---
schema_version: "library-reader-v1"
source_type: "original_book"
title: "人工智能和资本主义的未来"
author: null
edition: null
publisher: null
year: null
isbn: "9780745338613"
library_document_id: "a2eab101-adfc-4a0d-8142-63a4d11cee02"
source_sha256: "15c6139a50a0b26b71d6ea595faa034f40b5f04c9236d1b2341bc033fc1d60fc"
pdf_physical_page_count: 225
pdf_physical_page_index_base: 0
range_start: 80
range_count: 10
html_url: "https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02?start=80&limit=10"
markdown_url: "https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02.md?start=80&limit=10"
previous_url: "https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02.md?start=70&limit=10"
next_url: "https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02.md?start=90&limit=10"
---

# 人工智能和资本主义的未来

> 页码规则：PDF_PHYSICAL_PAGE_INDEX 是从 0 开始的权威文件物理页索引；PRINTED_PAGE 是页面中识别出的印刷页码，可能为空，二者不得混用。

## PDF_PHYSICAL_PAGE_INDEX=80 | PRINTED_PAGE=72

Automation, supply chains and financialization gutted factories, stalled wage growth and extinguished industrial militancy in the global North. Combined with political measures to weaken unions by governments such as those of Reagan and Thatcher, they broke the strength of the mass worker. As once prosperous cities became industrial rust belts, employment within the zones of advanced capital increasingly moved to the service sector, a term covering manifold fragmented and divided forms of work, from finance to trucking to domestic care, some well remunerated, others pitifully paid, or at best with stagnating real wages. There was also a change in the gender composition of labour as households became increasingly dependent on women performing both paid and unpaid labour, a shift that many left organizations have even yet failed to come to terms with. An entire political culture based on the power of industrial labour, its unions and socialist parties sank, like Atlantis sliding beneath the sea.

With globalization, the social factory in some ways becomes a ‘planet factory’, with a hugely expanded circulatory apparatus of highways, container shipping, air transportation, warehouses and distribution centres, fibre optic cables and data centres, moving goods, people and information around the world. Proponents of globalization, and even some autonomists (Hardt and Negri 2001), speak of a ‘flat’ or ‘smooth’ planetary economic order. But in fact the capitalist planet was divided by sharp hierarchical zonings. Advanced regions, such as North America, Europe and Japan, housed business headquarters, financial activities, high-technology research and design, and retail sales and consumption complexes. Newly industrializing areas, especially China, became centres of relocated manufacturing. Petro-states, from Russia to Saudi Arabia, fuelled the carbon economy. Other regions sank into immiseration, with survival depending on informal or criminal activities. Class composition became hugely complicated in an increasingly unified world market that nevertheless divided workers whose standards of living and conditions of social reproduction remain universes apart (Woland, Blaumachen and Friends 2014). We therefore continue to refer to the ‘social factory internally constituted by regional arrangements, even while recognizing their global articulation.

72

inhuman power

## PDF_PHYSICAL_PAGE_INDEX=81 | PRINTED_PAGE=73

Globalization and digitization changed the balance of class power in capital’s favour. Around the world, labour’s share of GDP relative to that of capital steadily decreased, in rich and poor countries alike (Economist 2013). Then, in 2008, the planet factory burst into flames. Capital’s increasing focus on financial speculation culminated in the Wall Street subprime mortgage crash. Widespread contagion of financial systems, bank bailouts, austerity, recession in the global North and ramifying knock-on effects around the world followed. So too did social tumult. In 2011 a series of rebellions against inequality and insecurity – often in the form of public square occupations – ricocheted from Tunis and Cairo to Madrid and New York, and then to Istanbul and Rio de Janeiro, Tel Aviv and Kyiv.

