On the Ecological Complexity of Artificial Intelligence

1. Autonomy

4Let me start by turning to the media and design theorist Benjamin Bratton. In a piece published for an edited collection in 2017, Bratton walks us through many of the problems associated with determining the existence of machine cognition based on a false analogy with human intelligence. As he insists, “we would do better to allow that in our universe ‘thinking’ is much more diverse, even alien, than our own particular case. The real philosophical lessons of AI will have less to do with humans teaching machines how to think than with machines teaching humans a fuller and truer range of what thinking can be” (Bratton 2017, 72). For Bratton, the stakes of clearing up the confusion over the “intelligence” of machines are high. Not only will it resolve certain hard, metaphysical problems, but it will also, and equally, reshape our ability to intervene in the many legal, ethical, and ultimately geopolitical questions regarding the rights, capacities, and agencies of non-human intelligence. These latter questions become all the more pressing once we take seriously the material and infrastructural distribution of artificial intelligence operating at a planetary scale. This is a computational reality which begins to make the far-fetched dreams of those noösphere theologians of Silicon Valley who prophesized wrapping the Earth in a layer of synthetic intelligence, seem realizable, if not, in some strange way, already here (Gabrys 2016).

5Bratton, for his part, has done well to articulate the possibilities of a planetary-scale, synthetic intelligence with his theory of “The Stack” (2015). The stack is a concept borrowed from computer science, and it describes a data structure that contains relatively autonomous elements layered on top of one another. These elements can be added or removed only from the top (according to the last in, first out principle). Only now, for Bratton, The Stack serves to illustrate an infrastructural totality that functions at the planetary scale. So conceived, The Stack is an accidental megastructure containing various interdependent layers: Earth, Cloud, City, Address, Interface, User. Together, they constitute an emergent order wrapping the Earth in a new kind of synthetic skin (87). It would be a mistake to assume, however, that the architectures of this platform are purely algorithmic and abstract; indeed, they’re equally embodied, material and concrete. As Bratton writes: “As much as it is made of computational forms … The Stack is also composed of social, human, and concrete forces (energy sources…dashboards, cities and streets, rooms and buildings…” (Bratton 2015, 11). In short, the architectures of The Stack are designed to give the earth a new exoskeleton, wrapping every scale of terrestrial activity, from the geological and biological to the social and political, in the artifice of automation.

6More than this, however, The Stack displaces the old, Westphalian system of sovereignty (the 1648 treaty that effectively marks the beginning of the sovereign authority of the nation-state) to give birth to a new geopolitical order: it’s one where the sovereignty of nation-states and even markets yield to the territorial designs of planetary-scale computing (6). The Earth’s new computational epidermis is ultimately a system of terrestrial governance, a new modality of geopolitics, that doesn’t completely replace the old systems (doing away with the state-form, for instance), but it fundamentally reworks them so that “the machine [now functions] as the state” [8]). Or in other words, Stack architectures fundamentally shift the ontology of the political and economic state.

7Still, for Bratton, and increasingly others, this emergent geopolitical order is not a call to arms, like we might find in the work of Bernard Stiegler or Franco “Bifo” Berardi. It serves as a kind of blueprint for imagining new modalities for speculation, philosophy, and for the rational subject itself, all prepared to tackle the overlapping crises our species faces today. And this is because The Stack, which is essentially a form of synthetic intelligence operating at the planetary scale (see Bratton 2021), can serve as a new model for sapience, one that can operate at the scales and with the levels of complexity adequate to the shifting modalities of terrestrial governance. What this also means is that using human rationality as the gold standard for thought and cognition, is not only anthropocentric but is also fundamentally unimaginative. Managing the various and complex crises our species faces demands thinking at scales that exceed the human, and thus speculating about other parameters for rationality; these are parameters that may be alien to the human, but nevertheless demand our attention (see Bratton 2019). And The Stack opens up a different horizon for cognition, which we, as humans, may not yet be prepared for: “It may be…our predicament,” Bratton insists, “that we cannot design the next political geography or planetary computation until it… more fully designs us in its own image or, in other words, that the critical dependence of the future’s futurity is that we are not yet available for it!” (2015, 15).

