autonomy. The African perspective thus reveals that hyperwar is not only a risk of inadvertent escalation among nuclear-armed states; it is also a structure of global inequality in which the costs of algorithmic speed are distributed unevenly. What it may occlude is the possibility that some African actors could benefit from rapid AI-enabled intelligence or logistics in regional crises, and that not all external technology partnerships are purely extractive.

On AGI and superintelligence, the African vantage point is skeptical of the framing that dominates much Western and Chinese discourse. The existential-risk narrative, however important, can crowd out present and near-term harms: data extraction, labor exploitation in content moderation and annotation, algorithmic discrimination, and the concentration of compute in a handful of firms and states. African scholarship on decoloniality and epistemic justice suggests that “alignment” is not a purely technical problem of making a superintelligent system follow human values; it is a political problem of whose values, whose languages, whose histories, and whose futures are encoded. Mbembe’s necropolitics and Fanon’s analysis of colonial violence imply that systems optimized for efficiency may reproduce racialized and colonial hierarchies unless their design is accountable to those historically excluded. The perspective contributes an insistence that AGI governance include development, climate resilience, health, and food security as core security goods, not as peripheral concerns. It occludes, or at least underweights, the possibility that advanced AI could accelerate scientific discovery or public-service delivery in Africa, and it may not fully engage with the technical specifics of alignment research. It also risks overextending colonial analogies if every AI system is treated as a repetition of empire rather than a contingent political arrangement.

On AI governance, the African position is best understood as a demand for equitable governance rather than rejection of technology. The central claims are data sovereignty, capacity building, technology transfer, inclusive standard-setting, and human-rights-based regulation. African scholars and policymakers have increasingly framed AI as an issue of digital colonialism: the extraction of data and minerals, the location of cloud infrastructure, the dominance of foreign platforms, and the absence of African languages and contexts in training data. Oyěwùmí and Tamale’s epistemological critiques imply that governance frameworks built on Western categories may misrecognize African social realities, including gender, kinship, and community. Mamdani’s bifurcated state and Grovogui’s quasi-sovereignty imply that governance must strengthen African institutional capacity, not merely invite African signatures onto externally drafted rules. The perspective contributes a corrective to great-power AI governance: it shows that legitimacy depends on who can participate meaningfully, not only on who convenes the meeting. It also highlights that “safety” and “security” are not identical; a system can be safe for its operators and dangerous for populations subjected to it. What it occludes is the diversity of African state interests—some governments may prioritize surveillance and control over rights—and the fact that continental coordination is itself constrained by resource gaps and competing regional agendas. It may also understate the genuine security benefits that some African states derive from partnerships with external AI providers.

Finally, heterogeneity matters. African perspectives on AI are not uniform: coastal and landlocked states, resource-rich and resource-poor economies, and different regime types experience AI differently. Some governments may prioritize surveillance, border control, or revenue extraction; others may prioritize health, agriculture, or education. The African vantage point is therefore not a single doctrine but a family of critiques united by attention to asymmetry, extraction, and epistemic justice. It demands that global AI governance be evaluated by its effects on those with the least power to shape it, while remaining accountable to internal African debates about democracy, development, and security.

Taken together, the African perspective’s distinctive contribution is to reframe AI, AGI, and hyperwar as questions of global justice, epistemic sovereignty, and the political economy of technology. It reveals what the other four perspectives often miss: that the AI divide is not a temporary lag but a structured relation of extraction; that autonomous weapons and algorithmic targeting are continuous with longer histories of remote and racialized violence; that hyperwar costs are globally unequal; and that AGI alignment is inseparable from whose knowledge counts. Its occlusions are real: it can understate African agency, operational military detail, and intra-African variation, and it may not always specify how equitable governance would be enforced against powerful states and firms. But as a primary lens, it prevents the lesson from collapsing into a great-power arms-race narrative and insists that the future of AI security be judged by its effects on those least able to shape it.

2.7.2 Chinese Perspective

The Chinese perspective treats AI as a strategic technology at the intersection of national rejuvenation, economic development, and military modernization. From within, AI is not merely a weapons issue; it is a general-purpose capability that will shape industrial competitiveness, social governance, and the future distribution of global power. The concept of “intelligentized warfare” signals that the People’s Liberation Army sees AI as central to command, control, intelligence, and decision support, while military-civil fusion reflects a deliberate effort to dissolve boundaries between civilian innovation and defense application. This perspective contributes an understanding of AI as a whole-of-society project rather than a narrow military procurement category. It also challenges the assumption that the United States and its allies naturally define the legitimate uses of AI.

On governance, China emphasizes sovereignty, development rights, and non-interference. Beijing’s international proposals tend to frame AI governance as a multilateral project in which all states have equal rights to develop and use AI, and in which unilateral export controls or technology restrictions are themselves security risks. Domestic regulations on recommendation algorithms, deep synthesis, and generative AI reflect a state-led model that prioritizes content security, social stability, and national security alongside innovation. The Chinese contribution is to insist that AI governance cannot be separated from development inequality: rules written by advanced economies may freeze technological hierarchies and deny the Global South the benefits of AI. This resonates with African demands for equitable access, even though the two perspectives differ on human rights and political openness.

On the AI arms race, China often presents itself as responding to US initiatives rather than initiating competition. It warns against an unconstrained race in military AI and supports discussions on responsible AI, but it resists constraints that would lock in US advantages or infringe on sovereign defense choices. In the CCW context, Chinese positions generally emphasize compliance with international humanitarian law and the need for consensus, while avoiding commitments that would prohibit systems China may wish to develop. The perspective contributes a reminder that governance proposals are themselves strategic instruments: definitions of “autonomous,” “human control,” and “responsible use” can advantage some actors over others.