While these revolts initially indicated a burgeoning working-class recomposition, often drawing on new strata of post-Fordist cultural and technological workers and students, they were almost all unsuccessful; they fizzled out, were repressed, replaced one conservative regime with another, or imploded into civil wars and foreign interventions. They were followed by another wave of turmoil, driven from the right: the UK Brexit referendum, Trump’s 2016 US election victory, and the rise of neo-fascism across Europe, all propelled by resentments against capitalist globalization, racist hatred of migrants and minorities, and working-class anxieties around job losses and flagging wages.

Both the unsuccessful recompositionary moment of the Occupy movements of 2011 and the all too successful decompositionary neofascism of 2016 were symptoms of a weakened working class; but they also signalled an alarming instability within capitalism. The wave of corporate interest in AI, though long in the making, surfaced towards the end of this roller coaster of unrests. Machine intelligence was framed as the ‘answer’ to crisis. Thus consulting firm Accenture began a 2016 pitch for AI investment by acknowledging a ‘marked decline’ in the ability of capital to ‘propel economic progress’ and ‘sustain the steady march of prosperity enjoyed in previous decades’ before proclaiming that ‘AI has the potential to double … annual economic growth rates by 2025’ (Purdy and Daugherty 2016).

Paradoxically, the digital industries that enabled globalization were strengthened by its 2008 near-death experience. With Wall Street reeling and Main Street industries, such as the big three US automakers, mired in recession, ‘big tech’ was the one beacon of hope. Having survived the dot.com meltdown of 2000 by reincarnating as ‘Web 2.0’, digital capital was already on the ascent before the Wall Street meltdown. Google first offered stock market shares in 2004 and Facebook in 2006. Apple released the iPhone in 2007. It was, however, during the post-crash recession that these digital corporations, along with others such as Amazon, topped the lists of the most highly valued companies in the world, occupying positions previously held by giant energy and financial service corporations (Mosco 2017: 65). Even uprisings against global capital, hopelessly simplified as ‘Facebook revolutions’, could be interpreted as manifestations of a new digital order. The US return from recession after 2008 was a ‘tech-led’ recovery, during which the preconditions for ML were established: big-data-generating social media, web and smartphone use.

automating the social factory

73

## PDF_PHYSICAL_PAGE_INDEX=82 | PRINTED_PAGE=74

Very specific aspects of the crisis of globalization boosted interest in AI. Amid the noise of round-the-world protests, one clear message could be detected: the days of cheap global labour were fading. While revolts in the global North West failed, something else was happening in the South. New generations of industrial labour were finding their own political composition. In China, successive strike waves in the factories of the Pearl River Delta steadily drove wages up. In 2010 protest-suicides in the giant electronic plants of Foxconn, where Apple’s iPhones were assembled by a workforce under quasi-military discipline, flashed around the world. Foxconn’s exasperated owner, Terry Gui, threatened to replace his assembly workers with one million robots. With the first influx of migrants from rural villages to cities exhausted, China’s factories found it increasingly difficult to attract young people to assembly line drudgery. Average hourly wages for China’s manufacturing workers trebled between 2005 and 2016, approaching the hourly rates paid in countries like Greece and Portugal (Gao 2017). This was only one sign of wider problems for globalized capital: in India, striking auto workers burned factories; in Cambodia, garment workers confronted police in the streets. With the sudden implosion of the financial fix, and the slow erosion of the spatial fix of worldwide low-wage locations, capital started to look at a new, more intense technical fix: AI automation.

Other, yet more violent crisis-events drove AI forward. The ‘9/11 destruction of New York’s World Trade Centre in 2001 and the subsequent ‘war on terror’ had stimulated the development of ML technologies to manage vast surveillance systems. Wars in Afghanistan and Iraq continued this momentum, adding drones and other autonomous vehicles using rudimentary machine intelligence to the US counterinsurgency arsenal. Military interest escalated when popular uprisings in Ukraine, Syria, Libya and elsewhere collapsed into civil wars and foreign interventions. As the US, Russia, China, Israel, Saudi Arabia and Iran squared off in shadow-wars, research and application of ML for surveillance, hacking and cyber-security also accelerated (Sanger 2018; Scharre 2018: Dyer-Witheford and Matviyenko 2019). Just as the Cold War had incubated digital technologies that then found their way into civilian production, so too did New Cold War weaponry work its way back into the front lines of capital’s confrontation with labour, as drones and improved robots entered the workplace.

