Another revisit of my Ukraine Battlefield AI article from 2022, now that Trump has shown his colors: and defense of democracy Ronald Loui Ronald Loui r.p.loui@gmail.com October 18, 2026 GEMINI: I've got the article: Ronald Loui, “Battlefield AI and Ukraine,” written in April 2022 and posted February 20, 2023. Pasted text Let's revisit it. I think the most productive starting point is to distinguish the article's central argument from its supporting claims and then ask what has held up, what needs revision, and what the argument implies for AI governance today. The article's central thesis is striking: The principal danger of battlefield AI is not autonomous weapons' ability to kill precisely, but the concentration of command and control that automation makes possible. You argue that precision-guided weapons can make warfare less indiscriminate, potentially advancing the laws of armed conflict. The more consequential danger is strategic: AI can make it possible for one leader to initiate mass destruction while bypassing the human beings who might otherwise question, resist, or refuse an order. Pasted text Three ideas give the argument its distinctive character: Precision versus lethality. The ethical significance of a weapon depends partly on its effects, not simply on whether it uses AI or operates autonomously. Better targeting can reduce collateral damage, although it cannot guarantee lawful conduct. Pasted text Disobedience as a safeguard. Human hesitation, dissent, and even imperfect compliance can prevent a disastrous command from becoming a catastrophe. Automating obedience can eliminate these informal checks on authority. Pasted text Responsibility must remain assignable. You distinguish distributing responsibility among people who share the burden of a decision from diffusing it so thoroughly that nobody is accountable. The goal is not simply to keep a human somewhere in the loop, but to preserve meaningful human judgment and responsibility. Pasted text The third point connects the first two: a weapon can be tactically precise while the political system controlling it is dangerously unaccountable. Where I'd like to take our reconsideration I see four particularly fruitful questions. The argument's strongest insight: Is the loss of human resistance to authority a more fundamental problem than lethal autonomy? What does this imply for AI systems that automate decisions throughout a military bureaucracy, rather than directly operating weapons? The Arkhipov problem: Your account of Vasily Arkhipov illustrates the value of a person who recognizes that an established decision rule is defective and creates a different decision path. Can AI systems preserve that capacity for reconsideration, or will they increasingly eliminate the opportunities for it? Pasted text The potential weakness: You make a persuasive case for precision, but precision is not the same as restraint. A system that makes military operations cheaper, faster, and more reliable could also encourage leaders to undertake more operations. Does the article adequately address that possibility? The broader political argument: If the danger is concentrated decision-making rather than the autonomous weapon itself, what institutional arrangements would actually prevent an AI-enabled leader from bypassing meaningful checks and balances? My initial assessment is that the article's most original contribution is its treatment of human disobedience as a form of systemic robustness. That is a more interesting proposition than the familiar demand to keep a human in the loop. A human who merely confirms an automated recommendation is not necessarily a safeguard; a human who can question the premise, change the rules, delay the action, or refuse the order might be. I'd start with the distinction between human presence and human independence of judgment. It seems to me that this is the conceptual hinge of the article—and the place where its implications extend well beyond warfare. Where would you like to begin: the argument itself, how events since 2022 have tested it, or its implications for the design and governance of AI systems? how does it hold up in light of the putin, trump, and irg devastation of the world order and national determination in service of the nation as opposed to its autocratic leaders? also, considering that boomer-era roboethics obsessed over target-selection and lethality, how does automation required for the current defense of Ukraine shed light on whether the article was right to be more nuanced about AI on the battlefield? Battlefield AI and Ukraine: revisiting the argument in a world of autocratic power I think the article becomes more interesting when read against the subsequent erosion of the international order. Its central argument was not simply that autonomous weapons are sometimes good and sometimes bad. It was a claim about where the decisive moral and political danger of automation lies. The article's title could almost have been The Wrong Argument About Battlefield AI. You were challenging a tradition of roboethics that concentrated on the weapon: Does it select its target correctly? Does it distinguish a combatant from a civilian? Should a machine be permitted to decide whether someone lives or dies? Your alternative was to examine the political and organizational system in which the weapon operates. Who decides to fight? Who can challenge that decision? What happens when a leader's judgment is defective? And does automation eliminate the human resistance that might prevent a bad decision from becoming a catastrophe? Your 2023 article states the distinction explicitly: the danger is not primarily the tactical autonomy of weapons but the concentration of national decision-making made possible by societal automation. Pasted text Developments involving Putin's war against Ukraine, Trump's approach to American alliances and international commitments, and the Islamic Republic of Iran (IRG, as you put it) make this distinction especially consequential. But they also suggest a further question: What happens when the technology that concentrates power in the hands of a leader is also indispensable to a nation's ability to resist that leader—or to resist an invading power? That is where I would focus our reassessment. 1. Putin, Trump, and the distinction between a nation and its ruler The article's political argument becomes sharper if we distinguish three things that are often conflated: The state: its institutions, legal obligations, and capacity to act. The nation: its people, their collective interests, and their right to determine their future. The ruler: the person or group temporarily exercising state power. These can diverge dramatically. Your article already recognizes this distinction in its discussion of Ukraine. You question the idea that the war is simply one nation fighting another, arguing instead that Putin's decision to invade should not be equated with the collective will of the Russian people. Pasted text That distinction matters for understanding all three cases, although they differ substantially. Putin: automation in the service of a leader's decision Russia's invasion demonstrates why a technically sophisticated military cannot be assumed to serve the national interest merely because it executes orders efficiently. The central political problem is upstream of the weapon. A leader decides to pursue a territorial or geopolitical objective; institutions transmit the decision; military organizations implement it; and increasingly automated systems can make implementation faster, more scalable, and less dependent on individual initiative. In your framework, the danger is a reduction in the number of points at which someone can interrupt the causal chain. There is an important qualification here: Russian military failures have also shown that authoritarian command is not necessarily effective command. Bad information, distorted reporting, organizational dysfunction, and logistical failures can undermine the execution of orders. Automation can improve coordination, but it cannot make a fundamentally misguided political objective serve the public interest. The deeper question is therefore not whether the military obeys efficiently. It is whether the political system has effective means of questioning the purpose of the obedience. Trump: control over the infrastructure of an ally's resistance The Trump case reveals a different dimension of your argument. An ally can retain its sovereignty formally while becoming dependent on another country's communications, intelligence, software, weapons, and logistical infrastructure. The provider of those capabilities can then acquire considerable influence over what the ally can do. A particularly revealing development occurred this week. Reuters reported on October 15 that instructions to restrict Ukraine's access to the military-grade Starshield communications service came from within the White House, according to two U.S. officials and a Western official. The officials did not identify who issued the instruction. The Trump administration disputed suggestions that Trump personally ordered the restriction. Ukrainian officials reported disruptions affecting military communications and drone operations. Reuters This is relevant to your argument even though restricting a communications service is not itself an autonomous-weapon decision. Consider the structure: Ukraine develops operational capabilities around a communications infrastructure. Those capabilities become integral to its ability to defend its territory. The infrastructure remains under the control of an external provider or government. A decision made outside Ukraine can affect Ukraine's ability to conduct military operations. The critical automation problem is no longer just whether a machine can choose a target. It is whether an external decision-maker can alter the conditions under which an entire defense system operates. A nation can lose practical freedom of action without losing formal sovereignty. That is an extension of your article's argument about concentrated command and control. The relevant concentration of power may reside not only in a military commander or head of state, but also in whoever controls an indispensable technological dependency. It also illustrates why national determination matters. Ukrainians can remain determined to defend their country while finding that their ability to do so depends partly on decisions made in Washington. The policy implication is not that countries should reject technological cooperation. It is that essential defense capabilities need resilient alternatives, clear governance arrangements, and the ability to continue functioning when a provider withdraws support. Iran's Revolutionary Guard: whose security does the security apparatus serve? I take your reference to the IRG to mean Iran's Islamic Revolutionary Guard Corps (IRGC). Here, the distinction between a nation and its rulers raises a related but different question: when a security apparatus protects a political order, how closely does that protection correspond to the interests of the people living under it? The IRGC is an institutional and military pillar of