人工智能和资本主义的未来

Clegg 2014: 51). Yet for an essay that hinges on the role of ‘labour-saving technologies’ within capital, it says relatively little about machines. In a lecture that does take up this issue more directly, Benanav (2017) remarks that the conventional focus on technology is a ‘fetish’ discourse that synopsizes the complex forces of capitalism around the figure of the robot. And yet, as he acknowledges, automation is a fundamental part of the crisis. Bearing both parts of this paradox in mind, we will now relate Marx’s ‘general law’ and the problematic of ‘surplus populations’ to the wave of ML AI.¹⁹

‘AI Apocalypse Now’ and ‘Business-as-Usual’ theorists both, at least in their popular expression, operate with highly deterministic, one-dimensional logics. AI Apocalyptics follow a hockey-stick graph of exponentially accelerating technological change, Business-as-Usual theorists a model of a homeostatic self-equilibrating labour market. In contrast, Marxist analysis sees technological change and market dynamics as reciprocally related, combining to produce intermittent but recurrent system crises. And while it has its own teleological versions of the resolution of such crises, such as that of a falling rate of profit leading to a terminal crisis of capital, other strands allow for a more complex interplay of tendencies and counter-tendencies. These allow us to envisage a staccato unfolding of AI employment effects, in which working-class decomposition and recomposition are active elements. In such an optic we can see how the drive to AI automation may be retarded by capital’s success in establishing a cheap-labour economy, for example, through globalization, then boosted by the re-emergence of wage-raising labour struggles, so that, for example, a resurgence of wage demands in tightening US labour markets might spark capital’s actual adoption of AI options in prototype or under research.

As such automation gains ground, it in turn increases the reserve army of the unemployed, and intensifies precarious work, lowering wage-rates in many sectors. As Jason Smith (2017) points out, under capital people have to sell their labour-power to survive: wage-labour has ‘nowhere to go’, so we can anticipate both ever greater expansions of a service sector, commodifying all kinds of personal interaction, as well as the proliferation of increasingly arcane forms of self-employment. These developments would continue even while a range of occupations are partially automated – so that, for example, autonomous truck convoys are accompanied by one or two safety-drivers, or fundamentally automated tasks such as routine medical diagnoses maintain a ‘human veneer’ (T. Lee 2018) of workers to wrap results with manifestations of compassion and care. The issue at this point would not be that there were no jobs, but rather that jobs would be subject to a persistent, creeping downward pressure on wages and conditions from advanced machinic competition. This plateau would again temporarily halt capitalism’s incentive to automate, until either new wage-raising struggles or innovations reducing technological costs ignite a new round of substitution of fixed for variable capital. Workers’ movements for improvement in wages and conditions will be constantly liable to an automating response, and indeed will provide a major catalyst for its continued forward movement. Such a jagged process would be overlain by capital’s regular business cycles, and by its more intermittent giant spasms of overproduction (to which of course AI-related job loss would contribute) with their familiar pattern of major surges in unemployment and increasingly prolonged jobless recoveries.

142

inhuman power

This suggests a process of AI employment effects different from both the sudden-onset, across-the board ‘Apocalypse Now’ and the moreor-less steady-state ‘Business-as-Usual’ models: we might call it a ‘Slow Tsunami’ of market-driven technological change gradually flooding out the labour market, driving remunerated work to diminishing – and, in terms of the logic of capital, more and more economically insignificant – islands of human-centric production. This is would be the situation metaphorically represented in the many sagas in which humankind is pursued across the universe from one refuge to another by implacable machinic adversaries: Cylons come to mind. Ameliorated by reforms such as a universal basic income (whose merits and demerits we discuss in the Conclusion), this process could be very protracted. For a sense of possible time scale, consider that capital’s foundational process of ‘primitive accumulation’, with its large-scale eviction of populations from the land to form urban proletariats, is generally considered to have occurred over centuries rather than decades, and indeed, on a global scale, to still be incomplete. Despite futurist insistence on the speeded-up nature of contemporary social change, one could imagine a capitalist phase of ‘futuristic accumulation’, shedding, rather than amassing, wagelabour, but in a similarly uneven and protracted fashion.