8The wager, then, seems to be this: “intelligence” and “rationality” (which really amount to the same thing in this argument) have, for far too long, been circumscribed by the material constraints of human biology; and the latter in no way exhaust the possibilities for the former. Intelligence is more expansive and more complex than its contingent realization in Homo sapiens, and recent advances in computing give us a taste of other parameters for sapience. While the common-sense discourse on AI, which continues to stoke fear and uncertainty in the general public, insists on using human intelligence as criteria for intelligence generally, such Turing Tests are underwritten, and not surprisingly, by deeply rooted forms of anthropocentrism. Not only do such criteria fail to capture the way AI researchers actually regard the intelligence of their machine-learning algorithms (as if they wanted their “humanness” reflected back to them [Bratton 2017, 75]); but using such classificatory schema also repeat all the old sins of speciesism: they fail to affirm a wide variety of nonhuman intelligences, whether they be expressed in nonhuman animals, the biosphere, or in algorithmic machines.

9It’s not hard to see why this line of thought would have wide appeal, especially among those looking for histories of technology that fall outside of anthropocentric frames; or among those searching for ways to affirm nonhuman thought and sensibility through a myriad of posthuman discourses, including animal and plant studies, Queer ecology, new materialism, and so on. What’s more, a claim that runs throughout this work, and is developed in explicit philosophical terms by Reza Negarestani, Ray Brassier, Luciana Parisi, Peter Wolfendale, and other so-called “neo-rationalists,” is that algorithms enact forms of intelligence that are, in some fundamental sense, “alien” to what we tend to regard as human intelligence. Machine learning algorithms synthesize and model data in ways that have no analogue in the human world; humans ascribe meaning to them, but in themselves, algorithms are mechanical procedures devoid of sense and meaning. And yet, it’s just this recognition that the algorithmic manipulation of symbols has a rationality to it that makes it possible to conceive of paradigms for reason that are not dependent on human faculties. This would mean that rationality itself is fundamentally autonomous from the constraints of the human organism. Reason, then, is a set of functionalities expressible in different material substrates, a realization that tares down many of the barriers to the nonhuman world that have been firmly in place at least since the work of Immanuel Kant. It’s also no surprise that this capacity to think the existence of thinking outside its correlation to human thinking pairs nicely with the much wider research program in post-continental philosophy to resist what Quentin Meillassoux famously calls “correlationism,” or the constant correlation between “Being” and “human thought” in modern philosophy.

10This isn’t to say that the intelligent acts humans perform are of little interest. Quite the contrary: Reza Negarestani, who has perhaps developed this line of thought further than anyone, affirms that his work is a kind of “enlightened humanism.” But it’s a perverse humanism, no doubt: it “extracts,” what he calls, the “essential core of humanism,” namely, rational agency, from its contingent, evolutionary manifestation in Homo sapiens; and it affirms that there are other substrates capable of realizing this program. Such a functionalist argument for the autonomy of intelligence recognizes that reason is an “abstract protocol” only incidentally and not essentially tied to carbon-based matter, and can be realized just as easily in silicon or some other material. And so, what has, since antiquity, been viewed as the most human activity, namely, rational activity, is actually what is most inhuman in the human—it’s the “alien vector” within humanism. Thus, and this is crucial: the intelligent acts performed by AI systems are less anti- or non- human than they are inhuman actions, in as much as they express what is already alien within the human (reason). For Negarestani, this points the way toward a much more ambitious collective, philosophical program that tries to make good on Hegel’s unfulfilled promise to realize collective, self-conscious rationality; only now, this is actualized through affirming the inhuman forces of reason in the human and fully artificializing ourselves as a species. This is philosophy conceived as a program to “reengineer” the human on a more rational, inhuman template (Negarestani 2017).

11Admittedly, the latter thesis about inhuman engineering as a revived Hegelianism is less interesting for our purposes than are the strong claims about the autonomy of rationality, and more specifically, the assertion that intelligence is irreducible to the historical and/or evolutionary contingencies of material systems. What’s more, by affirming the synthetic rationality realized in, for example, a neural network, the claim is that we are actually getting much closer to grasping reason as it is in itself, outside its correlation to the biological particulars of the “human,” than we are when we study its enactment in the embodied cognition of humans. And it’s precisely this recognition, the recognition that cognition is irreducible to its temporary manifestation in Anthropos, that Bratton, Negarestani, Parisi, and others use to set the philosophical scene for designing new planetary subjects of reason.