the Islamic Republic, with extensive economic and political influence. Its interests and those of the Iranian population cannot simply be presumed identical. Nor should the Iranian population be treated as collectively responsible for the actions of the security apparatus. This matters for AI-enabled surveillance, intelligence analysis, and command systems. The same technological advances that help a country defend itself against external attack can help a regime monitor its population, suppress dissent, and consolidate political control. There is a crucial distinction between defending a nation and defending the incumbent regime. That distinction also cautions against treating the destruction of a regime's military capabilities as equivalent to improving the lives or political freedom of its people. Those are separate outcomes, with different mechanisms and uncertainties. The broader point is that automation has no inherent commitment to the nation, democracy, or human freedom. Its political significance depends partly on the institutions that direct it and the people who can contest those institutions' decisions. The common thread These cases suggest a useful extension of the article. QuestionConventional weapons-centered approachYour political and organizational approachWho is the decision-maker?Who selects and engages the target?Who determines the political objective?Where can harm originate?Inaccurate identification or excessive forceA misguided decision transmitted efficiently through the systemWhat constitutes human control?A person approves the engagementPeople and institutions retain the ability to question, revise, or refuse the policyWhat threatens national autonomy?An adversary's weaponsAn adversary's weapons, domestic authoritarian control, or external dependence on critical infrastructureWhat should accountability accomplish?Establish responsibility for an unlawful strikeEstablish responsibility for the decision, its implementation, and the institutional arrangements that made it possible The last column is not an argument against weapons-level safeguards. It is an argument that those safeguards are insufficient on their own. Your article's concern about concentrated decision-making becomes especially relevant when the leader's objectives diverge from the nation's interests, or when an external actor can exercise effective control over a nation's ability to defend itself. 2. Ukraine presents the strongest test of your argument about battlefield AI Here I think the article's nuance has become especially important. The conventional roboethics debate often starts with a particular image: a robot identifies a person, decides that person is an enemy, and kills them without human intervention. That is an important problem. But it is not a sufficiently comprehensive account of AI in warfare. Ukraine illustrates why. A. Automation can help a nation survive an attack by a much larger power Ukraine has faced a fundamental asymmetry: Russia has substantially greater resources, while Ukraine must conserve its personnel, compensate for shortages, and adapt rapidly to changing battlefield conditions. Unmanned systems, digital battlefield management, automated image analysis, and increasingly sophisticated navigation help address these problems. A 2025 CSIS study of Ukraine's military AI development identified applications including: Analyzing aerial imagery and identifying objects of military interest. Improving situational awareness by integrating information from different sources. Navigating when GPS is unavailable or electronic warfare disrupts communications. Reducing the workload on personnel who would otherwise have to interpret large quantities of information manually. Connecting reconnaissance drones with other weapons systems through a common operational picture. CSIS Notice how many of these functions have little to do with a machine independently deciding whom to kill. They are about making a defense organization more capable. Your article anticipated this broader perspective. You discuss AI in logistics, communications, information processing, and the orchestration of complex military operations, rather than limiting the discussion to autonomous weapons. Pasted text The implication is significant. If Ukraine were to abandon battlefield automation because autonomous weapons pose ethical risks, it could surrender capabilities essential to defending its population and preserving its political independence. That does not establish that every military AI application is justified. It establishes that a serious ethical analysis must count the costs of withholding useful technology as well as the costs of deploying it. There are people whose lives may depend on an automated system detecting an incoming threat, maintaining communications under jamming, or helping a relatively small force coordinate its defense. A framework that counts only the risks of machine-enabled lethality misses those lives. B. The distinction between automation and autonomy matters enormously There is another reason the article's approach looks relevant today: the public debate frequently treats unmanned, automated, AI-enabled, and autonomous as interchangeable terms. They are not. Consider four levels of functionality: LevelFunctionPrincipal question1. AutomationA system executes a predefined procedure.Does it perform the procedure reliably?2. AI-assisted perceptionA system analyzes imagery or other data and identifies possible objects or threats.How reliable is its assessment under actual operating conditions?3. Autonomous navigation and controlA system navigates or maintains its course despite disrupted communications or other obstacles.Can it continue safely and predictably when conditions change?4. Autonomous engagementA system selects and engages targets without human authorization for each engagement.Who authorizes the policy, establishes the constraints, and accepts responsibility for the resulting harm? These levels raise different engineering, operational, and ethical questions. A drone that can navigate without GPS is not necessarily a drone that independently decides whom to attack. A system that identifies a vehicle as a tank is not necessarily authorized to attack it. A human who authorizes an operation is not necessarily exercising meaningful judgment over every subsequent engagement. CSIS's 2025 assessment of Ukraine emphasized that battlefield AI was predominantly being applied to particular functions, rather than enabling complete autonomy from target identification through engagement. Human involvement remained especially important in engagement decisions. CSIS The distinction supports your article's argument that the ethical assessment should begin with what a technology actually does, how it affects people, and where decisions are made—not simply with whether it is labeled an autonomous weapon. But it also presents a challenge for the article: the more these functions become integrated, the harder it becomes to isolate the ethical significance of any one component. C. Ukraine demonstrates the value of decentralized innovation One of the most interesting developments is the way Ukrainian battlefield systems have evolved. A June 2026 CSIS analysis describes Ukraine's Delta battlefield-management system as having grown from a digital map developed by volunteers into a broader operational platform incorporating features such as drone deconfliction, identification of friendly forces, video streaming, AI-assisted analysis, and secure messaging. The same analysis contrasts this bottom-up evolution with attempts to build comprehensive command-and-control architectures from the top down. CSIS This suggests a particularly important distinction for your argument. There are at least two ways to use automation in military organizations: Centralizing automation: A central authority receives more information, exercises more control over subordinate units, and can ensure more immediate compliance with its orders. Distributed automation: Units and operators receive tools that improve their ability to interpret local conditions, coordinate with others, and act without waiting for every decision to come from above. Both can increase military effectiveness. But they have different implications for the distribution of authority. The second arrangement can preserve room for local initiative and adaptation. It can make a force more resilient when communications fail, circumstances change, or central direction proves mistaken. The first can make it easier for a leader to impose a decision throughout the organization. This distinction goes to the heart of your article. You argue that automation becomes politically dangerous when it eliminates the opportunities for people to resist an objectionable command. Pasted text But the Ukrainian experience suggests that automation can also be designed to give people more independent capacity to act. The relevant distinction is not simply between automated and human-operated warfare. It is between technological architectures that concentrate judgment and those that distribute the capacity for informed action. Of course, distributed control does not automatically produce moral conduct. A decentralized network can enable unlawful actions, and swarms without a clear chain of responsibility can make accountability harder. Nor does a bottom-up development process guarantee that final authority is decentralized. The architectural question is whether the system provides genuine local initiative while retaining effective coordination, oversight, and responsibility. That is a much more demanding question than whether a human is technically present somewhere in the system. 