With the emergence of AGI, however, the futuristic accumulation of surplus populations would slowly or quickly engulf the human species. While capital does not care what material its labour-powers come in, it cares about the productivity of labour, and AGI would be far more productive than baseline humans. An AGI need not engage in time-consuming superfluous behaviours like breathing and eating. AGI hardware could be endlessly augmented and could, therefore, do whatever humans can do, but with more efficiency and precision, and faster. Indeed, an AGI need not be limited to any morphology or even a body at all; learning how to use new and different bodies for environments with extreme pressure, cold or heat, it could likely divide its attention between numerous bodies or entities; indeed it could exist as a factory or even an entire supply chain. If a supply chain could speak, would we understand it?

perfect machines, inhuman labour

143

In Marx’s analysis surplus populations are relative because of their fluctuation in size, as workers are alternately expelled and incorporated into capitalist production. With the advent of proletarian AGI, this population would become absolute, coextensive with a human species rendered obsolete to the valorization of value. Humanity would become a ‘legacy system’, outdated hardware unsuitable for running the inverted world of capital. The status of humans in such a situation might be comparable to the current status of wild animals, tolerated on the fringes of capital so long as they do not detract from valorization, or so long as they are not usable as raw material in production processes. In contrast to the malice of the machines in the Terminator series, in this scenario humans would simply no longer be of interest to capital. According to this view, we have not seen capitalism yet.

144

inhuman power

Conclusion: Communist AI

At the end of the second decade of the twenty-first century, AI development is dominated by capital, led by some of the world’s most powerful oligopolistic corporations, enabled by and assisting nation states seeking instruments of economic competition in the world market and weapons for their military and security forces. ML, advanced robotics, predictive analytics and other fourth industrial revolution technologies are strengthening capital vis-à-vis labour, and elite sections of labour relative to others, and are hence likely to increase inequality along lines of class stratification that are also lines of gender and race. Deployment of AI around the social factory renders work and life in general increasingly opaque through the surrender of decision-making to proprietorial algorithms, while at the same time increasing levels of surveillance, precarity and the corporatization of education. While speculative assessments of the labour-market consequences of AI vary wildly, almost every prediction sees serious issues of technological unemployment on the horizon, and many admit the possibility of an escalation, quick or slow, to a more general crisis of work. Despite the social tumults that will attend all these circumstances, expectation that widespread AI adoption as a general condition of production would automatically lead to the end of capitalism is misplaced: on the contrary, there is as good a chance that it will open the way to a capitalism that continues without humans.

We have contrasted this analysis of the AI question with two left perspectives that we named ‘minimalist’ and ‘maximalist’, but that could equally well be called ‘this isn’t really happening’ and ‘this is really happening – let’s speed it up’.

The ‘this isn’t really happening’ view on AI suggests that predictions of increasing AI powers and widespread adoption are hugely exaggerated, largely amounting to investment-attracting hype and worker-intimidating bluster. Past predictions of ‘a jobless future’ have been falsified: AI will endlessly repeat its ‘spring/winter’ cycles of high hopes and disappointing actualities. Whatever technological realities underlie AI-capital’s promises and threats are not sufficient to alter the sober realities of capital’s continued large-scale dependence on exploited human labour. There are therefore no immediate major implications of AI for socialist or communist struggles.

This position raises important points about uneven and contradictory technological change within capitalism. Straightforward extrapolations from technological powers to social consequences are mistaken. However, to the degree that the ‘this isn’t happening’ view is based on a general scepticism about capital’s technological capacities, the past provides some ratification (no jet packs yet!) but also dramatic counterexamples (nuclear weapons, the internet, biotechnologies). Our analysis of actually-existing AI-capitalism shows increasingly pervasive corporate and state use of narrow, usually ML-based, AI technologies that are already objects of social struggle. More are being actively researched. That many of these will fail is certain; an AI bubble will probably burst. But in the longer term more intense and wider AI deployments seems likely, not just because of advances in technical innovation, but because of these innovations’ intersection with capital’s changing structural conditions, including the increased difficulties of advanced capital in accessing global cheap labour; tightening, and hence wage-raising, post-recession job markets; growing economic and military antagonisms between ascending and descending imperial powers; and the ruling class’s search for technological control over unrests arising from heightening social inequalities and deepening ecological catastrophe. There are compelling historical examples, from the fire and blood of primitive accumulation to the tumults of capital’s successive industrial revolutions, of how combined technological and social logics remake or unmake entire modes of production.