74

inhuman power

## PDF_PHYSICAL_PAGE_INDEX=83 | PRINTED_PAGE=75

As significant to the AI industry was the widespread perception that US and European capitalist liberal-democracy was in trouble. Precarious workers occupying Wall Street and rust belts voting for Trump or Brexit brought home the point that the advanced capitalist economies were not delivering growth sufficient to both give corporations big profits and raise working-class living standards. Despite utopian promises about computers and networks, digital technologies had not been able to repeat this amazing double act of Fordism. Silicon Valley’s response was to double-down on its cybernetic bet. ML and advanced robotics would be magic bullets for renewed economic growth (Purdy and Daugherty 2017; Swanson and Mandel 2017). Such promises are appealing. What they obscure is that for AI investments to be undertaken by capital, they have to ensure not just economic growth, but profitable growth, a surplus well over anything accruing to labour. And this might require making labour more, not less disposable, and less, not more capable of claiming improved wages and working conditions and of exercising any type of autonomy vis-à-vis capital. It could, in short, entail further decomposition of the working class, which arguably never recomposed after the 1970s.

THE LABOUR OF AI

It seems paradoxical to speak of the labour of AI, a technology that purportedly eliminates human work. Yet AI itself has to be produced before it is applied in other industries. This is both a precondition for the AI-driven transformation of other types of labour, and provides some clues as to the direction that a more general recomposition of work may take. There are many different kinds of AI research currently in play, and while there are broad similarities in their production processes (e.g.

automating the social factory

75

## PDF_PHYSICAL_PAGE_INDEX=84 | PRINTED_PAGE=76

dependence on highly trained software engineers and programmers) each has its particularities. Here we focus on ML, the leading form of AI in the second decade of the twenty-first century.

Catherine Dong (2017) explains that ‘Machine learning engineering happens in three stages: data processing, model building and deployment and monitoring.’ The ‘meat’ in this ‘sandwich’, she says, is the building of the model. This is where software engineers develop the algorithms that recognize given input data, be it digital images of cats, or of pedestrians crossing streets, or configurations of search terms, or purchase records, or social media topics. This learning process is tested, tweaked and improved over thousands or millions of training examples, and then further tested on raw data, where the pattern signal is mixed with noise (images of cats with those of dogs, antelopes and elephants; pedestrians alongside wheelbarrows, letter boxes and unicyclists, and so on). At the end of this process the AI should be capable of identifying its target object with an acceptable level of confidence and of establishing statistical correlations between different patterns (more pedestrians crossing streets in daytime than at night; cat images posted more frequently by women than men, etc.). It is these capacities that give ML its uncanny, inhuman ability to predict associations and outcomes.

There are many techniques to construct algorithmic models – ‘linear/ logistic regression, random forests and boosted decision trees’ – familiar to computer scientists for some time (Dong 2017). These are used to create the friend-suggestions, ad-targeting, recommendations and search rankings that are part of everyday internet use. However, the most sophisticated ML modelling involves ‘deep neural networks’ that are ‘hard to train ... require[ing] more time and computational power’ and ‘a combination of intuition and trial and error’ (Dong 2017). While courses and do-it-yourself instructions in ML are proliferating, and libraries of open-source tools widely available, deep expertise with neural networks is rare; capable computer scientists may be numbered only in the thousands worldwide (Economist 2017b). This accounts for the intense corporate bidding wars for such talent and the high salaries and benefits it can command. ML thus intensifies the basic software development labour process in which core activities are conducted by small groups of highly skilled techno-scientific workers.

