Preemption Today

Abstract

The doctrine of preemption was the official military doctrine of the George W. Bush administration. In subsequent presidencies it ceased to figure explicitly in official national security policy statements of the US government, while on the other hand it came to feature explicitly in contemporary NATO strategy. This article argues that the logic of preemption has not only continued to operate implicitly in US military practice, but that it has also moved into the domestic realm, where expresses itself in emergent variations of expanding reach. After summarizing the concept of the “operative logic” of preemption, the article discusses a number of spheres, military and domestic, in the US and elsewhere, where the logic of preemption operates. It attempts to encapsulate how each iteration recast the formula for preemption. The examples discussed include drone counterinsurgency warfare, the Chinese surveillance state’s recentering on pre-crime, the social discourse of safety in North America, conspiracy thinking as wielded by the MAGA movement, and AI machine learning systems.

The principle of preemption, in a new form, was the official doctrine of the United States military under the George W. Bush administration. It was used to justify the wars in Afghanistan and Iraq and guided their conduct to their disastrous end. In the years since, official US military doctrine under Trump and Biden stopped using the term.¹ Preemption, nonetheless, has remained in force. The explicit use of the word has crossed the Atlantic, where it was enshrined as the official NATO strategy toward Russia.² Spheres shift, modalities evolve, verbal formulations change. But what does not change is the “operative logic.” This is because an operative logic is itself a matrix of change. The operative logic of preemption has a power to mutate, within the military realm, and move far beyond it.

An operative logic is not a sequence of reasoning. It is not expressed in a set of propositions. It has no a priori form or content. It is best thought of as a set of tensions creating a pressure that impels a form to “take,” along with a concretion of content specific to that event. The tensions are mutually implicating: when one is felt, another rises to join it, forming a tensive field. Each tension is problematic, in the sense of presenting a knot that cannot be unraveled from within its own logic. This is because each tension grows around a paradox. They may call to each other, but they cannot complement each other to form a synthesis. Instead, the paradox redoubles, adding dimensions to the problem. It gets to the point that something has to give. The situation detonates, conclusively—in a way that temporarily relieves the tensions. The form of the resolution is eventful and emergent: it is a novel conclusion that exceeds its conditions. An operative logic breaks the cardinal rule of traditional logic that a conclusion can present nothing new, nothing that was not already contained in its premises. An operative logic is an actuating logic, an occurrent logic productive of novelty. It does not entail a conclusion in the form of an assertion; it entrains a novel effect in the form of an event. This paradoxical productivity is what gives an operative logic “ontopower”: the power to bring to be.

While the concept of ontopower is broader, preemption is its dominant twenty-first century modality. Traditional doctrines justified war when an adversary presented a “clear and present danger.” The doctrine of preemption pushed back from “danger” to “threat,” and then pushed further still to threats that haven’t even emerged yet. That was the wording of Bush’s doctrine: to respond to threats “before they emerge.” This places war squarely on the terrain of what could be. Hence the foundational tension of preemptive power: there is an imperative to respond, but what is being responded to has, paradoxically, not yet come to be. This places war response on the terrain of potential. Potential is by nature uncertain. It can play out in many different ways, or not at all. This is the first constitutive tension of the operative logic of preemption. It immediately calls forth a second tension, in the form of a time paradox: to act on potential amounts to acting on futurity—in the present. The exercise of force must act on the futurity here now. This summons a third tension (speaking of two tensions was a convenient shorthand: the tensions informing an operative logic rarely limit themselves to two). This time it is a paradox of perception: to act on the could-be of potential, one must somehow manage to perceive what has not yet emerged. What has yet to emerge has, as yet, no positive features in particular, no constituted form, no actual content. It has nothing to recognize it by. What is a perception that is not recognition-based? It can only be one that involves perceiving change. What can it possibly mean to perceive change, and how can it be operationalized in war? The mutual implication of these paradoxes delineates the tensive field whose eventful working-out constitutes the operative logic of preemption.