On hyperwar and decision compression, China is concerned about US concepts of mosaic warfare, algorithmic command and control, and the acceleration of the kill chain. Chinese analysts worry that AI-enabled decision support could compress crisis timelines and create first-mover advantages, undermining strategic stability. The response is not only technological but doctrinal: crisis management, escalation control, and the preservation of retaliatory capability. This contributes a stability-focused lens that complements Western discussions of deterrence and African concerns about collateral damage.

On AGI, China tends to treat superintelligence as a long-term strategic horizon rather than an immediate governance emergency. The emphasis is on maintaining national capability, preventing technological dependence, and shaping global rules before others do. Alignment is framed partly as a technical challenge and partly as a question of whose values govern global AI systems. The perspective’s occlusions are significant: opacity around military AI programs, the integration of AI into surveillance and social control, and a tendency to use sovereignty claims to resist external scrutiny. It may also understate the risks of rapid military AI deployment in favor of competitive positioning. Nevertheless, the Chinese view correctly identifies that AI governance is inseparable from development, industrial policy, and the politics of technological hierarchy.

2.7.3 Russian Perspective

The Russian perspective frames AI primarily through military modernization and strategic stability. Moscow views AI as part of a broader military-technical revolution that can offset conventional inferiority, enhance precision and intelligence, and sustain Russia’s status as a major military power. Russian discourse emphasizes the modernization of armed forces, the integration of AI into command and control, and the preservation of a credible deterrent. This is not simply an arms-race mentality; it is shaped by a strategic culture that sees technological lag as an existential vulnerability and by a geopolitical position in which Russia perceives itself as encircled by NATO and US technological superiority.

On narrow AI, Russia’s interest lies in ISR, targeting, logistics, and decision support that can compensate for numerical or industrial disadvantages. The contribution of this perspective is to highlight how AI can be used by a state that is not a technological leader to sustain strategic parity. It also underscores that military AI is not only about autonomous weapons; it is about the speed and quality of decision-making in a contested environment. The occlusion is that Russian programs are opaque, and public discussion often mixes genuine modernization with propaganda. There is limited transparency about testing, safety, or human control, which makes external assessment difficult.

On LAWS, Russia has generally resisted legally binding prohibitions that would constrain its options, while supporting discussions on responsible use and compliance with international humanitarian law. The Russian argument is that definitions of autonomy are politically contested and that premature bans could disadvantage states that are not already dominant in AI. This perspective contributes a skeptical view of humanitarian disarmament processes: it asks whether such processes are neutral or whether they encode the preferences of technologically advanced states. Its weakness is that it can treat all constraints as containment, thereby avoiding accountability for how autonomous or semi-autonomous systems might be used.

On hyperwar, Russia is particularly concerned with strategic stability and escalation management. Russian analysts have long worried about US precision-strike capabilities, missile defense, and now AI-enabled decision compression. The fear is that faster kill chains and algorithmic recommendations could create incentives for preemption or could blur the line between conventional and nuclear escalation. Russia’s contribution is to keep the nuclear and strategic-stability dimension central: AI is not just a tactical issue but a factor in crisis stability. This is especially important for the lesson’s discussion of AGI and NC3, even if Russia’s own modernization contributes to the same risks.

On AGI, Russia tends to treat superintelligence as a distant but strategically significant horizon. The emphasis is on not falling behind and on ensuring that any global governance framework does not cement US or Chinese dominance. Alignment is less prominent in Russian discourse than sovereignty and parity. The perspective occludes the ethical and human-rights dimensions of AI governance and may overstate the coherence of Russian technological strategy. It also underweights the possibility that cooperation on AI safety could serve Russian interests by reducing catastrophic risk.

On governance, Russia favors state-centric, sovereignty-preserving frameworks and is skeptical of Western-led normative initiatives. It supports multilateral discussion but resists binding constraints that would limit military applications or domestic control. The contribution is a reminder that AI governance is inseparable from geopolitical competition and that major powers will not accept rules they perceive as disadvantaging. The occlusion is that this stance can block even modest transparency or risk-reduction measures and can conflate legitimate sovereignty concerns with resistance to accountability.

2.7.4 Western Perspective

The Western perspective is the most internally divided of the five. It combines a strong commitment to AI leadership, a vibrant LAWS and ethics debate, and a growing self-critique of governance failures. From within, AI is simultaneously an engine of economic growth, a military necessity, and a source of risk requiring regulation. The United States and its allies see AI as central to deterrence, intelligence, and operational advantage, while also recognizing that unregulated AI could undermine democratic values, human rights, and strategic stability. This duality is the perspective’s strength and its tension.

On narrow AI, Western militaries have been the primary innovators and exporters of AI-enabled ISR, targeting, logistics, and C2. The contribution is a rich body of doctrine, experimentation, and ethical review, including debates over human-machine teaming and the appropriate level of autonomy. Western analysts such as Scharre and Horowitz have helped define the policy agenda by distinguishing between different levels of autonomy and by emphasizing that the key issue is not full autonomy but the compression of human judgment. The occlusion is that Western innovation often assumes a technological and industrial base unavailable to others, and that “responsible AI” frameworks can become marketing language for systems whose effects are felt elsewhere.

On LAWS, the Western perspective contains a genuine debate. Humanitarian and legal scholars, along with some military ethicists, argue for meaningful human control and, in some cases, for bans or strict limits on lethal autonomous weapons. Military practitioners and strategists often counter that autonomy can reduce civilian harm, improve precision, and preserve force protection. The CCW process reflects this division: Western states have participated actively but have not converged on a binding prohibition. The contribution is a sophisticated legal and ethical vocabulary; the occlusion is that Western powers may exempt their own systems from the strictest interpretations while demanding constraints on adversaries.

On hyperwar, Western analysis has been at the forefront of identifying decision compression as a strategic risk. Kissinger and colleagues have warned that AI could destabilize deterrence by accelerating decision cycles and by introducing opacity into command and control. The concept of hyperwar captures the possibility that machine-speed interactions could outpace human political control. Western contributions include scenarios, wargames, and policy proposals for human oversight, crisis communication, and AI safety. The occlusion is that the same states developing these warnings are also investing heavily in the systems that create the risk, and that alliance politics can amplify competitive pressures.