While the ‘meat’ of ML is the making of the algorithmic model, it is ‘the bread of the sandwich process’ – what happens before and after training the algorithm – ‘that holds everything together’ (Dong 2017). These two layers of the ML process involve, at one end, ‘cleaning and formatting vast amounts of data to be fed into the model’ and, at the other, ‘careful deployment and monitoring of the model’. As Dong observes: ‘Any incompatibility from any stage of the ML development process – from data processing to training to deployment to production infrastructure – can introduce error’ (2017). In practice, ML software engineers often spend more time dealing with ‘preparing and monitoring’ than in actual model making. Later in this chapter we will address the sources of such data sets, from unknowing mobile phone users to closely monitored distribution centre workers and the sleepless ‘safety drivers’ in driverless vehicles. But regardless of how data is collected, it nearly always has to be cleaned for use in AI training: standardizing its format, eliminating errors, detecting outliers, supplying additional information, adding features used in processing and above all labelling objects that ML will learn to recognize. Amongst data scientists, jokes such as ‘80% of the job is cleaning data and the other 20% is complaining about how the data has been cleaned’ abound (Lohr 2014).

76

inhuman power

## PDF_PHYSICAL_PAGE_INDEX=85 | PRINTED_PAGE=77

Much of this ‘data munging’ or ‘data wrangling’ (Johnston 2015) is outsourced, often as offshored microwork, in low-wage zones from Venezuela to India and Indonesia, at rates ranging from a few dollars to a few cents an hour (Nakashima 2018a). At first this was handled through crowdsourcing platforms such as Freelancer, Mechanical Turk or Upwork. Then data cleaning became a specialized business. Start-ups like Mighty AI and Figure Eight (formerly CrowdFlower) develop software that ‘makes it easier to label photos and other data, even on smartphones’ (Nakashima 2018a). Playment, a Bangalore company, ‘gamifies’ such tasks and calls its 30,000 on-call workers ‘players’. In China, firms specializing in ‘refined data processing’ work with a Taylorist division of labour. Some ‘project groups’ are responsible for labelling ‘nodes on the human body in complex yoga-like postures’, or annotating ‘motorcycles, bicycles and pedestrians’ showing their travel directions for autonomous vehicle software; others ‘scan the contours of buildings or obscured objects, used to train radar to detect real-world objects’ (Yu 2017). The work is demanding and tiring.

The other layer of the ML ‘sandwich’, deployment and monitoring, also draws on global clickwork. ML algorithms may not work as anticipated, or encounter problems requiring intervention. How much purportedly automated systems depend on human backstops was highlighted in 2017, when Expensify, a company offering to automatically scan photos of receipts to extract data for expense reports was discovered to have bridged failures in its text analysis software by sending customers’ personally identifiable receipts to workers on Amazon’s Mechanical Turk (MTurk) crowdsourcing platform. This kind of ‘human safety net’ behind AI-powered services is very common (Gent 2017). Sarah Roberts (2016, 2017) has identified the role of generally poorly paid and precarious human content moderators in filtering social media feeds to purge violent and pornographic content. Google is reported to have an army of 10,000 ‘raters’ watching YouTube videos. Microsoft operates a Universal Human Relevance System, checking the results of search algorithms with millions of micro-tasks a month. Facebook and Google remediation of ML outcomes was recently intensified by ‘cyber-war’ scandals (Alba 2017; Dyer-Witheford and Matviyenko 2019). The microwork of data cleaning and monitoring are closely related; both may be performed by the same low-paid contract home-worker on different shifts (Nakashima 2018c).

automating the social factory

77

## PDF_PHYSICAL_PAGE_INDEX=86 | PRINTED_PAGE=78

Such activities illustrate ‘the paradox of automation’s last mile: as AI makes progress, it also results in the rapid creation and destruction of temporary labor markets for new types of humans-in-the-loop tasks’ (Gray and Suri 2017). There is a further twist in this paradox: data cleaning and monitoring have themselves become targets for ML automation. Amazon’s Mechanical Turk Human micro-tasking is referred to by the corporation as ‘artificial artificial intelligence’ (Reese and Heath 2016); the irony is that ‘turkers’ train real artificial intelligences that may one day replace them (Hook 2016). Facebook and Google executives are explicit that their goal is to eventually automate content moderation (Alba 2017; Nieva 2018). Data scientists have declared data cleaning ‘a machine learning problem that needs machine learning help’ (Ilyas 2018). Clickworkers are contributing data sets of their own digitally recorded labour as training material for future AI.