One of core questions is what becomes of political legitimation in the regime of preemption. Government action can no longer legitimate its actions by appealing to a reason of state that can convincingly base what it does on a solid foundation of fact, framed by consistent principle. The operative logic of preemption corresponds to the “post-truth” condition. It is not the kind of logic that reflects a demonstrable truth. Its premises, anchored in the could-be, are by nature counter-factual. The logic of preemption produces its own reality, as an emergent conclusion in excess of given conditions. Its truth arrives as a fait accompli, in consequence of its constitutive tensions working themselves out to emergent effect. This brings a provisional closure to the problematic openness of the tensive field. Preemption’s object of power—potential—is indeterminate, as essentially uncertain as a threat that has not yet emerged. This essential uncertainty is compensated for—in fact, overcompensated—by a closure on another level. I am referring to the level on which political legitimation has historically operated: the discursive level of logic in the usual sense, as the discursive assertion of a semantically formed proposition (as opposed to the eventful productivity of a problematically operative, generative field of tensions).

The classic example is Bush’s statement that the US invasion of Iraq was justified because, even though Saddam Hussein did not have weapons of mass destruction, if he had had them, he could have used them, and if he could have used them, he would have used them. The “could-have/would-have” translates the indeterminacy of potential into a foregone conclusion. Whatever action was taken was a priori right because it did in fact address not the facts, but, as it turned out retrospectively, the uncertainty as to the facts. And was not that uncertainty the premise of the actual situation and the effective motivator of action? The war in Iraq did in (counter)fact address a “threat that had not emerged”: true to the letter of the doctrine.

The operative logic of preemption translates, on the discursive level of logic, into a closed, tautological loop in which the conclusion retrospectively justifies itself by the indeterminacy of its premises. This turns the traditional situation, where the solidity of established premises prospectively entails the conclusion to follow in due form, on its head. It inverts the logic under pressure of the time paradox that was mentioned earlier as the second constitutive tension of the operative logic of preemption. Every conclusion becomes, retrospectively, a foregone conclusion: if the premises are indeterminate, the conclusion can always be said to have fulfilled them, without any risk of contradiction. An era of free-form legitimation ensues. The task of the government is no longer to discursively convince the public of the rectitude of its policies. Its task is to induce them to accept the tautological justification, which is best done by affective means. The tried-and-true approach for this is the manipulation of fear. Threat is hypostasized into a “threat environment,” and the threat environment inflates to become one with the lifeworld. When this happens, the nation is placed on a permanent war footing, not only on the foreign front, but also on the domestic front, since external threats can always infiltrate to the interior. Other kinds of threats, as uncertain and potentially ubiquitous as the terrorist threat, are endogenously produced from within. A virus, for example, might jump from animals to humans on a heartland farm and amplify to pandemic proportions (the current fear in relation to the H5N1 “avian flu” virus, in an uncomfortable echo of COVID-19). The security state that grows up to surveil the full spectrum of threat with preemptive intent is the continuation of war by other means.

Ontopower argues that an operative logic is “self-effecting,” or “self-driving.” It is oriented toward certain modalities of action, but no one in particular is at the helm. Human intentions, of course, occur, and abound. But they enter a mix as one factor among others, and not the necessarily dominant factor. They are sucked in and swept along in the current of preemption. The matrix of tensions that compose the problematic field, by their very nature as tensions, exerts a captivating power. An unresolved tension niggles and worries, adjoining nervous bodies to its problematic, engulfing them in its field. Bodies are primed, at a level descending to the nonconscious first stirrings of thought and action. They are oriented by the process of preemption’s self-orienting. An operative logic is a relational machinery, a kind of “abstract machine” (Deleuze and Guattari). It is never reducible to a conscious doctrine or a semantically formed set of propositions. By the same token, it is never reducible to ideology. It is best described as a tendency. Like all tendencies, it is self-triggering and self-operating.