On AGI, the Western perspective is split between acceleration and caution. Some actors treat advanced AI as a transformative technology requiring massive investment and strategic dominance; others treat alignment and safety as existential priorities. Bostrom’s work has shaped the discourse by framing superintelligence as a potential existential risk requiring global coordination. The contribution is a serious engagement with long-term risk and with the technical problem of alignment. The occlusion is that existential-risk discourse can be culturally specific, can overstate the near-term feasibility of AGI, and can divert attention from present harms such as bias, surveillance, and labor exploitation. It may also reflect the interests of large technology firms that benefit from both the urgency and the proposed governance frameworks.

On governance, Western states have produced the most elaborate AI governance architectures: the US executive order on AI, the EU AI Act, export controls on advanced chips, and various alliance-based initiatives. These frameworks emphasize risk management, transparency, human rights, and safety. The contribution is institutional capacity and legal sophistication. The occlusion is that governance often follows market and security interests, that tech firms exercise substantial influence, and that Western standards may be exported as universal norms without adequate adaptation. The Western perspective thus contains a genuine self-critique: it recognizes that leadership in AI is not the same as legitimacy in AI governance, and that self-regulation by powerful firms is insufficient. But it often struggles to translate that critique into constraints on its own strategic behavior.

2.7.5 UN Perspective

The UN perspective frames AI governance through human rights, inclusive participation, multilateralism, and capacity building. It does not treat AI primarily as a military competition among great powers, but as a global issue affecting development, peace, security, and human dignity. The UN’s comparative advantage is legitimacy: it provides a forum in which all states, including those without advanced AI capabilities, can articulate interests and norms. This perspective is essential for preventing AI governance from becoming merely an extension of great-power rivalry.

On narrow AI in military systems, the UN perspective emphasizes compliance with international humanitarian law, human rights, and accountability. It does not deny states’ security needs, but it insists that military AI must be governed by existing legal frameworks and, where necessary, by new rules. The contribution is to keep the human cost of AI-enabled warfare visible, particularly for civilians in conflict zones. The occlusion is that the UN has limited enforcement capacity and depends on member states to implement and comply with norms.

On LAWS, the UN perspective is most developed through the Convention on Certain Conventional Weapons process. The CCW provides a forum for discussing lethal autonomous weapons, and many UN officials, including the Secretary-General, have urged states to adopt new international rules to ensure human control over the use of force. The contribution is a sustained multilateral effort to prevent an unregulated arms race in autonomous weapons and to preserve human accountability. The occlusion is that consensus requirements allow a small number of states to block progress, and that the process can be slow relative to technological change. The UN perspective thus offers legitimacy without guaranteeing effectiveness.

On hyperwar, the UN emphasizes prevention, diplomacy, and the protection of civilians. It is less focused on operational military AI than on the systemic risks of escalation and on the need for confidence-building measures, transparency, and human oversight. The contribution is to frame decision compression as a threat to international peace and security, not merely as a technical problem. The occlusion is that the UN cannot compel major powers to slow their AI development or to share sensitive information.

On AGI, the UN perspective treats superintelligence as a potential global risk requiring inclusive governance. The Secretary-General’s advisory initiatives, the Global Digital Compact, and the Pact for the Future include commitments on AI capacity building, human rights, and international cooperation. The contribution is to broaden the agenda beyond safety and security to include development, equity, and human rights. The occlusion is that these commitments are often aspirational and depend on funding and political will that may not materialize.

On governance, the UN perspective is the most explicitly inclusive. It emphasizes capacity building for developing countries, bridging the AI divide, and ensuring that global AI governance is not captured by a few states or firms. This aligns with African demands for equitable participation and with Chinese emphasis on development, though the UN frames these through human rights and multilateral legitimacy rather than sovereignty alone. The contribution is a normative architecture that can accommodate diverse perspectives. The occlusion is that the UN cannot overcome the competitive incentives of major powers, and that its processes may produce lowest-common-denominator outcomes. Nevertheless, the UN perspective is indispensable because it provides the only forum in which AI governance can be debated as a common human concern rather than as a zero-sum competition.

2.7.6 Comparison Matrix

2.8 WORKED EXAMPLE (I do)

Case: The 2020 Nagorno-Karabakh war and the 2023–2024 Ukrainian drone deployments. The puzzle is how AI-enabled warfare has changed the battlefield without yet amounting to fully autonomous war. The evidence points to a shift in the sensor-to-shooter chain: more distributed sensing, faster target nomination, cheaper precision strike, and heavier reliance on human-machine teaming. The case is not a demonstration of artificial general intelligence; it is a demonstration of narrow AI and automation compressing decision time and altering cost-exchange ratios.

Levels of analysis. At the individual level, the decisive change is cognitive compression. Operators, analysts, and commanders face accelerated kill chains in which machine-assisted detection and classification reduce the time available for human judgment. Trust in automation becomes a battlefield variable: excessive trust can produce automation bias, while excessive skepticism can waste the speed advantage. At the state level, Azerbaijan and Ukraine show how military effectiveness depends on procurement ecosystems, doctrine, training, and industrial adaptation rather than on a single platform. Azerbaijan integrated Turkish and Israeli systems into a coherent campaign; Ukraine adapted commercial drones, FPV platforms, and software under wartime pressure. At the systemic level, the wars are embedded in regional and great-power competition: Turkey, Israel, Russia, NATO members, and Iran all shape the supply, doctrine, and escalation environment. At the transnational level, the battlefield is sustained by global supply chains, commercial electronics, open-source software, private firms, sanctions evasion, and transnational norms about acceptable targeting and autonomy.