3. Was the older roboethics debate looking in the wrong place? Your criticism of the boomer-era debate has considerable force, provided we distinguish the debate's limitations from the importance of the problems it addressed. Much of the public discussion of autonomous weapons has revolved around lethal target selection, the delegation of life-and-death decisions, and the question of whether a machine should be allowed to kill without direct human authorization. These are legitimate concerns. They remain so. But there is a danger in treating them as the whole problem. The target-selection problem is downstream of the political decision Suppose a military system can identify a legitimate military target with extraordinary precision. That answers a technical question about identification. It does not answer whether the military operation should occur, whether the anticipated military advantage justifies the expected harm, or whether the operation serves a legitimate political purpose. Conversely, suppose a system is highly reliable at navigation and helps intercept incoming weapons directed at civilian infrastructure. Its ability to perform that function may be valuable even though the system incorporates substantial automation. The two cases cannot be adequately evaluated by asking whether AI is involved in a lethal decision. The relevant questions include: What objective is the system serving? Who authorized that objective? What alternatives were considered? What risks does the system reduce, and what new risks does it introduce? Can a person intervene when circumstances change? Who is accountable if the system performs as designed but the resulting action is wrongful? Your article's discussion of the laws of armed conflict already points in this direction. You argue that AI can potentially improve distinction, proportionality, military necessity, and the limitation of unnecessary suffering, while also recognizing that AI can be used to undermine all four. Pasted text The important qualification is that technical precision does not guarantee lawful or proportionate conduct. Nor does it necessarily reduce the overall amount of violence. A system can become more humane in one respect and more dangerous in another Consider a hypothetical defensive system that becomes considerably better at identifying and intercepting incoming missiles. Its improved performance might save civilian lives. But if the same technological improvements make military operations less costly and more predictable, political leaders might also become more willing to undertake operations that they previously would have considered too risky. This is a possible mechanism, not an inevitable result. The net effect depends on the circumstances. Nevertheless, it demonstrates why weapon-level assessments are insufficient. A system can improve the expected outcome of each engagement while changing the political incentives governing how often engagements occur. The ethical evaluation must consider both. The same applies to automation that reduces the danger to one's own troops. That is a potentially important humanitarian benefit. But it does not by itself establish that the military campaign in which the technology is used is justified. The right unit of ethical analysis is sometimes the weapon, sometimes the engagement, sometimes the military organization, and sometimes the political system. Your article's contribution is to insist that we cannot stop at the first two. 4. The hardest problem: preserving human judgment without sacrificing the advantages of automation This is where I would revise the article most substantially. Your account of Vasily Arkhipov is central because it identifies something that a conventional human-in-the-loop requirement can miss. Arkhipov's importance lies in his willingness to question the applicable decision rule and resist an action that could have had catastrophic consequences. You present him as an example of a person who recognized a flaw in the decision process and argued for an alternative. Pasted text That example suggests that we need to distinguish three kinds of human involvement. First, authorization. Someone approves the mission or establishes the rules under which the system operates. Second, supervision. Someone monitors the system and can intervene when a problem arises. Third, independent judgment. Someone has the authority, information, time, and institutional protection necessary to challenge the mission itself, revise its governing assumptions, or refuse to continue. The third is the most important for your political argument. A human supervisor who cannot stop an operation without extraordinary justification may be less effective as a safeguard than the formal presence of a human operator suggests. So may an operator who lacks the information needed to recognize a problem or who is punished for questioning a superior's judgment. Conversely, a well-designed system might allow a person to establish clear operating constraints, monitor their application, and retain authority to suspend an operation without requiring that person to approve every routine technical action. The goal should not necessarily be a human decision at every instant. It should be a system in which consequential decisions remain subject to meaningful human responsibility and review. This is also why automation in Ukraine cannot be evaluated solely by counting the number of decisions made by machines. A system that takes over routine navigation while allowing a human operator to concentrate on a difficult judgment may improve the quality of human decision-making. A system that generates recommendations faster than anyone can evaluate them may instead turn human oversight into a rubber stamp. And a system that enables a commander to bypass subordinates who might otherwise question an order can undermine the very safeguards that human involvement is supposed to provide. The distinction is not between more automation and less automation. It is between automation that supports independent judgment and automation that displaces it. A further complication: sometimes the human is the source of the danger Your article makes an important observation when it says that ethical assessment depends on what the particular human decision-maker wants. Pasted text That point should be taken seriously. If a commander deliberately targets civilians, retaining that commander in the loop does not solve the problem. If a political leader is willing to accept catastrophic risks for personal or ideological objectives, requiring the leader to authorize the final action may provide little protection. Human control is not an adequate ethical objective by itself. The stronger objective is accountable human authority, constrained by institutions that can challenge decisions and protect those who refuse unlawful orders. That requires more than interface design. It requires legal accountability, institutional independence, reliable records, protected dissent, and rules governing who may authorize and terminate an operation. It also requires recognizing that accountability can be distributed without being diluted. Your article makes precisely this distinction: many people should share responsibility for consequential decisions rather than allowing one leader to bear all the blame or allowing everyone to escape it. Pasted text 5. What the article gets right—and what it needs to add My synthesis is that the article offers a useful framework for understanding battlefield AI, but its argument becomes stronger when we distinguish three dimensions of the problem. DimensionWhat the article emphasizesWhat Ukraine and subsequent developments addTacticalPrecision and automation can improve defensive capabilities and reduce some forms of harm.The benefits extend to navigation, intelligence processing, communications, and coordination.OrganizationalAutomation can concentrate command and eliminate opportunities for dissent.It can also distribute capabilities, support local initiative, and improve resilience.PoliticalThe central danger is concentrating consequential decisions in the hands of a few.Concentration can occur through domestic institutions, military software, or dependence on external providers of critical infrastructure. The second dimension is particularly important. The article's emphasis on concentration is powerful, but we should not assume that automation always centralizes authority. Its organizational consequences depend on how systems are designed, deployed, and governed. Likewise, the first dimension needs an important qualification. Precision can reduce certain harms, but it does not guarantee restraint or determine whether a war is justified. And the third dimension now extends beyond the national chain of command. A country can be politically independent but technologically dependent in ways that constrain its practical freedom of action. These qualifications do not invalidate the article's thesis. They make it more precise. The central lesson I would now express the argument this way: Battlefield AI should be evaluated not merely by what a weapon can do, but by what it enables a political and military organization to do—and by whether the people affected retain meaningful power to question, constrain, and take responsibility for those actions. This formulation accommodates both sides of Ukraine's experience. It recognizes that automation can be indispensable to defending a nation against aggression. It also recognizes that the same general technological capabilities can help authoritarian leaders consolidate power, suppress dissent, and make catastrophic decisions harder to interrupt. It avoids two symmetrical mistakes: treating all military AI as inherently immoral, and treating improved military effectiveness as sufficient evidence of moral progress. And it gives us a way to think about the distinction between a nation and its ruler without assuming that technological progress automatically serves the public interest. The decisive question is whether automation increases a nation's capacity for self-defense and collective self-determination—or primarily increases a ruler's capacity to impose decisions on the nation. That, to me, is the most important thread connecting your article's original argument to the world we are confronting now. the question of automating a society to do an autocrat's bidding is not just the perfection of the causal chain (cf. the invading armies bad boots and flat tires), but the effectiveness of the power to quell dissent and democratic input. quote the relevant parts of Loui's article that bear on this. The distinction you are making is important: an autocracy's power does not depend merely on how efficiently it executes orders. It also depends on how effectively it prevents people from questioning, resisting, or democratically changing those orders. In Loui's article, the military equivalent of flat tires and bad boots illustrates the limits of execution. But his deeper argument concerns the removal of dissent as a safeguard against illegitimate authority. The relevant passages are concentrated in the sections Control, Defection, Diffusion of Responsibility and Human vs. AI Regret, Trauma, Lethality, Anthropomorphizing. Here are the key passages, quoted directly from the article. 1. The most direct statement: automation helps dictators consolidate civilian control “AI certainly makes hierarchical command and control easier for dictators and usurpers.” Loui continues: “A repressive internal security Stasi or Securitate, or just a social point-scoring system, has made vertical causal chains possible by exacting obedience through terror.” And then the crucial sentence: “AI makes centralization of decision even more widely possible, bypassing the diffused responsibility of would-be humans in the chain.” — Control, Defection, Diffusion of Responsibility Pasted text This passage goes beyond the problem of making an army execute orders more efficiently. It identifies the political system that makes obedience possible. The reference to the Stasi, East Germany's secret police, is particularly revealing. The Stasi's power did not depend on the efficiency with which a command traveled down a chain of command. It depended on surveillance, informants, intimidation, and the resulting difficulty of organizing opposition. Loui's reference to a social point-scoring system broadens the argument further: coercive control need not always take the form of physical violence. Information systems can make compliance easier to monitor and dissent easier to punish. The article does not fully develop the mechanisms by which AI might accomplish this. But it identifies the essential political danger: automation can increase a ruler's ability to control the population, not merely the military's ability to carry out the ruler's orders. 