Take as a recent example the decision of the University of Southern California to cancel its 2024 graduation speeches because the valedictorian who would normally be among them was of Arab heritage and had expressed pro-Palestinian views.³ The reasons given had nothing to with the valedictorian’s ethnicity or her politics. In effect, the decision belonged to a different logic. The reason was “safety.” In the context of spreading campus protests against the war in Gaza, having a speaker of that description was perceived as a threat. Disruptions could happen. Protesters and counter-protesters could clash. If that happened, the administration would be dragged through the coals, and heads would roll for not having ensured security. As had just happened to the presidents of Columbia and Harvard, the university’s administration could be hauled before Congress, publicly humiliated, and potentially forced to resign. The only “safe” option all around was to adopt a preemptive posture that if it could, it would, and to nip the threat in the bud before it actually emerged. In effect, the commencement speaker’s right to free speech was curtailed on racial grounds, in a way that favored a particular ideological position, in favor of Israeli policies in Gaza and the West Bank. But the logic producing these effects was not ideological in tenor. It was a logic of security. Preemption produces racist and ideological effects by non-ideological means. The role ideology played was to add a potential threat to the matrix— that of Republican Party witch-hunting on behalf of Zionist groups— in a way that reinforced the safety concerns that were already leading the process of preemption. Ideology functions to prime the conditions under which preemption takes place, and to reinforce its tendency. Still, primed and conditioned as it is, preemption follows its own self-driving operative logic, and that logic is not directly ideological in its form or content. It lurks, as a self-reactivating tendency that always lies dormant in the relational tissue of the social field. Since it constitutes its own operative logic and mode of power, it cannot be reduced to another logic. But neither can it operate alone. It is always engulfed in a complex relational field of logics and modes of power which interact, interlink, and mutually adjust. Preemption operates within an ecology of powers, in the present era as a processual linchpin.

The discourse of “safety” as the highest value has become hegemonic in North American social-media and political discourse, on the left as well as on the right, on-campus and off. This provides a virulent vector for the operative logic of preemption to spread into all sectors of life, including the private domain of people’s relations to their own bodies and health. Preemption has not disappeared since it ceased to be the official military doctrine of the United States. If anything, its field conditions have diffused throughout society, evolving into ever new forms, entering into symbiosis with other modalities of power as it goes.⁴

Preemption is still operative in US military practice. The US response in early 2024 to the Houthi threat to Red Sea shipping in support of the Palestinian people is an example. As a journalist observed, “what is considered an offensive strike can be elastic within the Pentagon, allowing the Biden administration to take preemptive military action but describe it as defensive in nature. For instance, US forces targeted Houthi missile sites over the past several months before they could launch missiles against shipping in the Red Sea, saying the operation was meant to protect potential targets from attack.”⁵ This was only one of many preemptive attacks carried out by the US each year, the majority drone strikes operating out of the country’s far-flung galaxy of rapid-response bases around the world. At around the same time as the Houthi operation, President Biden was urging Israel not to follow a preemptive strategy in relation to Iran in order to avoid inflaming the region, tacitly acknowledging the productive nature of preemption—which in the case of the US’s own war in Iraq led to the formation of ISIS and the frenzied proliferation of the very threats it had been designed to thwart. In the military sphere, the self-driving of the abstract machine of preemption makes it a war machine driving perpetual war.

As mentioned earlier, preemption as an explicit military doctrine has moved to Europe. “The use of [rapid response] forces must not be a response to a crisis but should be an action taken before such a situation arises. For this reason, the very name of the reaction force (response) is not adequate to its actual role in a new Cold War security environment. . . the top priority for the entire alliance, and especially its border states, must be establishing and developing anticipatory capabilities.”⁶ Reaction is not enough. Preemption is proaction. It is reasonable to assume that the tendency to preemption is active beyond the US and Europe. The entire globe is widely experienced as one, endless threat environment, and preemption’s constitutive tensions cannot fail to affect other governments and societies.