Mechanism. The mechanism is not “AI fights the war.” It is the compression and redistribution of the OODA loop. In Nagorno-Karabakh, loitering munitions, unmanned aerial vehicles, and precision artillery created a reconnaissance-strike complex that made Armenian air defenses and armor vulnerable once detected. The key effect was not merely destruction but paralysis: forces could not maneuver without being seen and struck. In Ukraine, the mechanism is more decentralized. Cheap drones, FPV systems, electronic warfare, and rapid software iteration allow both sides to find, fix, and finish targets at low cost, while counter-drone measures force continuous adaptation. Narrow AI appears in target recognition, route optimization, electronic-warfare support, and logistics planning, but the lethal act usually remains human-mediated or semi-autonomous. The battlefield changes because the cost of precision has fallen and the tempo of adaptation has risen. This produces attrition, dispersion, and a premium on concealment, mobility, and electronic resilience.

The principal-agent problem remains central. Algorithms recommend; humans authorize; but compression reduces the quality of authorization. The risk is not only unlawful targeting but also escalation through speed: a system that can respond faster than political leaders can deliberate may turn tactical contact into strategic crisis. Hyperwar, in this sense, is not science fiction; it is the danger that algorithmic tempo outruns human control.

Perspective reconstruction. The African perspective sees these wars as evidence that the military-technological frontier is concentrating in a small set of states and firms, widening the AI divide and embedding new dependencies. It stresses that African states are often consumers, not producers, of the relevant technologies, and that lessons from Eurasian battlefields may be imported without attention to local political contexts. It demands equitable access to defensive capabilities, capacity building, and a voice in global AI governance, while warning that AI-enabled warfare can deepen asymmetry between technologically advanced and technologically dependent states. The Chinese perspective reads the case through sovereignty and strategic competition: military AI must remain under state control, and the lesson is that technological dependence is a security vulnerability. It emphasizes development, controllability, and the dangers of a US-led AI arms race, while treating multilateral constraints with suspicion if they appear to freeze Chinese advancement. The Russian perspective treats the wars as confirmation that precision, unmanned systems, and AI-enabled reconnaissance are now central to modernization and strategic stability. It emphasizes parity, deterrence, and the need to prevent technological surprise, but tends to understate the escalation risks created by autonomous and semi-autonomous systems. The Western perspective focuses on the tactical revolution: human-machine teaming, precision strike, and the legal and ethical problems of algorithmic targeting. It sees Nagorno-Karabakh and Ukraine as proof that AI-enabled warfare is already here, and it uses the cases to argue for responsible use, verification, and alliance adaptation. The UN perspective treats the wars as warnings about the speed of technological change and the erosion of human control. It stresses civilian protection, accountability, and the need for multilateral guardrails, while acknowledging that the UN cannot stop competitive adoption.

Explicit adjudication. The Western perspective best explains the immediate battlefield mechanism: it captures how sensing, targeting, and precision strike have compressed the kill chain and changed force employment. But it is weakest on distribution: it treats the wars as cases of technological adaptation rather than as symptoms of a global AI divide. The African perspective best explains the structural consequence: it shows that the same technologies that make advanced militaries more effective also deepen dependence and inequality. The UN perspective best captures the normative stakes but cannot explain why states adopt these systems despite risk. The Russian and Chinese perspectives correctly identify competition and sovereignty but understate the principal-agent and escalation problems. The strongest overall explanation is therefore layered: the Western perspective explains the mechanism, the African perspective explains the distribution, and the UN perspective explains the governance imperative. No single perspective is sufficient, because the case is simultaneously tactical, structural, and normative.

SECTION 2.9 PROBLEM PAIR (We do)

Tasks

1. Mechanism identification. Reconstruct the causal chain by which Israel’s Lavender and Gospel systems reportedly converted raw intelligence into nominated targets during the 2023–2024 Gaza campaign. Specify the inputs, the scoring logic, the human decision node, and the kinetic output, and identify where in that chain the “human in the loop” actually sits.

2. Principal-agent analysis. Apply the principal-agent framework to algorithmic targeting. Who is the principal, who is the agent, and what information asymmetry makes meaningful human oversight difficult? Distinguish “human in the loop,” “human on the loop,” and “human out of the loop” as defined in the LAWS debate.

3. Legal reconstruction (Israeli/Western). Steel-man the argument that algorithmic targeting can satisfy international humanitarian law—distinction, proportionality, and precaution—better than manual processes, and identify the specific evidentiary claims on which that argument rests.

4. Critical reconstruction (human rights/UN/African). Steel-man the counter-argument that these systems erode accountability, scale error, and shift the burden of proof onto civilian populations. Identify the strongest piece of evidence each side would need to win the argument.

5. Comparative placement. Compare the Gaza case with the Nagorno-Karabakh (2020) and Ukraine (2023–2024) worked example. In which case did AI function as a targeting generator, in which as a platform enabler, and what difference does that make for legal responsibility?

6. Falsification. Specify three observable findings that would falsify (a) the claim that algorithmic targeting reduces civilian harm relative to manual targeting, and (b) the claim that it necessarily increases civilian harm. State which data sources could in principle adjudicate between them.

MODEL ANSWER

The Lavender and Gospel systems illustrate a shift from human-driven to algorithm-assisted targeting at scale. According to investigative reporting, Lavender generated lists of suspected Hamas operatives by applying pattern-recognition models to communications metadata, social-network data, and behavioral indicators, producing a “score” for each individual; Gospel performed a parallel function for buildings and infrastructure, nominating structures as military objectives. The causal chain runs: intelligence ingestion → feature extraction → probabilistic scoring → threshold-based nomination → human approval (often compressed to seconds) → strike authorization → kinetic effect. The critical node is the human approval step. In principle it preserves legal accountability; in practice, the volume of nominations, the opacity of the scoring, and the operational tempo convert approval into a formality, with the human operator ratifying rather than evaluating the machine’s output. This is the principal-agent problem in its algorithmic form: the state (principal) delegates target nomination to a model (agent) whose reasoning it cannot fully inspect, while the human “approver” lacks both the time and the information to correct systematic errors.