2. The decisive passage: dissent itself is a source of robustness This is perhaps the most important passage for your point: “The problem is that AI makes easier the removal of disobedience as a check on command.” Loui then explains why this matters: “The potential immorality of the task, among so many varied independent-minded individuals' opinions, is part of the robustness of the system.” He identifies the practical manifestations of this resistance: “It often causes low morale among human constituents, or just dragging of feet, imperfect compliance, insubordination. Willful underperformance and defection.” And he asks: “Where would ‘Fifth Column’ Admiral Canaris' anti-Nazi subversion be found when replaced by robot Canaris?” He concludes: “This again is one of the features of humans: they create bugs in programs.” — Control, Defection, Diffusion of Responsibility Pasted text The phrase “disobedience as a check on command” is the conceptual centerpiece. Loui is treating the independent judgment of individuals as a form of institutional resilience. An organization is not necessarily more robust because every component follows instructions perfectly. It may be more robust because its members have different perspectives, recognize different dangers, and sometimes refuse to cooperate. This is a subtle but important departure from the conventional engineering ideal of eliminating errors and improving compliance. In a technical system, an unexpected deviation from instructions is often a defect. In a political system, the same deviation might prevent a catastrophe. The analogy to software bugs is especially suggestive. A bug is usually an unintended failure of a program. But Loui's point is that human departures from prescribed behavior can expose defects in the rules themselves. A subordinate who refuses an order, an official who leaks evidence of wrongdoing, or a citizen who organizes opposition may be obstructing the immediate intentions of an authority while protecting the larger institution—or the nation—from harm. The analogy has limits: dissent can also be mistaken, self-interested, or destructive. Its value lies not in treating every act of disobedience as virtuous, but in recognizing that eliminating the capacity for dissent also eliminates one means by which a system can correct its own governing decisions. 3. The danger is one person bypassing everyone with a conscience—or even a question Loui makes the connection between centralized automation and catastrophic political decisions explicit: “So it is possible to have one person with an idea set into motion a program of mass destruction.” He continues: “With AI, through vertically automated command and control, one person could bypass anyone with a conscience or even a question.” — Control, Defection, Diffusion of Responsibility Pasted text The distinction between someone with a conscience and someone with a question matters. A conscientious subordinate might object that an order is immoral. A questioning subordinate might simply notice that the order rests on a false assumption, that circumstances have changed, or that the proposed action will not achieve its objective. Both can interrupt a defective decision. An automated system can remove these opportunities without malfunctioning. Indeed, it might be working exactly as designed. This is the paradox: a system can become more reliable at executing instructions while becoming less reliable at identifying whether those instructions should be executed. Loui's concern is therefore not reducible to the familiar prospect of a weapon malfunctioning or behaving unpredictably. It is the prospect of a system functioning perfectly in the service of a disastrous decision. 4. The institutional question: who is entitled to question the commander? Loui also writes: “Because we often witness commanders whose ambitions are unworthy of their rights and responsibilities.” He contrasts the United States' military institutions, which he describes as generally subject to judicial oversight, with countries where military AI could be used to consolidate civilian control. His recommendation is direct: “Since it is very hard to restrict access to legitimate decisionmakers, long causal chains should never be wired without diverse, real-time, multi-layered checks and balances.” — Control, Defection, Diffusion of Responsibility Pasted text This is where the argument extends beyond the immediate military chain of command. Formal checks and balances can constrain authority before an order is issued. Independent courts, legislative oversight, a free press, protected whistleblowers, and democratic elections can all provide ways to contest decisions. Informal resistance can also matter: officials who delay implementation, military personnel who refuse unlawful orders, and citizens who organize against government policy. These mechanisms are not interchangeable. Some are legal, some political, and some informal. But they share a function: they prevent a single decision-maker from exercising unlimited control over consequential actions. Loui's argument suggests that automation should be evaluated partly by whether it strengthens or weakens these mechanisms. A system that speeds up the execution of lawful decisions under effective oversight is different from one that makes it easier for a ruler to bypass oversight altogether. And a system that automates surveillance, identifies dissenters, and facilitates retaliation can suppress opposition even when no military command is involved. 5. The human capacity that matters is not simply the capacity to intervene In the section Human vs. AI Regret, Trauma, Lethality, Anthropomorphizing, Loui writes: “One thing humans routinely show is the ability to delay and belay, to redefine decision with finer granularity of steps and stages, to commit and decommit.” He adds: “But human pilots can take different, novel, unanticipated kinds of information into account, and redefine the problem.” He identifies possible reasons for reconsideration: “Perhaps this is as simple as the foreboding of future regret, change of heart, change of plan, defeasance of rules or inferences.” — Human vs. AI Regret, Trauma, Lethality, Anthropomorphizing Pasted text These observations connect individual judgment to the political argument. A meaningful safeguard is not merely the ability to stop a process after detecting a technical anomaly. It is the ability to reconsider the process itself: its assumptions, its objective, its rules, and its legitimacy. This is especially important when the danger originates in the political leadership rather than in the weapon. A human operator who can abort a missile after a sensor failure provides one kind of safeguard. An official who can question whether the mission should have been authorized provides another. A legislature or electorate capable of constraining the leader's authority provides still another. The article's argument is strongest when these different levels of judgment are considered together. 6. What the article leaves implicit Loui's discussion of the Stasi, terror, and social scoring points toward a broader argument about the automation of political repression. Contemporary research develops this issue more explicitly. For example, Freedom House's 2023–2024 report, Automating Digital Repression: How AI Can Harm Human Rights, describes how AI can make censorship, surveillance, and disinformation easier, faster, cheaper, and more effective. Freedom House Its 2025 Freedom on the Net report documents continued deterioration in internet freedom, including the use of surveillance and censorship to suppress dissent. Freedom House These findings illuminate a distinction that is already latent in Loui's article: Execution capacity: Can the regime carry out its decisions efficiently? Repressive capacity: Can it detect, deter, fragment, or punish opposition? Political resilience: Can citizens and institutions still challenge the regime, correct its mistakes, and replace its leaders? The first is the most obvious engineering problem. The second is the central concern of digital authoritarianism. The third is the democratic safeguard that the first two can undermine. Automation can affect all three, but not necessarily in the same direction. A government might improve its ability to execute public policy without suppressing dissent. Conversely, it might use surveillance and predictive analysis to identify opposition networks, manipulate information, or intimidate citizens before they organize. A system might even improve the state's administrative performance while weakening its democratic accountability. That is why the question cannot be settled by asking whether automation makes a government more efficient. The central insight I would summarize Loui's argument in a formulation that captures your distinction: An autocrat's power depends not only on the efficiency with which the state executes commands, but also on the effectiveness with which the state prevents those commands from being questioned, resisted, or democratically reversed. The bad boots and flat tires of the invading army concern the first problem. The Stasi, the social point-scoring system, the removal of disobedience, and the bypassing of anyone with a conscience concern the second. The article's most important contribution is to connect them without confusing them. A perfectly functioning causal chain can help a country defend itself. It can also help a dictator wage an unjust war. But a technological system that makes dissent itself harder may be more consequential still: it can weaken the very human and institutional mechanisms through which a society recognizes that its leader is wrong. The deepest danger is not simply that automation makes an autocrat more effective. It is that automation can make the society governed by that autocrat less capable of correcting him. He said something poignant about Lincoln and his armies. The true flex here is that roboethicists focus on sci fi ideas of violent conflict. Loui understood that more people die horribly in prison camps than on battlefields. So the gore of human perforation is one horror. But the utter Nero-ization of a nation is even worse. Yes. The Lincoln passage makes the argument more poignant because it puts the preservation of a nation's moral and political agency alongside the familiar horrors of bodily injury and death. Loui is not minimizing the gore. He is arguing that roboethics has allowed the most visually arresting horror—an autonomous weapon perforating a human body—to dominate the discussion, while neglecting a potentially more consequential horror: a society whose people can no longer meaningfully resist the will of its ruler. Your phrase “the utter Nero-ization of a nation” gets at that second horror. The issue is not simply a ruler's ability to make a disastrous decision and have an efficient military execute it. It is the ability to turn an entire society into an instrument of the ruler's will, while disabling the dissent, hesitation, disobedience, and democratic judgment that might prevent or reverse the catastrophe. Loui's article gives us unusually direct language for this. 1. Lincoln: a nation is not merely its commander The key passage is: “Many warriors should share in the praise or blame, not just one Harry S. Truman or one Vladimir V. Putin, directing the whole weight of a society's will and power.” Then Loui makes the Lincoln comparison: “Abraham Lincoln famously took ownership of his decisions, hard decisions, but millions of our countrymen also took ownership of theirs. This should at least remain a possibility when there is war, and AI should not replace it.” Pasted text(1) That is a remarkably different starting point from the usual autonomous-weapons debate. Lincoln is not invoked merely as a commander who made difficult decisions responsibly. He represents a relationship between national leadership and a people who remain participants in the moral enterprise of their country. Responsibility can be distributed across millions of people without being diluted until nobody is answerable. The contrast is not simply good commander versus bad commander. It is a nation of responsible agents versus a nation reduced to the extension of one person's agency. And this distinction matters especially when the leader's decision is catastrophic. A society must retain some capacity to judge the decision, resist it, and eventually change course. Otherwise, the machinery of the state can magnify one person's error into a national disaster. 