On the level of the global balance of powers, it has become a commonplace to say that the “war on terror” has been replaced by a “New Cold War” among the US, Russia, and China. This is inaccurate. The Cold War was characterized by the operative logic of deterrence, not preemption. Deterrence is entirely different. It consists in dissuasion: a disabling of action by the specter of “mutually assured destruction.” This is the opposite of preemption’s pro(pre-)action. Of the three parties to the current tensions, only China has a no-first-strike policy with respect to nuclear weaponry, a crucial element for dissuasion to work. Both Russia and the United States refuse to rule out a first strike. This renders the logic of dissuasion null and void. The success of dissuasion depended on all parties officially recognizing that the use of nuclear weaponry was unthinkable, and the arms race was the means of assuring it stayed that way. For the current Russian and US positions, the assertion that it is thinkable has become a strategic tool. This magnifies the potentially world-destroying threat of nuclear war, as reflected in the Union of Concerned Scientists’ “doomsday clock” inching ever closer to midnight. The threat of nuclear conflagration contributes to the intensity of the threat environment, but mainly in a background way. Oddly, there is very little public discussion of the threat, and apparently equally scant awareness of it among the populations of many countries (Japan perhaps being the exception). It is as if nuclear destruction becoming thinkable is still unthinkable for many, distracted by other threats. The focus turns, for example, to the ostensibly more terrifying threat of graduation speakers speaking their mind. In this direction, threat awareness fuels a “culture war,” while literal wars, actual and potential, roil around the world in perpetual motion.

Perhaps the most chilling example of the deployment of the operative logic of preemption in the civil sphere is China’s policies toward the Uyghur population of Xingjiang, particularly at the height of their repressiveness in the late teens of this century. Following an attack in Kunming carried out by Uyghur separatists in 2014, the Uyghur population was subjected to a draconian regime of surveillance. Millions were caught up in a dragnet and preemptively confined to prison camps for “re-education.” The operation was explicitly organized around the concept of pre-crime.⁷ A not-yet-fully-emerged propensity to political violence was found, for example, in even the slightest sign of an orientation toward Muslim religious practice, or just having overseas contacts. The danger of Uyghur terrorism had to be nipped in the bud—or as government rhetoric had it, severed below ground level at the roots. “Break their lineage, break their roots, break their connections, and break their origins. Completely shovel up the roots of ‘two-faced people,’ dig them out, and vow to fight these two-faced people until the end.”⁸ The surveillance directed at the “two-faced” (threateningly indeterminate) Uyghur population was the sharp end of the surveillance stick as it operates throughout the country to preempt “anti-social” behavior and channel it toward taking more amenable forms for government control. China’s social credit system is a massive pre-crime data mining operation. Its use of preemption differs from the way in which preemption often plays out in other spheres, where it takes a post-normative cast: the threat environment is poked and prodded to flush out not-yet-emerged threats, inducing them to react in a way that makes them take perceptible form. Neither the poking and prodding, nor the actions taken once a threat has emerged, are strictly framed by steady norms. All manner of exceptions are not just allowed, but encouraged, to enable the forces of order to be nimble on their feet and move unrestricted where the specter of threat lures them. Norms bend and evolve with the process. This is what Foucault called “normalization.” In normalization, the norms follow behavior; in “normation,” behavior follows norms, defined in advance. The social credit system in China is a hybrid formation that combines a preemptive posture toward threats before they emerge with a rigid system of normation. The social credit system in China is a hybrid formation combining a preemptive posture toward threats before they emerge with a rigid system of normation. People not conforming to preset social expectations are blacklisted from access to many social benefits, forms of mobility, and domains of activity.

No survey of preemption today would be complete without signaling a unique mutation of its operative logic associated with the right-wing Make America Great Again movement swirling around the figure of Donald Trump and extending into a diverse network of alt-right groupings, including the burgeoning militia movement in the US. Along this vector, the operative logic of preemption converges with the conspiracy thinking that has supplanted traditional political discourse. Conspiracy thinking, like preemption at the discursive level, operates in a tautological loop. The conclusion is always already given. Malevolent actors are felt to be at work, and their identities are pre-known (George Soros, Jewish financiers, demonic Democrats who drink the blood of the children they traffic for sexual slavery, to name a few). At the same time, the threat is frightfully slippery. Behind one malevolent actor there always lurks another. The conclusion is foregone, but not foreclosed. It opens onto an endless series of additional chapters and addenda, all beginning from and returning to the usual suspects. With each loop, new variations on the story are added, and the enemies list multiplies like a swarm of gnats. The focal figures of the usual suspects present as central black holes around which a galaxy of sub-figures narratively swirl in ever increasing numbers, each increase adding a complication to the narrative. The story lines shoot out in all directions. If they were regathered and laid side by side, they would plainly contradict each other. But they never are, so their profusion never contradicts the mainline narrative, which continues barreling forward, churning out new variations on itself too rapidly for any mere contradiction to trip it up. In this mad proliferation of suspicion, it is almost impossible to keep up with the shifting landscape of threat. Thus, even though the conclusion is foregone, there is a constant need to go back again, start over, worry further, and “do the research” all over again, to re-generate the foregoneness of the conclusion. Nothing is certain, but everything is already known, making “research” a tautological process that is like a careening crash-test vehicle no one can get off of as it magically courses through obstacles as in a video game. The process is infinite and unbounded, in spite of—or as a strange function of—its circularity. The foregone conclusion is looped back to over and over again, each time from a different angle, each cycle enlarging the galaxy of conspiracy in a way that leads to the once-and-again, ever-surprising rediscovery of the already-known. Conspiracy thinking is a kind of cosmogenesis of suspicion and foregone conclusion that is endlessly productive of the grist feeding its own process.