The Israeli and Western legal defense rests on several claims. First, that algorithmic systems can process vastly more data than human analysts, improving distinction by identifying patterns invisible to humans. Second, that the systems are advisory, with a human retaining legal responsibility for each strike, thereby satisfying IHL’s requirement of human judgment. Third, that Hamas’s embedding in civilian infrastructure makes manual targeting equally or more error-prone, so the relevant comparison is not algorithm versus ideal human but algorithm versus human under combat stress. The evidentiary burden here falls on demonstrating that the human approver exercised genuine judgment—reviewing the basis for each nomination, assessing proportionality, and rejecting nominations where civilian risk was excessive.

The critical reconstruction challenges each claim. On distinction, pattern recognition trained on incomplete or biased data may systematically misclassify civilians as combatants, particularly where “combatant” is inferred from behavioral proxies (phone usage, movement patterns) rather than direct participation in hostilities. On proportionality, the speed of the kill chain compresses the deliberation that IHL’s precautionary principle requires. On accountability, the opacity of the model and the classification of its parameters as military secrets make it nearly impossible for victims, courts, or oversight bodies to reconstruct why a particular individual was targeted. The African and UN perspectives add a structural critique: these systems are developed and deployed by technologically advanced states against populations with no comparable capacity to audit, contest, or retaliate, reproducing a colonial asymmetry in which the Global South bears the costs of Northern innovation. The strongest evidence for the critical side would be documented cases of misidentification at scale; the strongest for the defense would be audited data showing lower civilian-casualty rates per strike than in comparable manually targeted campaigns.

Comparatively, Nagorno-Karabakh (2020) and Ukraine (2023–2024) represent platform-enabled warfare: drones and loitering munitions extended human reach and precision, but target selection remained substantially human. Gaza represents targeting-generation: the algorithm itself produces the list of persons to kill. The legal difference is significant. In platform-enabled cases, responsibility attaches to the human operator who chose the target; in targeting-generation cases, responsibility is distributed across the model’s designers, trainers, data curators, and approvers, none of whom may have made an individual targeting decision. This diffusion is precisely what makes algorithmic targeting legally and ethically novel.

Falsification requires data that neither side currently possesses in public form. The claim that algorithmic targeting reduces civilian harm would be falsified by evidence that (i) the false-positive rate for combatant identification exceeded that of manual processes in comparable contexts, (ii) human approvers rejected fewer than a trivial fraction of nominations, or (iii) post-strike assessments revealed systematic misidentification of protected persons. The claim that it necessarily increases harm would be falsified by (i) audited evidence of lower civilian-casualty ratios per engagement, (ii) documented instances of approvers rejecting nominations on proportionality grounds, or (iii) transparency measures enabling external audit without compromising operational security. Adjudication would require access to nomination logs, approval records, strike assessments, and model documentation—precisely the materials states classify. The epistemic gap is itself a finding: the accountability architecture of algorithmic warfare is structurally resistant to the verification that IHL presupposes.

SECTION 2.10 INDEPENDENT MEMO (You do)

(a) Case description. In late 2023, a coalition of African states at the UN General Assembly tabled a resolution calling for equitable access to AI technologies and a global fund for AI capacity-building in the Global South. The resolution passed, but implementation stalled. Meanwhile, a major African military—facing insurgency in its northern regions—procured an AI-enabled surveillance and targeting system from a foreign vendor whose training data and model weights remain proprietary. Civil society groups allege the system misidentifies civilians as combatants. The state’s parliament demands an independent audit, but the vendor cites national-security and commercial confidentiality. The state must now decide whether to suspend the system, renegotiate terms, or seek alternative suppliers. The case sits at the intersection of the AI divide, algorithmic accountability, and sovereign procurement autonomy.

(b) Required memo headings (use these as prompts).

• Mechanism: Identify the causal chain linking the foreign vendor’s proprietary model to targeting outcomes on the ground. Where does human judgment sit, and what information asymmetry prevents the state from auditing the system? Specify the principal-agent relationships among the vendor, the military, the parliament, and affected civilians.

• Perspective reconstruction: Reconstruct how each of the five perspectives—African, Chinese, Russian, Western, UN—would frame this case. For each, state the core normative claim, the preferred policy prescription, and the strongest objection from within that perspective.

• Historical analogue: Identify one historical case (pre-2020) in which a technologically dependent state faced a comparable accountability gap in military procurement, and specify what is genuinely analogous and what differs given the opacity and speed of AI systems.

• Falsifiers: List three observable findings that would falsify your central claim about whether the state should suspend, renegotiate, or replace the system. State what data you would need and why it is or is not obtainable.

(c) Self-check. Before submitting, verify that you have distinguished description (what happened), explanation (why it happened), prediction (what will happen), and prescription (what should be done)—and that no section conflates them.

SECTION 2.11 HINGE QUESTION

Question. Which of the following best captures the principal-agent problem as it arises in algorithmic targeting systems such as Lavender?

A. The algorithm acts as principal and the human operator as agent, reversing the traditional command hierarchy and placing legal liability on the machine.

B. The human operator cannot meaningfully evaluate the algorithm’s output because of model opacity and operational tempo, so approval becomes a formality rather than a judgment.

C. The state delegates targeting to a private contractor, creating an accountability gap between public authority and commercial interest.

D. International humanitarian law lacks jurisdiction over non-state actors, so algorithmic systems used by states fall outside existing legal frameworks.

Correct answer: B.

Distractor diagnoses.

• A misidentifies the principal-agent relationship. The state and its commanders remain the principal; the algorithm is an instrument, not a legal agent. Legal liability does not transfer to a machine under current IHL, and framing the algorithm as principal obscures the human choices embedded in its design and deployment.