This brings us to the ‘it’s really happening, let’s speed it up’ position, aka the left-accelerationist or postcapitalist view of AI, which fuses impulses ranging from the philosophical to the pragmatic. This school of thought has behind it the weight of much of Marx and Marxism in seeing socialism as the inheritor of capitalist modernity and its techno-tools. Fully automated luxury communism, or postcapitalism, envisages a transition to socialism by reducing or eliminating the need to work and supplying a universal basic income. The xenofeminist wing of accelerationism sees a reduction in unwaged domestic toil by applying fourth industrial revolution technology to the home. The productivity generated by new technologies will create conditions of abundance that in large degree smother the need to replace markets with complex social planning mechanisms. Where planning continues to be required, for example in regard to ecological questions, intelligent machine networks monitoring outputs and inputs will be capable of solving the ‘calculation problem’ that baffled previous socialisms.¹ This might seem like utopian speculation, but such ideas are in play in recent left electoral politics. Left accelerationists (Srnicek and Williams), postcapitalists (Mason) and luxury automated communists (Bastani) are an important intellectual grouping for Jeremy Corbyn’s Labour Party in the UK (Dinerstein and Pitts 2018; Pitts and Dinerstein 2017). We admire and respect this group’s grasp of the scale and speed of AI-related technological change in contemporary capitalism. Such changes lie at the core of an untidy nexus of social conflicts emerging around AI-related precarity, work monitoring and speed-up; surveillance; military and paramilitary AI applications; algorithmic discrimination; smart cities; and oligopolistic concentrations of digital power. However, we are critical of the left-accelerationist/postcapitalist strategy for addressing the rise of AI-capital (since one of the authors of this book has written in a similar vein [Dyer-Witheford 2014], these comments can be taken as self-criticism as well as criticism!). To explain our objections, we take a step back to discuss some theoretical considerations arising out of what has become known as ‘the reconfiguration debate’.

146

inhuman power

THE RECONFIGURATION DEBATE

Major passages in Marx’s work predict that capital’s compulsive technological development will not only eventually kill it, but also leave a machine legacy for socialism to inherit and put to emancipatory use. Yet many Marxists have found a tension between these promises and other sections of Marx’s writing emphasizing the dominative, life-crushing powers of capitalist machinery. These issues have come to the fore in the ‘reconfiguration debate’, which began with an exchange between Alberto Toscano (2011) and Jasper Bernes (2013) on the growing importance to capital of vast logistical systems – systems that are, we note, one of the prime sites of actual and probable AI deployment, be it in automated distribution centres, robo-trucking or drone-delivery schemes.

Toscano (2011) criticized the anarchist group The Invisible Committee for looking at logistical systems, such as high-speed train systems and electronic networks, solely as targets for ‘sabotage’ and ‘hacking’ without a longer-term perspective on the ways in which such technologies could be ‘refunctioned’ as components of a future non-capitalist social order.

conclusion: communist ai

147

Bernes’s response came in a paper arising out of the blockade of the Port of Oakland in 2011, a moment in the wider US Occupy movements. He pointed to how logistical infrastructure, such as a high-tech port facility, is implicated both in the digital deskilling of work to mindless button-pushing, and in capital’s search for cheap labour. For workers to take command of such systems, ‘to seize, in other words, the control panel of the global factory’, would, Bernes (2013) asserted, be to assume management of a system constitutively hostile to them. The very design of end-to-end logistics is predicated on the ‘high-volume and hyper-global distribution’ of the commodities whose necessity was precisely what should be thrown in question by social revolution. Finally, Bernes posited that the idea of taking over global systems evaded the necessarily local nature of revolts against capital, revolts which would, at least initially, be isolated within a still-dominant world market, and have to make do with technologies at hand, rather than immediately taking over global infrastructures. Invocations of ‘the literal deus ex machina of supercomputers’ to establish a new social order would thus be largely beside the point (Bernes 2013).