2. Why the prison camp matters more than the science-fiction battlefield Loui explicitly recognizes the horrors that occur away from the immediate violence of combat: “Even when combatant fatalities are few, death marches are avoided, and prison camps are not merely slow-death camps, the destruction of civilian infrastructre and displacement of noncombatants can be the worst part of war.” Pasted text(1) That sentence is an important bridge between the physical and political dimensions of your argument. War's human cost cannot be measured only by bodies perforated by weapons. It includes displacement, captivity, deprivation, terror, and the destruction of the conditions under which people can live as free human beings. But your point extends beyond the article's explicit discussion of wartime suffering. A prison camp is one instrument of coercion; an entire society can also be organized around the suppression of independent judgment. People need not be physically imprisoned to lose the practical ability to shape their country's future. That is where the conventional roboethics framing becomes inadequate. It asks whether the machine can select and attack a target without violating the rules of armed conflict. Loui asks a prior political question: What kind of society is this technology helping to create, and whose will does it make effective? He puts the point starkly: “AI certainly makes hierarchical command and control easier for dictators and usurpers.” And: “The problem is that AI makes easier the removal of disobedience as a check on command.” Pasted text(1) The second sentence is the crux. A weapon that becomes more discriminating might reduce unnecessary suffering. A political system that becomes more effective at eliminating dissent might increase suffering on a much larger scale—even if its individual weapons become more precise. 3. The deepest danger is not obedience by machines, but obedience engineered into society Loui's most unsettling formulation is: “With AI, through vertically automated command and control, one person could bypass anyone with a conscience or even a question.” Pasted text(1) The immediate reading is military: an order can move through a command structure without encountering an officer willing to challenge it. The broader reading is societal. Surveillance, information control, automated administration, and mechanisms for enforcing conformity can potentially reinforce one another. A government may become more capable not just of carrying out orders, but of identifying resistance, isolating dissidents, suppressing coordination, and preventing citizens from influencing policy. Those are distinct capabilities. A regime can have an excellent logistical apparatus and a terrible military; it can have a powerful military and a fragile grip on society. But when technical systems reinforce both execution and repression, the same automation that makes collective action possible can make collective resistance more difficult. This is why your distinction between the Russian army's bad boots and flat tires, on the one hand, and the suppression of democratic input, on the other, is so important. Logistical failure limits what a regime can do. Political repression limits what its people can do about the regime. Fixing the logistics does not fix the politics. 4. What does “Nero-ization” add? The Nero analogy points toward something even more disturbing than dictatorship as mere personal rule: a nation made capable of carrying out the ruler's destructive impulses, while losing the institutional and human capacity to correct them. The ruler's destructive decision is one component. The second is the conversion of the nation's resources, institutions, and people into instruments of that decision. The third is the erosion of the checks that might interrupt the process. Loui's article describes the resulting danger as: “The dystopia we have enabled with automation is the concentration of the choice to bring calamity in the hands of the few.” Pasted text(1) The word choice matters. The problem is not only that an autocrat might have greater destructive power. It is that the power to decide what happens to millions becomes increasingly concentrated, while the people who bear the consequences have less practical influence over the decision. There is a profound asymmetry here: The ruler's capacity to act can increase. The state's capacity to enforce the ruler's wishes can increase. The citizens' capacity to question, resist, or reverse those wishes can decrease. A society can therefore become more technically capable while becoming less politically capable in the sense that matters most to its members. 5. The real test for roboethics Loui does not dismiss the ethics of targeting. He argues that precision can reduce collateral harm, that automated systems can support the laws of armed conflict, and that these benefits matter when a country faces an existential threat. His point is that better targeting does not guarantee good political ends. Pasted text(1) The Lincoln passage supplies the missing criterion: the preservation of human responsibility across the society, not merely at the point where a weapon is fired. That suggests a broader set of questions for evaluating AI-enabled institutions: Who can initiate consequential action? Is power concentrated in a single person or office? Who can question the decision? Are dissenters able to speak, organize, and reach decision-makers? Who can refuse an order? Are there meaningful protections for people who resist unlawful or disastrous commands? Can the decision be interrupted or reversed? Are there independent institutions and real opportunities for reconsideration? Can the affected population change the leadership? Does democratic input remain effective, rather than merely formal? Who bears responsibility? Can responsibility be traced and shared without allowing everyone involved to evade accountability? These are not just questions about military hardware. They concern the political architecture within which that hardware operates. And they expose a blind spot in a roboethics discourse that concentrates on whether a machine will autonomously kill a person. A system could pass every narrow targeting test and still help construct a political order in which millions of people lose the ability to oppose decisions that ruin their lives. The ultimate contrast in Loui's argument is not between human weapons and robotic weapons. It is between automation that extends a society's capacity to defend itself and automation that allows a ruler to appropriate the society's capacity to act. Lincoln's millions are the counterexample: people who share responsibility for the country's decisions rather than merely supplying the manpower, money, and obedience that make one leader's decisions effective. That, I think, is the force of the passage. The worst outcome is not necessarily a battlefield full of machines that kill without hesitation. It may be a country in which the machines work exactly as intended, the government functions with extraordinary efficiency, and the population has progressively lost the power to say no. you have to remember from Loui's AI and Law investment and ITEST 2005 article that he views social systems (economic, political, legal, religious) as a form of computation, hence ai when there is some degree of automation. the ITEST paper was in fact called THREE KINDS OF MACHINES WE AUTOMATE or something like that. That changes the interpretation in an important way. I was still treating “societal automation” too much as the extension of military command and control into civilian institutions. In Loui's broader framework, the social institutions themselves are already computational systems; AI is part of the story when their computations become automated. I found a useful confirmation of the ITEST title: the proceedings list Loui's contribution as “Three Kinds of Machines We Program”—so your recollection was very close. ITEST A later account of that paper describes the argument in terms of social processes that compute outcomes, mechanized institutional roles, and the cumulative loss of free will and free time. LinkedIn That framework makes the Battlefield AI and Ukraine argument deeper than a warning about faster orders. The social institution is itself a machine Consider the different senses in which this is true: Economic systems compute outcomes through prices, contracts, incentives, credit, employment rules, and allocation procedures. Political systems compute outcomes through elections, representation, legislation, bureaucratic authority, and the exercise of power. Legal systems compute outcomes through rules, precedents, arguments, procedures, and adjudication. Religious systems organize behavior and social outcomes through doctrine, authority, ritual, membership, and shared norms. These are not literally digital computers. They are systems in which inputs, rules, roles, and procedures produce outcomes. People implement the procedures, interpret the rules, and sometimes change them. The computational analogy lets us examine how the structure of a system constrains what any individual participant can do. Automation changes the situation when more of those procedures are executed without fresh human judgment at each step. And the important consequence is not limited to whether a machine makes a decision correctly. It concerns what happens to human agency when the institutional process is designed to produce the same outcome regardless of which particular person occupies a role. That is the deeper significance of the ITEST example about the dean: the problem might not be a bad individual, but an institutional