• C identifies a real but secondary issue. Contractor involvement can complicate accountability, but the core principal-agent problem in algorithmic targeting is informational—opacity and speed—not commercial delegation. A state-operated system with the same opacity would raise the same problem.

• D misdiagnoses the legal gap. IHL applies to state conduct regardless of the technology used; the difficulty is not jurisdictional absence but the evidentiary and procedural challenge of verifying compliance when the decision chain is algorithmic and classified.

SECTION 2A — DATA LAB

Table 1 — State AI Investment and Compute Capacity (selected, 2023)

Note: Figures are approximate and vary by definition of “AI investment” (private venture vs. total R&D). The US lead is structural, not merely cyclical, and is concentrated in frontier-model compute.

Table 2 — Autonomous and AI-Enabled Military Systems by State (selected, 2024)

Autonomy levels follow Scharre’s (2018) spectrum: human-in-the-loop, human-on-the-loop, human-out-of-the-loop.

Table 3 — UN CCW Process on Lethal Autonomous Weapons Systems (LAWS)

Source: UN CCW LAWS reports; UN Secretary-General statements.

Table 4 — AI Incident and Safety Trend (selected indicators)

Incident counts are approximate and reflect reporting bias; they indicate a rising trend in documented harms, not a complete census.

Five Numbered Analysis Questions

1. Trend. Table 1 shows a roughly 8:1 US–China private AI investment ratio in 2023, yet Table 2 shows China fielding competitive autonomous systems. What does this divergence suggest about the relationship between commercial AI investment and military AI capability? Is the gap closing, and on what time horizon?

2. Level of analysis. The Lavender/Gospel systems (Table 2) are deployed at the state level by Israel, but their effects are felt at the individual level (targeted persons) and the transnational level (international humanitarian law norms). Using the four levels of analysis from Lesson 1, explain which level best explains the proliferation of such systems and which best explains the constraint on their use.

3. Perspective reconstruction (all five). Reconstruct how each of the five perspectives would read Table 3’s stalemate: (a) African — as evidence of a governance gap that excludes non-possessors; (b) Chinese — as a sovereignty-preserving outcome; (c) Russian — as strategic-stability management; (d) Western — as a failure of norm entrepreneurship; (e) UN — as incremental multilateralism. Which reading is best supported by the evidence in Tables 1–4?

4. Interpretation. Table 4 shows rising AI incidents alongside rising national AI strategies. Does this correlation imply that governance is failing, or that detection and reporting are improving? Distinguish description from explanation, and identify what evidence would adjudicate between the two readings.

5. Limitation. All four tables rely on open-source, self-reported, or advocacy-compiled data. Identify the single most important data gap that, if filled, would most change your assessment of whether AI is destabilizing or stabilizing in the military domain, and explain why.

SECTION 3 — DUAL-CODED VISUAL ANCHOR

Visual 1 — “The Compression Spiral”: OODA-Loop Collapse and Autonomy Escalation

What to draw: A spiral (or nested concentric arcs) on a horizontal time axis running left-to-right, labeled “Time to decision (seconds → milliseconds → microseconds).” Three nested loops, each smaller and faster than the last: (1) outer loop labeled “Human-in-the-loop (HITL): minutes–hours” with a human icon and a “veto” gate; (2) middle loop “Human-on-the-loop (HOTL): seconds” with a human icon watching a screen; (3) inner loop “Human-out-of-the-loop (HOOTL): milliseconds” with no human icon, only an algorithm node. Arrows show escalation pressure pushing inward. A vertical dashed line marks “Flash-war threshold” where human cognition cannot keep pace.

Labels to place: “Observe–Orient–Decide–Act (OODA)”; “Algorithmic speed”; “Human cognitive limit (~250 ms reaction)”; “Escalation pressure”; “Flash war zone.”

Perspective overlay: Annotate the spiral’s outer rim in five colors: African (red) — “who is excluded from the loop?”; Chinese (blue) — “sovereignty over the loop”; Russian (green) — “strategic stability at speed”; Western (purple) — “norms of human control”; UN (orange) — “CCW / Pact for the Future.” Place a small legend in the corner.

Visual 2 — “The Principal-Agent Problem in LAWS”: Misalignment and Accountability Gap

What to draw: A flowchart with two boxes: “Principal (State / Commander)” on the left and “Agent (Autonomous System)” on the right, connected by a solid arrow labeled “Delegation.” Below, a dashed arrow returns labeled “Accountability?” with a question mark. Between them, a jagged “misalignment zone” (shaded) containing three failure modes: (a) “Objective mismatch (reward hacking)”; (b) “Context shift (training ≠ battlefield)”; (c) “Emergent behavior (unpredictable coordination).” Below the boxes, a horizontal bar labeled “Governance gap” with segments: “CCW GGE (non-binding)” | “National policy” | “International humanitarian law (existing)” | “No treaty.”

Labels to place: “Principal”; “Agent”; “Delegation”; “Accountability gap”; “IHL (distinction, proportionality)”; “Command responsibility.”

Perspective overlay: Five small flags or color-coded tags at the bottom: African — “accountability for victims in the Global South”; Chinese — “state responsibility preserved”; Russian — “no new constraints on modernization”; Western — “human control as norm”; UN — “GGE guiding principles.”

Visual 3 — “AI Governance Landscape”: Concentric Regimes and the African Demand

What to draw: A set of concentric circles (or a Venn diagram) centered on “AI in military use.” Innermost circle: “Hard law (none binding on LAWS).” Second: “Soft law / norms” (CCW guiding principles, Pact for the Future, REAIM). Third: “National regulation” (US EO 14110, EU AI Act, China’s Interim Measures, UK AI Safety Institute). Outermost: “Industry / technical standards” (ISO, IEEE, model evals). Around the circles, place five perspective markers: African (outermost, pointing inward) — “demand for equitable participation and technology transfer”; Chinese — “sovereignty and non-interference”; Russian — “strategic stability, no disarmament of conventional AI”; Western — “leadership + self-regulation + alliance coordination”; UN — “multilateral framework, SDG alignment.”