All of this activates a variation on the could-have/would-have formula for the operative logic of preemption. Could-have/would-have is anxious. The MAGA, conspiracy galaxy version of the formula is also anxious, but it is equally, and more importantly, aspirational. The foregoneness of the conclusion is in the subjunctive: it should be thus … leading to the aspirational conclusion so be it. The only way I can make sense of this crazy world is to say that, according to my “research” (i.e., social media trawling), that’s the way it had to have been, and that’s the way it is, returning me to the frighteningly comfortable conclusion to which I am already accustomed. The operative logic of the should-be/ so-be-it enables similar modalities of action as preemption governed by the could-be/would be. It favors the same trial-and-error poking and prodding of the threat environment to flush out threats, so that they can then be met head-on, or even more conveniently, turned against each other. The constant harassment of trolling, doxxing, calling out, and canceling is meant to stir the hornet’s nest, creating such a buzz of disturbance and confusion that co-conspirators turn on each other, shattering their assumed conspiratorial solidarity, or simply withdraw in frustration, fear, and anger. In far-right circles this is often designated by the term “owning the libs.”

MAGA rage, like all preemptive logics, is affectively driven, continually incited and modulated by the social media environment in its functioning as a threat multiplier and amplifier. Should-be/so-be-it forms a palimpsest of aspirational conclusiveness layered on top of ineradicable uncertainty.⁹

Finally, and perhaps most far-reachingly, certain cultural critics see an operationalization of preemptive productivity in the workings of AI, in particular the large-language-model and image synthesis machine learning systems behind such game-changing products as facial recognition, ChatGPT, and DALL-E.

Machine learning (ML) systems also start at the end, and restlessly loop back to the beginning to feed back forward to the end. This circular closure is often thought of as a hewing to a pregiven norm, both among critics and boosters of AI. Shoshana Zuboff quotes a 2013 Microsoft patent application entitled “User Behavior Monitoring on a Computerized Device.” The proposed AI module would “monitor user behavior in order to preemptively detect [in the words of the patent application] “any deviation from normal or acceptable behavior’.”¹⁰ The system could be used to nudge the user into moderating their behavior (sometimes referred to as “tuning”: priming, conditioning). Or it could be used for alerting others that a behind-the-scenes response or even an open intervention is necessary (“herding”). The patent focuses on the health care providers, who could use it, for example, to monitor people with mental health problems and alert a care-giver to an anomaly in behavior that is about to occur, so that it can be headed off before it fully emerges. But it also mentions insurance companies and law enforcement, and could undoubtedly be useful for social media companies in relation to online trolling. The system is presented as an all-purpose preemptive “behavior surveillance-as-a-service opportunity” system customizable to whatever applications a client may have. Zuboff sees this as dovetailing with earlier twentieth-century technocratic visions aspiring to “preemptively eliminate anomalies, driving all behavior toward preestablished parameters that align with social norms and objectives.”¹¹ However, in an era where normalization has taken over from normation, it is no longer the case that the parameters and social norms are preestablished once and for all. The way ML actually functions tells a different story.