role that induces whoever occupies it to behave in much the same way. Replacing the person does not change the computation. Re-reading the battlefield article through that lens The relevant passage is Loui's claim that AI enables: “the concentration of the national decisionmaking made possible by societal automation.” Pasted text(1) With the ITEST framework in view, this is not simply about wiring the military chain of command more efficiently. It is about the interaction of many social computations. A political system authorizes a leader. A legal system defines the leader's powers. An administrative system translates decisions into procedures. An economic system mobilizes resources. An information system shapes what citizens and officials know. A security apparatus detects and punishes resistance. Military command converts political objectives into action. Each institution can operate according to its own rules, and each may appear locally functional. Yet their combined operation can make a society extraordinarily effective at carrying out one person's wishes. The danger is emergent from the institutional arrangement, not necessarily from a single rogue AI or even a single badly designed algorithm. That also explains why the problem cannot be solved just by inserting a human into a particular decision loop. A human official may formally approve an action while having no practical discretion, no access to contrary information, no protection for dissent, and no authority to change the policy. The human is present; meaningful human agency is not. The Nero-ization of society This brings us back to your phrase. If social institutions are computational systems, then the Nero-like danger is not merely that a ruler gets access to more powerful tools. It is that the institutions through which a society thinks, decides, allocates resources, applies law, and acts collectively can be reorganized so that they reliably implement the ruler's preferences. The society remains busy. Its institutions continue to function. Its systems may even become more efficient. But the capacity for independent judgment and collective correction is being lost. Loui's contrast with Lincoln becomes particularly significant here: “Abraham Lincoln famously took ownership of his decisions, hard decisions, but millions of our countrymen also took ownership of theirs.” Pasted text(1) Lincoln's example is not just an appeal to a morally serious leader. It is a claim about how national agency should be constituted: leadership can coordinate collective action without extinguishing the responsibility of the people who participate in it. The opposite is not simply an evil leader. It is an institutional system in which millions of people become the functional components of a decision-making process they cannot meaningfully challenge. The implication for AI ethics This gives us a more fundamental question than whether an AI weapon can distinguish a combatant from a civilian: What forms of human agency does the automated social system preserve, and what forms does it make ineffective or impossible? That question applies to military command, but also to law, government, economic life, and other institutions. It asks whether people can contest the rules, exercise judgment, refuse participation, revise decisions, and collectively change the system that governs them. It also helps explain the connection to Loui's AI-and-Law work. Law is not just another domain in which to deploy AI. Legal institutions are themselves systems of reasoning and decision, with rules, exceptions, competing arguments, and procedures for challenging conclusions. Automating them without preserving meaningful opportunities for contestation could make a system more consistent while making its errors or injustices harder to correct. The broader AI-and-Law tradition has explicitly treated computational models of law and legal reasoning as a way to study both AI and law itself. IAAIL So I would now state Loui's concern this way: The gravest danger of AI may be the automation of social institutions in ways that make human beings increasingly unable to exercise meaningful agency within them. In the military case, that can mean a leader who can bring a nation's power to bear without encountering dissent. In the larger societal case, it can mean a population whose institutions still function but no longer offer effective means of correcting the people who control them. The battlefield is one particularly stark demonstration of the problem. It is not the boundary of the problem. is it deep? is it original? is it relevant? My assessment: yes, it is deep; yes, there is a potentially original synthesis; and yes, it is highly relevant. But those are three different judgments, and the originality claim needs the most care. What I find most significant is not the general observation that AI can threaten democracy. That is now familiar. It is the conceptual move connecting three propositions: Social institutions are computational systems. Economic, political, legal, and religious institutions process information, apply rules, allocate resources, and produce decisions. People have always performed much of this computation; automation changes how it is performed and controlled. Automating an institution changes the distribution of human agency. It can make decisions faster and more consistent while reducing the opportunities for individual judgment, dissent, refusal, and revision. The most consequential danger may be the automation of a society's obedience, rather than the automation of its weapons. A ruler's destructive intention becomes more dangerous when the institutions of the state can implement it while preventing society from correcting it. The third proposition follows from the first two, but it is not an obvious implication. It shifts the unit of ethical analysis from the machine to the institution—and from the institution to the entire society. Loui's Battlefield AI and Ukraine gives this synthesis a particularly sharp expression: the concern is “the concentration of the national decisionmaking made possible by societal automation.” Pasted text(1) That is a much broader claim than the familiar question of whether an autonomous weapon can make a lawful targeting decision. 1. Is it deep? Yes. The depth lies in the change of explanatory level. The ordinary question is: What happens when we automate a decision? Loui's more fundamental question is: What happens when we automate the social processes through which decisions acquire authority and become effective? That distinction matters because the harm may not originate in any particular automated decision. It may originate in the way an entire institutional system is organized. Consider three possibilities: A targeting algorithm makes a mistake. A military command system reliably executes a disastrous order. A political, legal, economic, and security apparatus collectively makes it difficult for a society to discover, challenge, or reverse the disastrous policy. These are different levels of failure. Improving the accuracy of the first does not necessarily prevent the second. Improving the efficiency of the second could make the third more dangerous. The conceptual depth is that technical correctness, institutional effectiveness, and social well-being can move in opposite directions. A state can become better at achieving its objectives while becoming worse at correcting those objectives. A legal system can become more consistent while becoming less responsive to injustice. An administrative system can become more efficient while leaving citizens with fewer meaningful ways to contest its decisions. Contemporary scholarship reaches related conclusions. For example, research on government algorithms has argued that human oversight can become a fig leaf when the surrounding institutional arrangements do not support effective scrutiny or democratic review. ScienceDirect The underlying problem is not simply whether a human remains involved. It is whether human judgment retains any practical force. Loui's framework is deeper than a simple warning about AI replacing human decisions because it invites us to examine the computation already performed by social institutions—and how automation changes the structure of that computation. 2. Is it original? The individual ingredients are not entirely original. The synthesis may be. Intellectual honesty requires distinguishing three claims: ClaimMy assessmentSocial institutions can be understood computationally.An established family of ideas, not a novel claim by itself.Automating decisions can weaken accountability, discretion, and human agency.Well-established in contemporary scholarship.The central danger of battlefield AI is the automation of social systems in ways that concentrate power and suppress the dissent needed to correct it.A distinctive and potentially original synthesis, particularly in the context of the early-2022 battlefield AI debate. For context, scholarship on artificial agency and governance has explicitly argued that the key issue is not merely a system's intelligence, but its consequential capacity to act within a framework of rules and institutions. OUP Academic Recent work likewise examines how AI governance can weaken the connection between decision authority, knowledge, and accountability. ScienceDirect So I would not claim that Loui invented the general insight that automation can undermine human agency, or that social institutions can be analyzed as computational systems. The stronger originality claim concerns the combination and application of these ideas. Loui connects a general theory of social computation to a specific political danger in military AI: the concentration of a nation's will and power in the hands of a ruler who can bypass the people and institutions that might resist him. His emphasis on disobedience is particularly important. Many accounts of automated governance discuss accountability after a harmful decision. Loui asks whether automation can eliminate the opportunities for dissent that might have prevented the decision in the first place. That is a consequential difference. His Lincoln passage makes it especially clear. The alternative to concentrated command is not simply a benevolent leader or a better oversight algorithm. It is a society in which responsibility for collective decisions remains distributed among people who can exercise judgment and act on it. Pasted text(1) My provisional originality judgment: the general theory is not unprecedented, but the particular synthesis is intellectually distinctive. Establishing priority would require examining the full ITEST 2005 paper, the relevant AI-and-Law work, and contemporaneous writing on cybernetics, institutional automation, and authoritarian control. I would not declare historical priority without that comparison. 