2. Critique

12It’s just this affirmation of the autonomy of synthetic intelligence that could not be further from the work in Critical AI Studies. As the AI scholar Kate Crawford argues, in seemingly direct opposition to what was claimed above:

AI is neither artificial nor intelligent. Rather, artificial intelligence is both embodied and material, made from natural resources, fuel, human labor, infrastructures, logistics, histories, and classifications. AI systems are not autonomous, rational, or able to discern anything without extensive, computationally intensive training with large datasets or predefined rules and rewards (Crawford 2021, 8).

13Thus, what Crawford, along with a growing number theorists hailing from critical media and technology studies, infrastructure studies, and post- and decolonial theory argue is essentially this: machine learning is part of much wider social, biological, and geophysical environments, and these environments tend to get disregarded or sidelined in the enthusiasm over the possibilities of “synthetic intelligence.” The challenge, according to Crawford, is seeing past, for example, our enthusiasm over the imperceptible patterns in data machine learning algorithms are capable of finding, to see the labor, materials, and infrastructure required to produce such results. For as she explains elsewhere, “each small moment of convenience” an Amazon Echo provides “…be it answering a question, turning on a light, or playing a song – requires a vast planetary network, fueled by the extraction of non-renewable materials, labor, and data” (Crawford and Joler 2018, n.p.; emphasis in original).

14The claim here is that AI techniques need to be demystified. Behind the “black box” of machine learning, and the opacity manufactured by Big Tech companies, there are vast political, economic, social, and material systems, spread across the global North and South, that are asymmetrically affected by something as (seemingly) straightforward as training an algorithm to make a recommendation based on user preferences. These planetary systems make up the hard and soft infrastructures concealed behind the slick and efficient operation of mathematical tools, and are best characterized, according to Mary Gray and Siddharth Suri, as “ghost work”: the invisible, but always present, laboring force necessary for AIs to build models, to make recommendations, and more. As Matteo Pasquinelli and Vladan Jolar put it:

All training datasets are implicitly, a diagram of the division of human labour that AI has to analyse and automate. Datasets for image recognition, for instance, record the visual labour that drivers, guards, and supervisors usually perform during their tasks. Even scientific datasets rely on scientific labour, experiment planning, laboratory organisation, and analytical observation (Pasquinelli and Jolar, 2020, 19).

15In light of this “invisible work force,” we require techniques for visualizing the physical systems of work and exploitation, of resource management and data extraction, of digital sovereignty and geopolitical alliance. And that is precisely what Crawford provides: a planetary cartography of “smart computing”—one that ranges over the lithium mines, the laboring minds and bodies, the elaborate technologies of data extraction and classification, as well as the modalities of sovereignty, security, and finance born out of artificial intelligence at scale. Such a cartography of AI is meant to be disenchanting: It’s what allows us to look under the hood of the car and actually see how the engine runs. And what’s revealed are indeed infrastructures, industries, and power dynamics; but ultimately, what comes into focus is that AI is “a manifestation of highly organized capital backed by vast systems of extraction and logistics, with supply chains that wrap around the entire planet” (18-19). On this reading, it’s the critique and redirection of the “empires of artificial intelligence” that’s ultimately at stake.

16And to be sure, an AI cartography comes to function, for Crawford, as a strategic intervention. This is because creating atlases is precisely what the “AI empire” does: “The AI industry,” she writes, “is making and normalizing its own proprietary maps, as a centralized God’s-eye view of human movement, communication and labor… This is a desire,” Crawford continues, “not to create an atlas of the world but to be the atlas—the dominant way of seeing” (Crawford 2021, 11). In this perspective, the old adage that the map is not the territory no longer holds water: the AI map is the territory. As algorithms map patterns, the landscapes of planetary movement, communication, work, trade, war, and so on, are simultaneously being charted. These are the complex, and very often invisible, architectures of the planetary that only become visible and thinkable on condition that we see artificial intelligence as physically embedded in the world and redesigning the Earth.

17Returning to the previous section, it seems clear now that focusing on the abstract, autonomous, and even alien qualities of computational systems conceals the very real, physical processes involved in training algorithms, in building models and recommender systems. It’s not by magic, or some mystical force, that algorithms are endowed with their powers; no: these capabilities are built step-by step from labor, resources, and data, all of which can be traced and visualized. It’s out of these planetary cartographies of AI that the difficult work in AI auditing can begin; these are the spheres of AI justice, ethics, and regulation that Ruha Benjamin articulated so well in her book Race After Technology. And so, to insist on the “autonomy” of these algorithmic systems is ultimately to disavow the planetary-scale dependencies that make the illusion of autonomy possible. What this means is that the continued mystification and even quasi-theologization of deep learning techniques only mask the real planetary costs of artificial intelligence. AI is not alien to human intelligence in the way some might think, then. It’s rather that the “ghost work” of human intelligence remains alienated from what it produces.