ML thus exemplifies what Tim Jordan (2015: 32–8) terms ‘recursion’, the process by which information capital feeds into itself. Fear of automation has worked backed to the elite strata of software developers. AI industry workers exhibit a paradoxical relationship to automation in their work; while they are wary of it, they also describe it as necessary and even desirable (Steinhoff 2019b). This then spurs further disagreement as to whether such technologies will enhance or extinguish the work of software engineers (Shanin 2018; Vorhies 2016a, 2017). In 2016, 550 developers were asked to identify the ‘most worrisome thing in their careers’ from a list of possible concerns: ‘I and my development efforts are replaced by artificial intelligence’ was the most popular answer (Thibodeau 2016). Some developers dismiss such alarms, believing AI industry growth will only increase demand for human coding and design. Others foresee software developers increasingly assisted by AI ‘bots’ that automate code checking, distribution and repair, and will increasingly ‘spin up’ applications in response to natural language commands without developers ‘even touching the keyboard’, escalating to a point where the ‘entire software development lifecycle will be only a high-level description’ (Despoudis 2017). Such discussions have been further stirred by the recent BAYOU ML applications, developed at Rice University for the US Department of Defense; this is a neural network that ‘trained itself’ on millions of lines of human-written code on GitHub, and can reportedly recommend chunks of code to developers on the basis of very brief keyword instructions, thereby taking initial steps towards ML-designed ML systems (Rayome 2018).

78

inhuman power

## PDF_PHYSICAL_PAGE_INDEX=87 | PRINTED_PAGE=79

The ML production process has a double significance for the analysis of class composition. First, it is an important part of an expanding occupational sector of software development crucial to digital capital. Discussion of this sector conventionally focuses on the well-paid, high-status labour of software engineers and programmers, of whom ML designers are the latest incarnation. This obscures the work conducted ‘behind the AI curtain’ by data cleaners and monitors (Gray and Suri 2017), part of a global clickwork-force that a UN survey found characterized by ‘low pay, poor prospects, and psychological toll’ (Berg et al. 2018; Geuss 2018a). All parts of ML’s segmented workforce confront a horizon where the very product they create may automate their labour, so that data scientists and data cleaners may both be working themselves out of a job. Second, while ML accounts for only a fraction of global labour, the patterns it discloses of digitally mediated labour, subject to powerful polarizations, and rendered insecure by automation, also appear where the new AI is applied across the wider social factory.

LEARNING THE WAY OF THE HUMAN

The social function of AI as fixed capital in production, like any other labour-automating technology, is the reduction of necessary labour and the concomitant increased extraction of surplus-value. In circulation it accelerates the realization of value, i.e. reduces circulation time, by enabling a tighter integration (coordination and scheduling) between factories, distribution centres and retail stores, and more precise targeting of both seller and advertisers to potential purchasers. It also reduces or eliminates the labour costs of all these activities. In the finance sector, automation speeds the speculative metamorphosis of money into more money. And as AI impinges on consumption and social reproduction, it both eliminates waged workers in industries such as health care, education and entertainment, and establishes new patterns of machine-mediated interpersonal relations (such as social media), or human-machine relations (such as elder-care robots).