Labels to place: “Hard law gap”; “Soft norms”; “National regimes”; “Technical standards”; “Equity gap (compute, data, talent).”

Perspective overlay: Use the same five-color legend as Visual 1. Add a small arrow from the African marker toward the center labeled “restructuring demand.”

SECTION 4 — CORNELL NOTES

Source log: Scharre (2018) Army of None; Horowitz (2018) “Artificial Intelligence, International Competition, and the Balance of Power”; Kissinger, Schmidt & Huttenlocher (2021) The Age of AI; Bostrom (2014) Superintelligence; UN CCW LAWS reports; Stanford AI Index (2024); Grovogui, Mbembe, Aning, Danso (African IR scholarship).

Summary (3–5 sentences): Military AI spans narrow systems already in use (ISR, targeting, logistics) and speculative superintelligence, with LAWS sitting at the contested boundary. The central risk is decision compression: as algorithms accelerate the OODA loop, human control erodes and flash wars become plausible. Governance is fragmented across national regimes and a stalled CCW process, with no binding LAWS treaty. The African perspective insists that this fragmentation reproduces the AI divide and demands equitable participation, while Chinese, Russian, Western, and UN framings each privilege sovereignty, stability, leadership, or multilateralism respectively. Whether AI stabilizes or destabilizes depends on whether speed outruns accountability.

SECTION 5 — FEYNMAN COMPRESSION

1. Explain without jargon. Imagine a chess clock that keeps speeding up. At first a human general has hours to decide; then minutes; then the computer moves on its own in milliseconds. That is “hyperwar”: the machine’s speed outruns human judgment, and a small spark can become a big fire before anyone can stop it. AI in war today is mostly narrow tools—better eyes, faster targeting—not robot generals, but the trend points toward machines making lethal choices. Countries argue about whether to ban “killer robots,” yet no binding rule exists, and the big powers are racing ahead. The deeper worry is a super-smart AI that no one can control; the nearer worry is a fast, dumb mistake.

2. Mark gaps. I am unsure how much Lavender/Gospel actually changed targeting rates versus human staff; I conflated “AI” with “automation” in places; I have not specified the mechanism by which AGI would destabilize (capability jump vs. misalignment); I treated the CCW stalemate as failure rather than as a sovereignty-preserving equilibrium for some states.

3. Reread verification. Re-reading Scharre (2018) confirms the spectrum of autonomy (HITL/HOTL/HOOTL) and the principal-agent framing; Horowitz (2018) confirms that AI is a general-purpose technology whose military diffusion depends on organizational adaptation, not just algorithms; Kissinger et al. (2021) confirms the “compressed decision” thesis; Bostrom (2014) confirms the alignment problem as a control problem. The African-scholarship framing (Grovogui, Mbembe, Aning, Danso) is consistent with the technology-gap and equity-demand narrative.

4. Analogy with explicit limits. The chess-clock analogy captures speed and loss of human control but breaks down on three points: (a) war is not zero-sum with fixed rules—adversaries adapt; (b) the “clock” is not neutral—it is owned by states with unequal compute; (c) a chess loss ends a game, whereas a flash war can escalate to nuclear use. The analogy is pedagogically useful for compression but must not be mistaken for a model of strategic interaction.

5. Rewrite. Military AI is best understood as a compression technology: it shrinks the time between observation and lethal action, eroding the human control that international humanitarian law presupposes. Narrow systems (ISR, targeting, logistics) are already deployed and have changed battlefields (Nagorno-Karabakh 2020, Ukraine 2023–24, Gaza 2023–24), while LAWS remain legally unregulated despite a decade of CCW talks. AGI raises a categorically different problem—alignment as global security—but its probability and timeline are deeply uncertain. Governance is fragmented (US EO, EU AI Act, Chinese measures, UN soft law), and the African perspective highlights that this fragmentation reproduces the AI divide: those who build the systems are not those who bear the risks. The stabilizing case (better information reduces fog) and the destabilizing case (speed outruns accountability) are both live; adjudication requires evidence on escalation dynamics, not just capability counts.