18With this characterization of alienation, the Marxist critique of artificial intelligence comes into focus. This reading is certainly implicit in Crawford, but it’s explicit in Matteo Pasquinelli (2019). As Marx makes clear in both the Grundrisse and Capital, technical machinery is not science’s “analysis” of nature, but the “analysis” of labor. Indeed, it’s the division of labor that drives technological development and not the other way around. However, once labor is transformed by the scientific nature of production, the collective knowledge of laborers is objectified in machines as fixed capital, and workers are alienated from their own collective knowledge (that has become the knowledge of science and technology). This means that the so-called intelligence displayed by machine learning algorithms is no way autonomous or free from human productive labor. And what a cartography of AI does, politically, is map the global exploitation and alienation of the labor required to design and build these systems.

19What’s intriguing here is that the so-called “autonomy of AI,” advanced by Bratton, Negarestani, Brassier, Parisi and others is just as often built on the back of a Marxist political project. Without going too far down the rabbit hole of post-Marxist political theory, it’s worth pointing out that it’s Marx’s Grundrisse, instead of his Capital, that serves as the foundation forthis political project. This argument, which has been repeated many times in Italian post-operaist thought, boils down to this: “The Fragment on Machines” in Marx’s Grundrisse provides a kind of blueprint for algorithmic autonomy.

20If we move beyond the so-called “Dogmatic Marxisms” that identify technology with “fixed capital” and “dead labor” (as we saw above), then we begin to see how, for Marx, “the evolution of machinery also indexes a level of development of productive powers that are unleashed but never totally contained by the capitalist economy” (Terranova 2014, 386). Within the dominant capitalist relations of production, the surplus labor that results from increased automation is reabsorbed into the capitalist economy. But for those who put inestimable faith in Marx’s “Fragment on Machines,” disposable time can and should be reclaimed, and it can serve as a basis for new forms of the commons. As the Italian media theorist Tizianna Terranova argues, “We need to ask … not only how algorithmic automation works today (mainly in terms of control and monetization, feeding the debt economy) but also what kind of time and energy it subsumes and how it might be made to work once taken up by different social and political assemblages—autonomous ones not subsumed by or subjected to the capitalist drive to accumulation and exploitation” (Terranova, 2017 389). In other words, the marching orders are clear: subvert the exploitive uses of computational technologies, and use them to build infrastructures, like Universal Basic Income, that could support the production of a post-capitalist commons. Of course, this argument very much depends on the autonomy of algorithmic technologies (from human labor at least), so that complicated infrastructural work can get done, freeing up humans to pursue other, collective activities (Williams and Srnicek, 2014).

21The point here is not to continue down this path of Marxist political thought; rather, it’s to suggest that both projects can be read as overtly political, even if the “critical environmentalists” accuse the “algorithmic autonomists” of disavowing the political nature of their work, in order prop up the autonomy of computational practices. What’s more, it may even be the case that these differences can be fought out in Marxist terms, and much more specifically over how the “machine question” is configured in the Grundrisse and Capital. But I won’t pursue this any further here.

3. Opacity

22Leaving these admittedly abstract discussions aside for the moment, I want to stage my intervention by appealing not to philosophy (just yet), but to recent work on auditing and governance in the social and legal sciences. Now, the legal scholar Frank Pasquale is well known for his work on algorithmic transparency and auding (Pasquale 2015). He insists that in a world where our systems of government and finance are increasingly run by machine-learning algorithms, the computational tools that are making decisions on our behalf need to be made available for audit. Indeed, as Pasquale, along with a host of other legal and policy scholars have made clear, algorithmic accountability is absolutely central to upholding democratic practices. The legal theorist Mireille Hildebrandt, for her part, argues that the (admittedly imperfect) system of constitutional democracy has two minimal requirements: on the one hand, people live under rules of their own making; and on the other hand, people have to be able to contest how the rules are made and implemented (Hildebrandt 2016). For this, decision making practices need to be transparent, so those who are “ruled” are not dominated, but participate in the democratic process itself. The rationality of law, and the granularities of decision making, require transparency and accountability.