automating the social factory

79

## PDF_PHYSICAL_PAGE_INDEX=88 | PRINTED_PAGE=80

How AI is introduced around the planet varies according to capital’s spatial organization. In North American and European economies, increasingly dominated by services, retailing and financialization, AI initiatives are strongly oriented towards circulation, although they also attempt to repatriate industrial production. In China, Southeast Asia and other manufacturing zones connected to export markets, AI is strongly oriented to automating industry, though the breakneck speed of economic development also incentivizes circulation-accelerating AIs deployed by companies such as China’s Baidu, Tencent and Alibaba. In the sacrifice zones of the planet, the AIs most immediately in play are those of domestic and foreign security forces deploying drones and surveillance systems to control immiserated populations. Here we focus primarily on how AI automation is changing the technical composition of labour in North America and Europe, but taking into account developments elsewhere. Capital’s AI deployment is incipient, at a stage of experimental introduction, accompanied by large promises. As Kim Moody points out, in the aftermath of the 2008 crash, business investment in new technologies has been subdued by low profit rates and repressed wage growth; while these conditions persist, the advance of AI automation is likely to be ‘bumpy and slow’ (2018a: 26). Over the same period, however, corporations have been testing, and using, AI applications in many spheres; as capital’s previous waves of technological innovation suggest, this process is likely to escalate if and when capitalists’ confidence or workers’ wage demands increase. Some prospects may prove far more difficult to implement than others, or even be completely blocked by unanticipated technological or regulatory problems – or, perhaps, working-class resistance. Yet even if only a portion are successful, AI will significantly change the balance of power between proletariat and capital.

80

inhuman power

## PDF_PHYSICAL_PAGE_INDEX=89 | PRINTED_PAGE=81

In the realm of production, AI in the global North West is hailed as a means of re-industrialization. Manufacturing capacity offshored to foreign locations, or threatening departure, is to be recovered by the injection of robots, sensors and process control computing, all enhanced by ML. While ‘[m]any industrial AI applications are still somewhat new and bespoke’, powerful companies such as GE, Siemens, Microsoft and Intel are ‘all making significant investments’ in applying ML in manufacturing to bring down labour costs, reduce product defects, shorten unplanned downtimes, improve transition times, increase production speed and conduct predictive maintenance (Faggella 2018b). Siemens has reportedly been using artificial neural networks ‘to monitor its steel plants and improve efficiencies for decades’ (Walker 2018). In Europe, this project marches under the name of ‘Industry 4.0’, while in North America it often goes under the broader heading of ‘the Internet of Things’.

Such industrial automation is most flamboyantly and problematically exhibited in Elon Musk’s Freemont, California factory producing Tesla vehicles (Bhattacarii 2017) – a facility planned not only to produce self-driving cars and trucks but also to itself be so highly automated as to resemble, in Musk’s words, an ‘alien dreadnought’ (DeBord 2017). This craft’s operations have, however, been plagued by scandals over injuries to the workers it hopes to dispose of completely (Wong 2017a, 2017b); production bottlenecks Musk concedes were caused by ‘excessive automation’; digital sabotage by disgruntled employees; and the crashes and burnings of auto-piloted vehicles that proved better at self-combustion than self-steering. Tesla indicates the difficulties of extreme automation (Wilson and Daugherty 2018). Nonetheless, rising trade protectionism continues to fuel plans for plants tethered to the homeland, even if they employ more patriotic robots than domestic workers. It is, however, an irony dictated by capital’s competitive logic that China’s corporations, facing labour shortages and wage pressures of their own, are moving in the same direction, and probably faster. Terry Gui’s vow to robotize Taiwanese-owned but China-based Foxconn has by no means been completely fulfilled, but nor has it proven totally empty; by 2018 so-called ‘Foxbots’ were a feature of production lines at Foxconn, as the company expanded its revenues while diminishing its workforce (Chan 2017; Wang 2018). Since 2013, China has been the world’s largest market for industrial robots. The majority are purchased abroad, but with growing domestic production, its factories are becoming as or more automated than those in the West (Bland 2016; Ying 2018).

automating the social factory

81

下一段：[继续读取](https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02.md?start=90&limit=10)