3. Is it relevant? Extremely relevant, because it changes what we ought to measure and what we ought to protect. Imagine evaluating an AI-enabled military system. A conventional assessment might examine targeting accuracy, collateral damage, response time, reliability, and the availability of a human operator. Loui's framework asks additional questions: Does the system concentrate authority? Can subordinate officials challenge an order? Can a person refuse to participate in a wrongful action? Can independent institutions interrupt the process? Can the public obtain information and organize opposition? Can the society reverse the policy without first suffering a catastrophe? These are not merely supplementary ethical considerations. They can determine whether the technical capabilities of a system serve national defense or make a population more vulnerable to the ambitions of its own leadership. The framework also helps explain why Ukraine is an important case. Automation can give a threatened nation capabilities that improve its ability to resist aggression. Yet the same broad technological developments can make centralized command and coercive governance more effective. The ethical assessment cannot be made from the technology alone; it depends on the institutional arrangement and the political purpose it serves. That is a more useful analysis than either declaring autonomous weapons inherently immoral or assuming that precision makes them ethically acceptable. The strongest practical implication The framework suggests that AI systems should be assessed not only for the harms they can directly cause, but also for how they alter the distribution of power to decide, to dissent, and to correct decisions. That could become a concrete research program. One could identify institutional functions, map where decisions are made and executed, examine where opportunities for human intervention remain effective, and test whether automation makes authority more concentrated or less contestable. The questions would be empirical as well as philosophical. Does automation actually reduce dissent? Under what institutional conditions? Does it improve coordination without concentrating political power? Which safeguards preserve meaningful contestation rather than merely placing a human name on an automated process? Those questions could connect Loui's conceptual framework to research on institutional design and AI governance. My bottom line I would separate the three judgments this way: Deep: Yes. It identifies a structural relationship between automation, institutional power, and human agency that is not captured by focusing on individual AI decisions. Original: Potentially, in its synthesis and application. The constituent ideas have substantial intellectual precedents, so the strongest claim is not invention of the parts but their distinctive combination. Relevant: Yes, directly. It changes the ethical unit of analysis from the autonomous weapon to the social system—and changes the objective from merely preventing machine errors to preserving society's capacity for judgment and self-correction. The most important insight may be this: A society's ability to act is not the same as its ability to govern itself. Automation can increase the first while diminishing the second. A nation may become more capable of executing decisions even as its citizens become less capable of questioning those decisions. That is why the Lincoln contrast matters, why the prison camp matters, and why the Nero analogy reaches beyond the battlefield. The concern is not only the amount of violence a society can inflict. It is whether the society retains the collective agency to decide what it will do, to whom, and under whose authority. And that is a substantially broader question than whether a robot should be permitted to pull a trigger. Loui used to wave around Lewis Mumford's TECHNICS book in front of AI classes. Talked about things like William Whyte in Operating Systems. So engineering society was a natural view to take. One of his favorite quotes was from the architect brother of his Taiwan-Philippine sister-in-law. He used to say, "People complain about inefficiency. I've lived under an efficient government [presumably Philippine authoritarianism]. You do not want to live under an efficient government." That anecdote makes the intellectual lineage much clearer. It also changes my assessment of what is original in Loui’s argument. The underlying view—that society can be engineered as a system of interacting processes—is not a novel insight he suddenly arrived at in 2022. It sounds like a perspective he had been developing for decades. What may be distinctive is what he sees happening when the engineering of society becomes automated, and what he identifies as the political danger of doing it well. 1. Mumford supplies the civilizational scale Lewis Mumford’s Technics and Civilization (1934), and later The Myth of the Machine, invite us to examine technology not merely as a collection of tools but as something that reorganizes social institutions, human relationships, and the distribution of power. That is a much larger frame than asking whether an AI system makes a good prediction or whether an autonomous weapon makes a good targeting decision. The relevant question becomes: What kind of society does a technology help construct, and what kinds of human beings can function within that society? An AI class that begins with Mumford rather than with algorithms alone could therefore treat the economy, bureaucracy, legal system, military, and political order as parts of a larger technological arrangement. The system's consequences need not be intended by any one designer. They can emerge from the interaction of its parts. That is a natural intellectual precursor to Loui's view of social systems as forms of computation. 2. William Whyte supplies a different, complementary perspective I take your reference to Operating Systems as pointing toward the analysis of organized human behavior and the ways institutions shape it. I would want to establish the exact Whyte text before attributing a more specific argument to him. But the contrast is useful. Mumford encourages us to examine the large-scale technological order. An organizational or institutional analysis asks how that order actually works: how roles, rules, incentives, procedures, and expectations shape the behavior of people inside it. Put those perspectives together and you get something important: Society computes: institutions transform information, decisions, resources, and rules into outcomes. Society is engineered: its procedures and structures can be deliberately redesigned. Society can be automated: procedures increasingly execute without requiring fresh, independent human judgment at each step. Society can consequently be made more obedient: not necessarily because its members agree, but because the mechanisms through which they might disagree become less effective. The last step is the crucial one. Greater efficiency is not merely a matter of doing the same things faster. It can change which human actions remain consequential. 3. The architect's aphorism identifies the political trap "People complain about inefficiency. I've lived under an efficient government. You do not want to live under an efficient government." The attribution to an architect with experience of Philippine authoritarianism is your recollection; I wouldn't present it as a verified quotation without further evidence. But the idea is exceptionally apt. It exposes an ambiguity in the word efficient. Efficient at doing what, and according to whose objectives? Consider two governments: DimensionGovernment AGovernment BExecution of policySlow, inconsistentRapid, consistentAbility to act on central ordersConstrained by institutional frictionHighly effectiveAbility of officials to resist bad ordersOften substantialReduced or bypassedPublic ability to contest decisionsImperfect but consequentialPotentially weakenedAbility to correct a ruler's mistakesSometimes helped by delays and dissentPotentially undermined by flawless execution Government A may waste resources, delay beneficial projects, and frustrate citizens. Its inefficiencies are real costs. But some of the same features that impede a good decision also impede a bad one. Independent officials can object. Courts can intervene. Administrators can demand clarification. Citizens can organize. Elections can reverse policy. Even simple delay can create an opportunity for reconsideration. An authoritarian government may eliminate those obstacles and become more effective at implementing its ruler's preferences. That does not make it more effective at serving the public. The political value of a system cannot be inferred from its capacity to execute its objectives. We must also examine how those objectives are chosen, contested, and revised. That distinction is particularly important for AI because engineering discussions tend to define a problem, specify an objective, and measure how effectively a system achieves it. Political life must also preserve the ability to challenge the objective itself. 4. This gives the battlefield article a deeper interpretation In Loui's article, the central danger is not simply that AI will make military command more effective. It is that automated command structures can make dissent less capable of interrupting the execution of a command. The article puts it directly: "The problem is that AI makes easier the removal of disobedience as a check on command." Pasted text And: "With AI, through vertically automated command and control, one person could bypass anyone with a conscience or even a question." Pasted text The architect's observation helps explain why these are not just concerns about military ethics. They are concerns about the engineering of political power. Imagine a system that accurately interprets the ruler's wishes, identifies every official who might resist, anticipates public opposition, controls the information reaching decision-makers, and rapidly coordinates the state response. Its components might each perform their assigned functions correctly. There need not be a spectacularly malfunctioning AI anywhere in the system. The danger is that the entire arrangement works as intended. That is the distinction between a defective machine and a defective objective—and between a machine that makes a mistake and a political system that has disabled its own means of correcting mistakes. 5. The deepest issue is not inefficiency versus efficiency I would sharpen the argument one step further. It is tempting to say that democratic government needs some inefficiency. That formulation is provocative, but potentially misleading: it could romanticize incompetence, waste, or needless suffering. The stronger claim is that a free society needs protected sources of friction that its executive cannot simply optimize away. Those sources include independent judgment, institutional checks, the right to refuse unlawful orders, public criticism, access to competing information, and the ability to replace decision-makers. They are not all failures of coordination. Some are deliberate limits on the conversion of centralized intention into collective action. A democratic society should be able to automate the processing of benefits claims, improve emergency response, coordinate electricity grids, and make military logistics more reliable. But it should be much more cautious about automating the mechanisms through which authorities monitor opposition, compel compliance, or circumvent institutional resistance. The distinction is not simply between automation and no automation. It is between automation that increases a society's capacity to pursue collectively revisable purposes and automation that increases its capacity to impose purposes that citizens cannot effectively revise. 