23But for Hildebrandt, with the rise of algorithmic governance, where algorithms actually do the work of anticipating and regulating the behavior of its citizens, the “transparency” of democratic practices is fundamentally threatened, since the decision-making process used is itself shrouded in trade secrets. Not resting content with such technological fatalism, Pasquale is more optimistic about the possibilities for demanding transparency, and offers a series of recommendations for how to realize an “intelligible society” increasingly run by algorithmic machines.

24However, the trouble with this story, as computer scientists like Paul Dourish (2016) and Jenna Burrell (2016) note, is that it fails to account for the different levels of algorithmic opacity. Burrell, for her part, outlines three such layers, which ultimately trouble arguments in favor of algorithmic transparency (Burrell 2016). The first two levels of opacity are straightforward enough, and are well documented in the legal scholarship: on the one hand, the algorithms that run our search engines, that determine our credit scores, or that assist Alexa in ordering the best-rated takeout within five city blocks, are all protected in trade secrets. But notice: even if these algorithms were visible, they are mathematically complex abstractions whose realization in code is unintelligible to just about everyone but the highly trained specialist. Thus, even if the first form of opacity could be overcome, algorithmic operations will still be far from transparent to the general public. Still, these forms of opacity do not necessarily lead us to the conclusion that AI is a “black box.” There are numerous policy recommendations, pieces of legislation, and training programs that could help shed light on the algorithms that govern our lives.

25Despite these efforts, there is yet another layer of opacity that presents a far greater challenge to the ideal of transparency. To understand this, it’s important to recognize that machine learning techniques do not require, in the way traditional techniques of statistical modelling do, that we identify the functional relationships between variables in advance. Instead, with machine learning, data are mined using algorithmic techniques (such as deep neural networks) to identify patterns in data that inform a model that can be tested and improved upon. For example, an algorithm for handwriting recognition is exposed to massive amounts of training data, but it is not programmed with the human concept of the alphabet or even with shape of certain letters. Instead, it works “bottom-up” to recognize that certain configurations of strokes are characteristic of certain letters (Dourish 7). And with this, the algorithm may very well be able to confirm handwriting samples. And as a human interpreter of outcomes, I am satisfied with how the algorithm performed.

26Still, there’s no way to know what the algorithm “knows.” In the case of social media feeds and trending topics, a computer scientist can certainly describe the algorithm—its weightings, and so on—and hypothesize why one topic trends and another one doesn’t, but they cannot actually peer into the algorithm and see how the sausage is made. The patterns an algorithm picks out in data are not only “ephemeral,” and so almost impossible to forensically trace, but they are statistical and probabilistic and so not in themselves meaningful in human terms. What we’re left with is /human interpretation imposed on mathematical processes of statistical optimization, but not a picture of why this decision was made at this specific time. This leads Burrell to conclude that: “When a computer learns and consequently builds its own representation of a classification decision, it does so without regard for human comprehension. Machine optimizations based on training data do not naturally accord with human semantic explanations…The workings of machine learning algorithms,” “can escape full understanding and interpretation by humans, even for those with specialized training, even for computer scientists” (Burrell 2016, 10).

27 Here’s why this matters for our argument: it’s hard to deny that ubiquitous, algorithmic auditing would go a long way toward creating what Pasquale calls an “informed citizenry” that takes an active role in understanding and intervening in how our data are harvested and used. Despite these efforts, there are aspects of algorithmic decision making that refuse to be objectified, and that stubbornly resist comprehension, even in the face of strict auditing practices performed by the most skilled computer scientists (or algorithms). And if this opacity is made explicit in the legal and social scientific research on algorithmic governance, then it’s no less true for the critical environmental discourse on artificial intelligence. Although Kate Crawford and others offer an ambitious program for mapping the geographies of labor, data, and resource extraction, their cartographies of AI will always be in some sense incomplete. It’s true, of course, that their ambition is not create the map of AI, and repeat the sins of the AI empire they’re trying to critique. However, their incomplete map is generally conceived of as a limitation in perspective (i.e., there is no perspective that captures all perspectives) and not in the conceptual framework itself.

28Recall that what’s at stake in these cartographies is bringing the material and infrastructural systems that support an algorithm into view—that is: seeing, mapping, and critiquing them. And while it’s surely the case that we can point to increasingly complex networks of material infrastructure, and see that they’re brought into relation with the exploited labor of knowledge workers, and also show how these elements feed into the selection of training data, we cannot say with any kind of certainty why these elements have come together, at this time, for an algorithm to make this recommendation. What role these elements play for the algorithm, their importance, and in conjunction with other factors, is largely off limits. It’s as if, on the conceptual model offered by Crawford and others, the inner workings of an AI model are visible if we look closely enough at the environments that influence it. There is a kind of “ontological flattening” at work here, where the inside is the outside, and the one reflects the other (Wolfe 2020). This is a level of “transparency” that’s not just a technical impossibility, but it’s also, I want suggest, an ontological one.