It is not possible in the short space available here to develop this point at length, given the highly technical nature of AI, its multiple varieties, and their rapidly changing landscape. The short story is that ML is a nonlinear process. The systems are trained using target outputs, essentially starting at the end. But the end is just the beginning. When presented with a prompt and asked to generate its own output, it has to begin all over again. It decomposes the input into sub-representational (“dividual”) features such as edges in an image or sublinguistic components of words (“tokens” in large language models). It then begins a feed-forward journey through multiple layers of processing, in the course of which patterns are detected and used to make a match (as in face recognition) or to construct a novel output (as in image synthesis or a chatbot response). Transformations are applied at each layer. These include data pooling and filtering, connective mappings, differential weightings, and even the forgetting of data to enable the system to refresh. The process is iterative. The output has the potential to be assessed, then fed back in to feed forward all over again, with tweaks applied to the transformations en route. AI is not called “generative” for no reason. The process, even when its task is “recognition,” is creative. This is because of the inescapable fact that there is, for example, no one profile of a face. A face is field of continuous variation, depending on angle, light conditions, movement, and the effect of any and all of these on the gestalt. If a face recognition operated in terms of representation—taking a set model and matching already-defined features—it would fail. It has to return to the dividual level, decompose, find patterns and then recompose, before it can attempt to match. The output is not a single portrait, but always a variation, no less so than in the more obvious cases of image synthesis and the novel responses generated in a chat box conversation. When the mission is to recognize, the match is never to an exact form, but to a set of parametric possibilities: a field of possible variations on the form, falling within certain limits.

As Louise Amoore argues, AI is possibilitic.¹² It is creatively possibilitic, not realistic, in the sense of slavishly reflecting something given. AI generates; it is actively constructive. This is why it is itself given to “hallucinate”: its mission is always, in one way or another, to produce something, which is to say to surpass the given. The nonlinearity of the iterative process as it convolutes its way through cycles of feedback and feedforward unmoors it from the simply given. AI unfounds itself so as to find not what it began with, but a new variation on it, in the form of an emergent analogue of it (an actively constructed match; a synthesized image that is as good as real; a simulated conversational response). It unfounds for refinding; it is refounding. Its product is always a derivative. In the course of the cycling through transformations among layers, anomalies can be introduced that throw the output off course. This is in addition to a fundamental element of uncomputability native to all digital processing, to which the work of Luciana Parisi has drawn attention.¹³ The risk of hallucination is a condition for the functionality of AI: it derives from the same source as AI’s capacity to be generative and surpass the given.

The nonlinear circularity of AI is not a hermetically closed loop. It is circular in the productive manner of all preemptive processes. It generates a definite output from the uncertain conditions of the world’s crowd of moving faces and their ambiguous, contradictory, complex, and ever-proliferating online expressions. AI makes do with indeterminacy. Its iterative circularity operates to filter uncertainty toward the generation of a productive conclusion. It adds a preemptive formula to the could-be/would be and the should-be/so-be-it. The first clause of its operative logic is Do what?, in the way that in colloquial English that phrase responds to something that has been put on the table and takes it up as a presumptive conclusion, reiterating it with an element of uncertainty that leverages an open-endedness out of it, preparatory to a riposte. AI’s Do what? cycles around to Take that! Do-what?/ take-that!: provide me an opening (prompt me), and I will return you a conclusion—or deliver you to a conclusion—that will constructively enter the slipstream of your life as an added factor in its subsequent unfolding. (You have been recognized! Stop in the name of the law. Your deepfake is ready for service! Proceed with trolling. ….)

Politically, two of the great potentials of AI in the wild are priming and profiling. As underlined by the notorious Facebook experiment influencing affective states through the management of users’ feeds and Cambridge Analytica’s attempts to influence voter preferences through misinformation,¹⁴ AI lends itself to the practice of the “nudge”: the quintessential preemptive art of priming. This refers to the act of inflecting behavior through cut-ins, cut-aways, and cues that are not necessarily consciously experienced as such, and precisely because of that can exert a formative pressure on experience, tweaking it to take form otherwise or move divert to a different path. Priming involves the characteristic preemptive trial-and-error modus operandi of poking and prodding to induce a taking-form, bringing to a culiminating expression an emergent variation on what might have been, in the form of a nonlinearly produced conclusion, cued in advance.