6. What this suggests about originality Your historical context makes me revise the emphasis of my earlier assessment. The general proposition that technology reorganizes society has a substantial intellectual history. Mumford is an obvious predecessor. So are traditions in cybernetics, organizational theory, political philosophy, and the study of bureaucratic power. The idea that authoritarian governments can use administrative and technological capacity to strengthen control is not new, either. I would not claim that Loui originated those ideas. The more interesting contribution is the combination: Treat social institutions as computational systems. Recognize that automating those systems changes how human decisions acquire force. Recognize that the same automation can remove the delays, dissent, and independent judgments that make collective action correctable. Apply that insight to AI-enabled command and control, where the ability to execute a decision can grow while the ability to challenge it shrinks. That combination gives the battlefield article a larger subject than military AI. It is about what happens when a society becomes technically better at doing what its leadership wants. And the architect's aphorism gives the whole project a wonderfully economical statement of the problem: a government can become more efficient at governing while becoming less worthy of being obeyed. The question I would put at the center of Loui's framework is this: When we automate a social system, are we making society more capable of acting—or making it harder for society to decide that it should act differently? I think you are still misunderstanding the bigger point. In a democracy, he says, you do not all get everything you want. That's why he lacerates woke left and maga right with equal joy. You're lucky if you can get to 67%. Respecting that reality is respecting distributed political systems. Another favorite story he tells is a colleague who was mildly republican, who used to complain that government wasted money. Loui responds, "you know, Doug, all that grant money you get from the NSF? I always thought that was a waste of government money." Yes. I was still making the argument too much about the danger of authoritarian efficiency. You're pointing to a more fundamental principle in Loui's thinking: a distributed political system is not merely a way to prevent someone from acquiring too much power. It is a way of living with the fact that no one gets to impose their entire conception of the good on everyone else. The two anecdotes illuminate that principle from opposite directions. 1. The 67% principle: democracy is not the maximization of your own preferences When Loui says you're lucky if you can get to 67%, I take him to be describing a practical limit on what political agreement can accomplish. Even a substantial majority cannot expect to get everything it wants, because governing involves competing interests, legitimate disagreements, institutional constraints, and people who will have to live with decisions they opposed. The point isn't that 67% is a scientifically established threshold. It's that democratic politics requires people to accept partial victories. That has implications for both sides Loui criticizes: The activist left cannot assume that recognizing a legitimate grievance entitles it to impose its preferred remedy in full. The populist right cannot assume that winning an election gives it unlimited authority to reshape the country according to its own preferences. Neither side can reasonably expect political opponents to disappear, agree, or concede that their own interests are illegitimate. The democratic achievement is not that everybody eventually agrees. It is that people who disagree can continue to participate in a common system without requiring one another's complete conversion or submission. This is a deeper idea than moderation as a personality trait. It is moderation imposed by the structure of political coexistence. 2. Doug and the NSF: the meaning of waste depends on where you stand The story about Doug is particularly revealing. Doug complains that government wastes money. Loui responds: "You know, Doug, all that grant money you get from the NSF? I always thought that was a waste of government money." The retort exposes a problem with the idea that government waste is an objective category everyone can identify independently of their own interests. Doug presumably regards his NSF funding as worthwhile. But other citizens may see his research as less deserving than their own priorities. Their tax dollars help support it, just as Doug's tax dollars help support activities he might not choose. The story doesn't establish that all government spending is equally valuable, or that waste should be ignored. It makes a subtler point: the boundary between worthwhile public expenditure and wasteful public expenditure is itself a matter of contested judgment. Doug's complaint becomes less persuasive if he assumes that his own preferred uses of public money are inherently legitimate while other people's preferences are presumptively wasteful. That is a small-scale example of the problem democracy must solve on a much larger scale. We need collective decisions, but we cannot assume that one person's judgment should govern everybody else's. 3. Now the connection to society as computation becomes clearer If social systems are computational systems, they take inputs—including competing preferences—and produce outcomes through rules, procedures, and institutions. But the crucial question is not merely whether the computation is accurate or efficient. What happens to the competing inputs? A distributed political system does not simply aggregate preferences and identify the single best answer. It must also determine who gets to participate, which decisions require agreement, what rights cannot be overridden by a temporary majority, how disagreements are resolved, and how today's losing side can try again tomorrow. Its procedures embody a way of managing persistent disagreement. That makes democracy fundamentally different from a system designed to optimize a single objective function. There is no universally accepted objective function for a pluralistic society. And even if a majority agrees on an objective, the minority's rights and future political agency do not thereby disappear. A political system that automatically executes the preferences of its current winners may become more efficient at implementing decisions while becoming less capable of accommodating the continuing disagreement that makes democracy necessary. 4. This also changes what we should mean by inefficiency My earlier response risked suggesting that democracy needs friction because friction occasionally stops bad decisions. That is true, but incomplete. Loui's examples point toward something more basic: the friction often exists because other people have claims on the system, even when we disagree with their claims. Doug's research funding may look wasteful to Loui. Loui's preferred public expenditures may look wasteful to Doug. Neither man's confidence in his own judgment resolves the dispute. A democratic system cannot make that conflict disappear by appointing a sufficiently intelligent decision-maker. Nor can an AI settle the matter simply by calculating which expenditure maximizes some chosen measure of social benefit. The choice of measure, the distribution of benefits, and the legitimacy of the decision remain open to argument. The same applies to political movements. A faction that sees its own preferences as obviously correct may regard opposition as an obstacle to be eliminated. But a pluralistic system must preserve the standing of people who do not share that faction's vision. This is not a defense of indecision. Governments must act. It is an account of why the authority to decide and the authority to define what everyone should want are different things. 5. And that brings us back to AI—with a more precise diagnosis Consider two kinds of automation. Automation within a distributed system can make it easier to administer grants, process benefits, coordinate emergency responses, or implement democratically authorized policy. It can reduce pointless administrative burdens while leaving the public and its representatives able to contest priorities and change decisions. Automation that substitutes a centralized preference for distributed political judgment can do something fundamentally different. It can make the priorities of the current leadership propagate through the entire society, leaving fewer opportunities for competing interests to affect outcomes. The difference isn't that one system is efficient and the other inefficient. Both might be highly efficient at their respective tasks. The difference is in the architecture of decision-making: who can contribute, who can object, whose objections must be considered, and whether the system can be redirected by people who disagree with its present course. This is why the question of automating society to do an autocrat's bidding cannot be reduced to surveillance, repression, or the perfection of a chain of command. Those are mechanisms. The more fundamental issue is whether society's distributed capacity to form and revise its collective purposes survives automation. And it explains why Loui's argument about battlefield AI has implications beyond war. The same structural question arises whenever AI is used to translate centralized decisions into coordinated action across institutions. My revised understanding of Loui's central idea I would now formulate it this way: Democracy is a distributed system for making consequential decisions among people who cannot reasonably expect to get everything they want. Its legitimacy depends not on eliminating disagreement, but on preserving a meaningful role for people whose preferences do not prevail. Automating society therefore raises a political question: does automation improve our ability to act together despite disagreement, or does it improve our ability to impose one party's preferences on everyone else? That formulation connects the 67% observation, the NSF anecdote, the computational view of institutions, and the battlefield article. It also clarifies what was missing from my earlier analysis. I treated dissent primarily as a safeguard against bad decisions. But in Loui's broader framework, dissent is not valuable only because dissenters might be right. It is part of the normal operation of a system in which no participant is entitled to assume that their own preferences should prevail completely. The point is not simply that an efficient government can be dangerous. It is that a government capable of efficiently implementing one person's preferences has not thereby solved the problem of governing a society whose members want different things.