4. Complexity

29This brings us back to the first section of this talk on “autonomy.” While I’m deeply critical of arguments that champion the artificialization of reason on an algorithmic template, I’m also aware of the need to respect the levels of opacity involved in machine learning. We cannot pretend to objectify the internal processes of a deep neural network in human terms. Although great progress has been made in the last several years to make machine learning models less opaque, a fully transparent model, which would account for the vast networks of human and nonhuman infrastructure required to build the model, is inconceivable. Still, the environmental critique of AI has shown why it’s also not in our interest to call these concealed operations “reason” or “intelligence”— this is not only because they’re too easily equated with forms of autonomy that disavow their environmental dependency, but also because such logics of autonomy still on some level uphold the values of the liberal subject of reason, and thus do not so easily escape the straightjacket of humanism.

30What we require, instead, is an ecological conception of computing that does not give into the temptation to reduce the system to its environment. Which means: we need to account for the vast networks of labor and resources that condition the material and functional operations of machine learning, but do so in a way that does not predetermine how those environments are organized for and by an algorithm. In short: we cannot prescribe what an environment is and will become for a system in advance. We need an ecological philosophy that’s complex enough to deal with systems that are irreducible to the vast environments that condition and make them possible.

31Naturally, the complexity sciences lend themselves to such theorization. And yet, here, too, we need to avoid some of the troubles haunting complexity conceived on mathematical template. There’s no doubt that complex dynamical systems have, for the last half century (or more), proven to be an especially powerful mathematical language for explaining systems that contain multiple stable trajectories; that undergo change induced by different causes; that experience large-scale transformation as a result of a very small shift in the environment; and that cannot be reduced their parts. These processes can be modeled using nonlinear differential equations, and this has proven to be a game changer for multiple scientific fields, including the biological and ecological sciences.

32The trouble with this picture, which I have written about elsewhere, is that while the environments and causal mechanisms have gotten more complex, so have the mathematics that model these mechanisms. The power of nonlinear dynamical systems lies in their ability to demystify what was once mysterious. At bottom, the sciences of complexity are tasked with discovering the algorithms on which physical, biological, and social systems run. And the portrait that’s painted for scientists, life and environmental scientists included, is an attractive one: the functionality of a system fits into a larger, mechanistic picture in which the distributed mechanisms of feedback and recursion that once baffled scientists sit alongside the complex behavior of other physical, social, and economic systems. And so, what we’re left with is a conceptual frame that tries to eliminate, as much as mathematically possible, the indeterminacy between and among systems. But it’s precisely this quest for mapping the causal mechanisms underlying all systems that eclipses the heterogeneity and irreducibility we’re looking to preserve in our ecological theory of computing.

33For this reason, if I invoke complexity here, then it’s because I’m drawing on a largely overlooked genealogy of the complexity sciences derived from the mid-twentieth century mathematician and theoretical biologist, Conrad Hal Waddington. Waddington, who is perhaps best known for having coined the term “epigenetics,” was just as responsible for the birth of complex systems in the biological sciences. And behind both of these conceptions is the process metaphysics of the mathematician and philosopher, Alfred North Whitehead. As a closer reader of Whitehead, Waddington showed how genetic and epigenetic systems are not isolated from one another in embryonic development; instead, development is understood as a complex system in “multidimensional phase space.” Though Waddington was especially careful to emphasize that the parameters selected to model a developing system are artifacts of the modelling environment. And because of this, we should never forget that the true complexity of a developing system can never be modeled, since it’s shaped by at least three scales of time that resist objectification in a model: evolution, life history, and physiology (Waddington 1957, 6). In other words, development is more complex than the sciences of complexity.

34Waddington’s intuition of irreducible complexity is no small part indebted to Whitehead. In fact, Waddington himself notes in 1975 that his conception of the “creode,” which is Waddington’s technical term for how an environment gives shape to a system’s trajectory in phase space, should have been more explicitly Whiteheadian in its original articulation. This is because, as Waddington explains and comes to realize later in his work, the specificity of a system’s trajectory is not determined by an external “switch” (Waddington 1975, 9). It’s rather determined by the concrete ways in which the environment becomes relevant to the system during the material process of development. Which means that the true complexity of a system’s development concerns how the environment shapes and is shaped by the system. In short: The environment does not pre-exist the system.