Toscano (2014) replied to Bernes, suggesting that the revolutionary reconfiguration of capitalist technological systems did not preclude a critical evaluation of their specific component sub-systems, and criticizing as romanticism Bernes’ suggestion that emancipation required technologies yielding ‘the transparency of the small community or commune’. The debate has subsequently ramified in a number of directions (Degenerate Communism 2014; Cuppini, Frapporti and Pirone 2015; Chua 2017), including a direct entry into discussions of left accelerationism. Srnicek and Williams, in their argument for a full-automation socialism, acknowledge the issues Bernes raises, but suggest that ‘repurposing’ capitalist technologies is a matter for pragmatic experimentation (2015: 145–53). Bernes (2018) has criticized this ‘mix-and-match’ theory of transition, in which revolution can ‘discard unusable technologies (nuclear weapons: bad) and cultivate useful ones (antibiotics: good)’. He suggests this is a view that thinks of technology as ‘discrete tools, rather than an ensemble of interconnected systems’, and reiterates his fundamental challenge to accelerationist thought: ‘The standard assumption among Marxists and many others is that, despite its toxic excretions, the more developed technology becomes, the easier it will be to produce communism. But what if these technologies actually make it harder?’

148

inhuman power

AI is arguably the ultimate test for the reconfiguration debate, because general AI would be the ultimate technological system. The question of ‘refunctioning’ AI is bigger than that of reconfiguring logistics systems, which have in some ways become a subset of AI, and at least as large as the problem of energy system design, with which AI is, as we will see, closely bound up. The issue is not the reconfiguration of specific algorithms, or the automation of particular jobs – though those are also what is at stake in disputes over the application of specific narrow AI’s – but the trajectory of a far more universal technological project, one that is, it can be argued, a particularly capitalist project. For while the dream of automata dates back at least to Aristotle, only capitalism built into itself a systemic imperative to recruit labour, replace it with machines, accelerate markets, and animate commodities so that their rendezvous with purchasers becomes increasingly self-propelled and auto-guided, up to and including the automation of the very act of purchasing. To simply say that no other economic system has enjoyed the technical-scientific capacities to do these things begs the question of the reciprocal interaction of economics and innovation. As we argued in the Introduction, the real subsumption of labour by capital means that capital develops and adopts technologies that fit its systemic requirements of valorization; this imperative can be baked into the very design of technology.

Under capital, processes of fetishism and reification, by which, in an inverted world, things assume the appearance of human powers while people are treated as things, become real abstractions: AI is the concrete manifestation of that abstraction. Through this process, AI embodies the contradictory potential of capital, as, to quote Fredric Jameson (1991: 47), ‘both the best and the worst thing’ that could happen to humans: it offers humanity freedom from the exploitation of labour for capital, but also capital freedom from a humanity that becomes a biological barrier to accumulation. In our view, left accelerationism ignores the second part of this dialectic.

A COMMUNIST ORIENTATION TO AI

Left accelerationism’s blind spot is reflected in a series of proposals that would, we think, do more to accelerate AI-capital than outdistance it with AI-socialism. These include support for a universal or guaranteed income as an answer to technological job loss; neglect of the ecological problems of intensive AI use; and a confidence in a pacific transition to socialism that overlooks the military and repressive aspects of AI. What is critical to a communist orientation to AI is the issue of the ownership and control of the means of production, a point that postcapitalist and left-accelerationist thinkers partially take up, but also obscure by their insistence on the possibilities of passage to high-technology socialism by parliamentary reforms undertaken within the framework of actually-existing AI-capitalism. We address these issues in turn.