SECTION 6 — SPACED-REPETITION FLASHCARDS

| Tag | Card Type | Front | Back | Perspective Note | | L54 | definition | What is narrow AI in military systems? | Narrow (weak) AI performs specific tasks—ISR, targeting, logistics, C2—without general reasoning; it is the current operational layer of military AI, distinct from AGI. | African scholars note the asymmetry: African militaries are consumers, not producers, of narrow AI systems. | | L54 | mechanism | How does decision compression create hyperwar risk? | Algorithmic speed compresses the OODA loop; when both sides use AI, escalation can outpace human deliberation, producing flash wars. | Russian strategic-stability framing warns of compressed timelines undermining crisis management. | | L54 | application | What did the 2020 Nagorno-Karabakh war demonstrate about AI-enabled warfare? | It showcased loitering munitions (Harop, Orbiter) and drone ISR enabling rapid, precise strikes against armor and air defenses. | Western analysts read it as proof of concept; African analysts stress proliferation to non-state actors. | | L54 | comparison | How do LAWS differ from remote-controlled drones? | LAWS select and engage targets without human intervention in the critical function; drones retain a human in the loop. | The CCW GGE debates this distinction; African states often stress human control. | | L54 | source | What is Scharre’s (2018) core argument about autonomous weapons? | In “Army of None,” Scharre argues autonomy is a spectrum and the key question is meaningful human control, not a binary ban. | Western defense perspective; African delegates often push for stronger prohibitions. | | L54 | debate | What is the principal-agent problem in LAWS? | The commander (principal) cannot fully predict or control the autonomous system (agent), creating accountability gaps. | UN CCW reports highlight this as central to IHL compliance. | | L54 | reconstruction | What is China’s position on AI governance? | China supports UN-led governance, opposes an “AI arms race,” and frames AI as a sovereignty and development issue. | Chinese perspective emphasizes multilateralism and development rights. | | L54 | occlusion | What does the Western AI-leadership framing occlude? | It tends to treat AI competition as a US-China dyad, marginalizing African and Global South agency and the AI divide. | African perspective centers the technology gap and equitable access. | | L54 | adjudication | How should one adjudicate between US and Chinese AI governance proposals? | Compare scope (binding vs. voluntary), enforcement, and inclusion of Global South voices; neither fully addresses the AI divide. | African scholars demand equitable governance, not just great-power bargains. | | L54 | visual | What does the Stanford AI Index track? | Private investment, model performance, and policy developments in AI, showing concentration in US and China. | African perspective: the index reveals the AI divide in investment and talent. | | L54 | definition | What is AGI and why is it a security problem? | AGI would match or exceed human cognition across domains; its security dilemma arises from first-mover advantage and alignment uncertainty. | Bostrom (2014) frames alignment as existential; African scholars stress whose values align. | | L54 | mechanism | What is the alignment problem? | Ensuring an advanced AI’s objectives match human values; misalignment could produce catastrophic outcomes even without malice. | Kissinger et al. (2021) warn AI may outpace human comprehension. | | L54 | comparison | How does the EU AI Act differ from the US executive order? | The EU AI Act is binding risk-based regulation; the US EO relies on voluntary commitments and agency guidance. | African regulators watch both as templates but lack capacity to implement. | | L54 | source | What does Horowitz (2018) argue about military AI diffusion? | Horowitz argues military AI diffusion depends on organizational capacity and regime type, not just technology. | Explains why African states lag in adoption beyond hardware. | | L54 | debate | What is the UN Secretary-General’s position on LAWS? | Guterres has called for a legally binding instrument prohibiting lethal autonomous weapons by 2026. | African and many Global South states support a binding treaty. |

SECTION 7 — METACOGNITIVE CALIBRATION

7.1 Confidence audit (score 0–100):

7.2 Three-question debrief:

1. Which of the five perspectives did you reconstruct most weakly, and what evidence would strengthen it?
2. Did you conflate description (what happened) with explanation (why) in the Lavender/Gospel case?
3. Which falsifier would most change your conclusion about whether AI-enabled warfare is stabilizing or destabilizing?

7.3 Error-log:

SECTION 8 — CHEATSHEET

SECTION 9 — REVIEW SCHEDULE

• R1 (Day 1, 30 min): Re-read Section 2.2–2.6 bolded causal claims; recite the five key ideas from memory; check flashcards tagged L54.
• R2 (Day 3, 45 min): Reconstruct all five perspectives on the central topic (AI, AGI, and hyperwar) without notes; compare against Section 2.7.
• R3 (Day 7, 60 min): Work the Problem Pair (Lavender/Gospel) memo from scratch; verify you distinguished description, explanation, prediction, prescription.
• R4 (Day 14, 45 min): Re-run the Data Lab analysis questions; update confidence audit scores in Section 7.1.
• R5 (Day 30, 90 min): Full synthesis: write a 500-word adjudication of whether AI-enabled warfare is stabilizing or destabilizing, citing at least three perspectives and two falsifiers.

SECTION 10 — COMPLETION CHECKLIST

• I can define narrow AI, LAWS, hyperwar, AGI, ASI, and alignment.
• I can explain the mechanism of decision compression and flash-war risk.
• I can describe the principal-agent problem in autonomous weapons.
• I can reconstruct the CCW LAWS process and the Secretary-General’s position.
• I can compare US EO, EU AI Act, Chinese regulations, and UN framework.
• I can apply all four levels of analysis to the Nagorno-Karabakh/Ukraine case.
• I can reconstruct all five perspectives on AI, AGI, and hyperwar.
• I can identify what the African perspective reveals that others miss.
• I can adjudicate between competing explanations with evidence.
• I can state at least two falsifiers for my conclusions.
• I have completed the confidence audit with honest scores.
• I have logged at least one misconception in the error-log.
• I can read the Stanford AI Index data and interpret the AI divide.
• I can explain Horowitz’s (2018) diffusion argument.
• I can explain Bostrom’s (2014) alignment concern.
• I can explain Scharre’s (2018) spectrum-of-autonomy argument.
• I can explain Kissinger et al.’s (2021) warning about AI and human comprehension.
• I can distinguish description, explanation, prediction, and prescription.

FURTHER READING (Optional)

Military AI and autonomous weapons:

• Scharre, Paul. Army of None: Autonomous Weapons and the Future of War. W.W. Norton, 2018.
• Horowitz, Michael C. “The Promise and Peril of Military Applications of Artificial Intelligence.” Bulletin of the Atomic Scientists, 2018.
• Horowitz, Michael C. “Artificial Intelligence, International Competition, and the Balance of Power.” Texas National Security Review, 2018.

AGI, alignment, and superintelligence:

• Bostrom, Nick. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014.
• Kissinger, Henry A., Eric Schmidt, and Daniel Huttenlocher. The Age of AI: And Our Human Future. Little, Brown, 2021.

Governance and policy:

• UN CCW Group of Governmental Experts on Lethal Autonomous Weapons Systems, reports and chair’s summaries.
• Stanford Institute for Human-Centered AI. AI Index Report (annual).

Chinese, Russian, and African perspectives:

• Kania, Elsa B. “Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power.” CNAS, 2017.
• Allen, Gregory C. “Understanding China’s AI Strategy.” CSIS, 2019.
• Grovogui, Siba N. Regimes of Sovereignty: International Morality and the African Condition. Palgrave Macmillan, 2016.

End of Lesson 54. Say Next when you are ready.