35It’s Whitehead who develops this idea in more explicitly ecological terms, however. For Whitehead, entities, whether they make up lines of code or the operations of human cognition, do not influence each other by bumping up against one another as they do for much of the history of Western metaphysics; rather, they actually enter into the constitution of one another. Each entity in the universe (or occasion of experience) becomes what it is because of the way it incorporates (or prehends) its environments into its own becoming. Hence, there is no dualistic opposition between self and other: the self just is its integration of its environments. And what’s more, there is no “environment” in isolation of other environments; each environment of entities, or what Whitehead calls a society, “requires a larger social environment” (Whitehead 1978, 100) that “provides a favorable environment for the subordinate societies which it harbors” (99). The human subject, just as much as the single-celled organism, emerge and are sustained over time by the incalculable number of wider ecosystems they incorporate.

36A lesser emphasized aspect of Whitehead’s philosophy, which tempers his cosmic “relationism,” is that there is no bare repetition of an other’s experience in a prehending occasion. Causal efficacy is not all that’s at stake. If this were the case, then nothing would happen, the world would not move. If one entity actually enters into the constitution of another, then it would have to do so uniquely. What this means, for Whitehead, is that each entity integrates its proximate and distal environments in an irreducibly singular way, so that each and every coordination of environmental data cannot be objectified in another occasion of experience. Thus, how each entity incorporates its environmental background into its own becoming cannot be prescribed from what has been. The directionality of each entity’s emergence is metaphysically irreducible to what is already settled. Every entity is therefore interior and exterior to every other entity at the same time; and this is as true for events of human cognition as it is for events algorithmic processing.

37Now, it’s thanks to this conjugation of Waddington’s theory of complexity and Whitehead’s metaphysics of ecological relation that a different conception of ecological computing begins to come into focus. With it, we gain a conceptual architecture that can account for the fact machine learning models are embedded in incredibly diverse physical, social, and political environments, and these environments actually enter into these models; they literally make it possible for algorithms to do their work. They are the complex infrastructures Crawford and others map. Still, these environments are coordinated in highly specific ways, each time, and on every occasion of algorithmic integration (or concrescence, for Whitehead). The concrescence of data, resources, and labor, to continue with Whitehead’s terminology, cannot be externalized into intelligible models, even though the algorithm’s concrescence is literally constituted by what is external to it. I cannot say, in other words, what the environment is for the system. On this template, the real ecological complexity of an AI model has to do with the fact that internality and externality are not opposed, in fact, they are required by one another.

38Surely, one of the most important lessons of the cartographic critique of AI is this: today’s machine learning algorithms are embedded in much wider environments that they use and exploit, and we need to make these practices visible and available for critique. This is also fundamental to the conception of ecological computing I’m proposing here. What’s different, and this is where I will finish up, is that blunt instruments of critique no longer hold water. What we tend to find in both camps—whether it’s the algorithmic autonomists or the critical environmentalists—is the reification of nonhuman reason or the environment; it’s all or nothing. Instead, the kind of critical program of AI that’s offered here proposes a radically situated philosophy of computing, which is to say: a theory capable of affirming that machine learning models, especially the ones that seem to have the most predictive power, are constituted by environments that do not pre-exist their highly specific ways of being coordinated by the model. In other words: the system constructs the environments that make it possible. As such, there are no environments “out there” to map, objectify, and understand —they’re different each time; nor is there some nonhuman, “xeno”-rationality that is autonomous from its material instantiation, which could be leveraged to imagine new forms of planetary design and governance. This is pure, Promethean fantasy.

39Instead, what we have to come to terms with is a planetary-scale computation that does not allow us objectify what the planet is. The planetary is different each time; indeed, with every model, there’s a new planet, or rather, a different coordination of biosocial and geophysical activity as Dipesh Chakrabarty might say (Chakrabarty 2021). Which means: there is no single planetarity, there are only multiple and diverging planetarities. Thus, if a critical project is to be imagined, then it would have to take shape against the backdrop of these situated practices of computing. This isn’t necessarily a call to reinvent the wheel, since there’s no shortage of situated critique on the Left, from deconstruction and feminist science studies to decolonial theory. But it does call for new attention to be paid to the practices of computing, especially when this genealogy of political thought too often resorts to sweeping critiques of technological homogenization and instrumental rationality. Situated political critique needs to be reimagined according to the situated design agencies of machine learning, which continually redraw the planet with no window onto its own design process. And so, to conclude, I will simply say this: although the theory I’m proposing here may not be able to say what the “planet” is for a neural network, it is nevertheless able to articulate the risks of identifying it as one. Perhaps, then, this bit of critique also offers a way forward for future philosophies of computing: They need to attend to the plurality of planets designed by machine learning algorithms.