conclusion: communist ai

149

The ‘AI Plus UBI’ Formula

Proposals for a universal basic income (UBI) as a transitionary component of an anti-capitalist programme have been discussed for many years (and supported by one of the authors of this book). One conclusion of these debates is that UBI’s political valence, whether as a permanent strike fund for labour or a streamlining of neoliberal welfare austerity, depends on the terms under which it might be instituted, e.g., generous or stingy, with or without a dismantling of other social benefits. The more basic issue, however, is that, introduced within capital, UBI does not disturb the ownership of the means of production, and in some ways endorses it, as a manifestation of the munificent largesse of the ruling class, offered within the context of an otherwise fully commodified economy (Clarke 2017). These issues have been recently highlighted by a sudden wave of enthusiasm for UBI among Silicon Valley capitalists, who are advancing the idea specifically as an antidote to the unemployment and precarity likely to be caused by AI. Facebook founder Chris Hughes (2018), venture capitalist Marc Andreessen, and web guru Tim O’Reilly, among other Silicon Valley luminaries, support UBI as the ‘social vaccine of the 21st century’; the tech incubator Y Combinator is running a basic income research programme in Oakland; and tech entrepreneur Dan Yang has announced an independent US presidential candidacy on a platform that includes a form of basic income (Ghaffrey 2018; Ito 2018). These proposals tend to envisage UBI, often at a fairly low subsistence level, as an addition to an otherwise nakedly laissez-faire market order. As several observers (Filoux 2018; Sadowski 2018; Rushkoff 2018) have remarked, these proposals do not challenge the right of capital to direct AI development, reaping billions, and do not disrupt the vast income inequalities, either between capitalists and workers (or non-workers), or between elite professional high-tech employees and menial workers. And while UBI is promoted as an aid to entrepreneurialism, in reality it would probably require supplementation by forms of precarious work, in that respect actually subsidizing AI-driven gig-economy ventures such as Uber, Mechanical Turk or Figure Eight. Further, UBI makes no provision for other public supports, which might make the free time supported by UBI something other than a miserable penury. In this form, UBI figures as a holding pen for what Harari, whose gloomy futurism is favoured by many Silicon Valley capitalists, ruthlessly characterizes as a ‘useless class’ (2016: 379). This is certainly not the UBI that left accelerationists and postcapitalists want, but it is likely the type of UBI they would get under the auspices of AI-capital. In this regard we agree with the point made by Alex Gourevitch and Lucas Stanczyk (2018): a basic income of some sort might be an important part of a mode of production beyond capital, but it is not a prelude to it – rather, a trap waylaying its emergence.

150

inhuman power

AI’s Dirty Secret

In the loudly proclaimed ethical and safety-conscious deliberations of leading AI-capitalists, attention is now given to AI as an ‘existential risk’ (Bostrom 2014: 4). Such risk arises largely because of the possibility of an AGI evolving into an ASI beyond human control. The issue is not malevolence (Skynet) but rather efficiency. Nick Bostrom’s (2014: 123–5) famous example is of an AI instructed to make paper clips that attains superintelligence and uses its ever-extrapolating powers to convert the entire universe into paper clips, obliterating humanity as collateral damage. The point is a serious one, even if currently (we hope) remote; not only could ASI’s ‘integral accidents’ – Paul Virilio’s (2000) term for malfunctions so intrinsic to a given techno-system they must be considered a feature, not a bug – be devastating, they might not be ‘accidents’ at all.

However, Bostrom’s prediction – and other similar warnings, such as that of nanotechnologist Eric K. Drexler’s (1987) earlier ‘grey goo’ scenario, in which out-of-control self-replicating nano-robots consume all biomass on Earth while building more of themselves – can also be interpreted in a wider sense. This is to understand them not literally but metaphorically, or rather both literally and metaphorically, as simultaneously identifying a concrete hazard and providing a parable of runaway economic growth and universal commodification. The real ‘paper clip is profit, and the manifest danger of AI is not only that of an ‘integral accident’ but equally or more of its intended use as a means of intensifying and accelerating the production and circulation of goods that is destroying the environment, annihilating species, and, for humans, heating the planet to civilizational, perhaps existential, limits. This untrammelled economic growth is of course the very profit-maximizing process that would lie behind an unlimited order of paper clips, or of self-replicating automata, so the specific and general form of ‘accident’, or rather efficiency, are related – capital itself constitutes an existential risk.

conclusion: communist ai

151