With respect to profiling, AI reinvents it from the bottom up, in its usual refounding way. There has been much concern over the propensity of large language models to reproduce racial stereotypes. It is most often said that this is due to biases in the training data, and that the systems are merely reflecting societal prejudices. This assumes that AI is reflective of a given, ignoring its “generative” nomenclature. It mobilizes a logic of pregiven categories, defined by fixed characteristics serving as identifying traits that count a body in or out of a given social grouping based on its possession of like characteristics. The issue is then which traits it is allowable to ascribe, and how to moderate online expressions to ensure that they hew to a certain norm. This is a profoundly ungenerative logic. When this idea of race is applied to profiling, it is manifestly dysfunctional. As Nahum Chandler points out, commenting on a text by W.E.B. Du Bois,¹⁵ there are no invariant traits definitive of a race. A race is a population, and a population is a field of variation, with shifting boundaries and zones of indistinction, as peoples mix and genes drift. If a pre-defined category is rigidly applied like a mold, it misses the complexity, inevitably leading to misses and misidentifications.

Profiling practiced in this manner does not pertain to the preemptive mode of power, as embodied in the surveillance state. It belongs to an older, disciplinary mode—still very much with us, but not definitive of the overall cultural condition. Power in this mode does not care about the misses and misidentifications because it is not really about correctly categorizing. It is about wielding categorization as a weapon to force a population into acquiescing to social boundaries and divisions designed to ensure the unequal distribution of life potentials (with regards to such things as mobility, enrichment, and expression). The disciplinary mode of power is also productive. But what it aims to produce is conformity more than variation, by means of the application (of a given norm) rather than through generative feedback-feedforward (toward a new iteration that a surpasses the given). Racial profiling of this traditional cast works as a mechanism of what was called earlier normation.

As Louise Amoore argues, large language models do not operate in this way. They may generate racist results even when the training corpus has been scrubbed of data repeating racial stereotypes.¹⁶ This is because what is reconstructed are patterns. Due to the effects of racialization, operating through previous applications of normation and any number of other mechanisms of social triage and conformation, accumulate in ways that pattern the social body. The patterns of such interconnected variables as home ownership, education, neighborhood, and health outcomes can operate together as a proxy for race. No racial category may be applied. Still, the racialization that is already at work in society is re-produced, in the strong sense of being produced all over again from scratch, through the multilayered reconstitution of symptomatic patternings: made to be, in new iterations. For when the results of an AI analysis are fed forward a step further, into the social policies they inform, the effect may well be to reinforce the patterns of racism the analyses reconstituted, even if that was not the explicit intent of the programmers or end users. Racism comes around again, in the guise of a new conclusion, effectively refounding it.

This is another instance of the production of ideological effects by non-ideological means. The process runs with the productively nonlinear circularity characterizing operative logics of preemption. It makes the situation more pernicious than the simple linear hypothesis of the prejudices of individuals skewing a data sample. And it makes it even more insidious than the bald-faced application of a racial category to disciplinary ends, as in police stop-and-search practices. It normalizes systemic racism, in a manner that is itself systemic, humming along in the background of the digital infrastructure. This is the kind of back-end, feedforward racialization that Jonathan Beller sees as integral to the operations of contemporary racial capitalism: a capitalism whose endemic production of inequality falls systematically along racial lines as a consequence of the accumulated effects of the histories of slavery, colonization, and segregation.¹⁷ AI can boost racism to the level of an ontopower, producing racist effects above and beyond the intentions of individuals.

All operative logics of preemption have this “machinic” character: they are self-driving. They are self-catalyzing, given the conditions (the tensions and productive paradoxes). The conditions demand the production of new, adaptive variations on the logic itself. It is not an option to simply opt out of them. Operative logics circle through us. Effects of their constitutive powers are already embedded in our daily lives, entering into our very constitutions. Because they are self-driving, merely critiquing or debunking them, as if they responded to traditional logic, will never be enough. Their embeddedness in our lives at the level at which they are made and remade—their participation in what has made us the way we have become to be—makes it clear that we cannot simply step away from them. The only way around the ontopower of processes like preemption is to contrive counter-ontopowers.

But that is another story.