References

1. Benjamin, Ruha (2019): Race After Technology: Abolitionist Tools for the New Jim Code. Cambridge, UK: Polity.
2. Bratton, Benjamin (2017): Outing Artificial Intelligence: Reckoning with Turing Tests. In: Alleys of Your Mind: Augmented Intelligence and its Traumas. Edited by Matteo Pasquinelli. Lüneburg: Meson Press.
3. Bratton, Benjamin H. (2015): The Stack: On Software and Sovereignty. Cambridge, MA: The MIT Press.
4. Bratton, Benjamin (2019): The Terraforming. Moscow: Strelka.
5. Burrell, Jenna (2016): How the machine ‘thinks’: Understanding opacity in machine learning algorithms. In: Big Data & Society. January–June: 1–12.
6. Chakrabarty, Dipesh (2021): The Climate of History in a Planetary Age. Chicago: University of Chicago Press.
7. Crawford, Kate (2021): Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven: Yale University Press.
8. Crawford, Kate and Vladen Joler (2018): Anatomy of an AI System: The Amazon Echo as an Anatomical Map of Human Labor, Data and Planetary Resources. https://anatomyof.ai/.
9. Dourish, Paul (2016): Algorithms and Their Others: Algorithmic Culture in Context. In: Big Data and Society 3(2). https://doi.org/10.1177/2053951716665128.
10. Gabrys, Jennifer (2016): Program Earth: Environmental Sensing Technology and the Making of a Computational Planet. Minneapolis: University of Minnesota Press.
11. Gray, Mary L. and Siddharth Suri (2019): Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Boston and New York: Houghton Mifflin Harcourt.
12. Hildebrandt, Mireille (2016): Law as Information in the Era of Data-driven Agency. In: Modern Law Review 79: 1–30.
13. Joler, Vladen and Matteo Pasquinelli (2020): The Nooscope Manifested: Artificial Intelligence as Instrument of Knowledge Extractivism. https://nooscope.ai/.
14. Marx, Karl (1990): Capital Volume 1. Translated by Ben Fowkes. London: Penguin Books.
15. Marx, Karl (1993): Grundrisse. Translated by Martin Nicolaus. London: Penguin Books.
16. Negarestani, Reza (2017): Intelligence and Spirit. Falmouth, UK: Urbanomic.
17. Pasquale, Frank (2015): The Black Box Society: The Secret Algorithms That Control Money and Information. Boston: Harvard University Press.
18. Nocek, Adam (2021): Molecular Capture: The Animation of Biology. Minneapolis: University of Minnesota Press.
19. Pasquinelli, Matteo (2019): On the Origins of Marx’s General Intellect. In: Radical Philosophy 2. O6: 43–56.
20. Terranova, Tiziana (2017): Red Stack Attack! Algorithms, Capital and the Automation of the Common. In: #Accelerate: The Accelerationist Reader. Edited by Robin Mackay and Armen Avanessian. Falmouth, UK: Urbanomic Media Ltd.
21. Williams, Alex and Nick Srnicek (2014): #Accelerate: Manifesto for an Accelerationist Politics. In: #Accelerate: The Accelerationist Reader. Edited by Robin Mackay and Armen Avanessian. Falmouth, UK: Urbanomic Media Ltd.
22. Waddington, Conrad Hal (1975): An Autobiographical Note. The Evolution of an Evolutionist. Edinburgh: Edinburgh UP.
23. Waddington, Conrad Hal (1957): The Strategy of Genes. New York: Routledge.
24. Whitehead, Alfred North (1929/1978): Process and Reality: An Essay in Cosmology. Edited by David Ray Griffin and Donald W. Sherburne. Corr. ed. New York: Free Press.
25. Wolfe, Cary (2020): What the ‘Animal’ Can Teach the ‘Anthropocene’. In: Angelaki: Journal of the Theoretical Humanities 3: 131–145.