人工智能和资本主义的未来

It was, however, in the sphere of circulation that capital’s AI initiatives developed most quickly in North America and Europe, accelerating the ‘logistics revolution’ that has linked transportation, warehousing, communications and retail sales in increasingly complex and seamless systems (Cowen 2014; Rossiter 2016). As we have seen, autonomous vehicles are amongst the most potentially lucrative forms of AI-capital. This prospect is usually discussed in terms of the self-driving personal car (both a consumer commodity and a means by which the commodity labour-power gets to work). The immediate prize, however, is automated commercial trucking. Robo-trucks are already used on-site in industries such as mining. Companies like Volvo, Daimler and Tesla are working on their graduation to long-distance hauling. In 2016, Uber carefully staged a demonstration (under the most favourable conditions) of beer delivery by a self-driving truck (Peterson 2016). Since then such vehicles (accompanied by a watch driver) have entered use in the delivery of refrigerators from a warehouse in El Paso, Texas, along the I-10 freeway, to a distribution centre in Palm Springs, California.

Self-driving trucks, operating 24/7, without rest breaks, in convoys with over-watch, or possibly remote, drivers, on long straight highways free of pedestrians, cyclists and traffic lights or perhaps in special lanes, are a lucrative AI prospect. It would automate an important sector of working-class employment. In the US, trucking is a \$676 billion industry, employing about 3.5 million drivers. In a slim-margin business, labour costs ‘account for up to 45% of total freight costs’ (Campbell 2018). A 2017 report from the International Transport Forum forecast that up to 4.4 million of the 6.4 million truckers in the US and Europe could be eliminated by autonomous technology (ITF 2017). The inclusion of self-driving trucks in national vehicle regulations has been opposed by the Teamsters Union, representing almost 600,000 truckers, which describes their introduction as ‘calamitous’ (Marshall 2017). Such opposition is only one of a number of barriers facing AI-automated trucking: others include not only continuing problems in refining auto-driving technologies (as witnessed by a series of accidents in 2018), but also the need for adaptions of highways systems and the articulation of legal responsibilities (Moody 2018b). So, while robo-trucks are potentially one of the most profitable of AI-capital ventures, they are also one of the most challenging. There are, however, other more immediately available opportunities.

82

inhuman power

Less dramatic than cyber-rigs rolling down highways are the invisible software agents transforming capital’s circulatory processes. Around the globe, day and night, call centres send outgoing messages soliciting sales for everything from phone plans to museum memberships, and answer incoming calls maintaining, repairing and reinforcing customer relations (Brophy 2017; Woodcock 2017). In Europe and North America, they were promoted as replacing jobs lost to deindustrialization, though such work was then substantially outsourced to India and the Philippines. But as wages for repetitive phone-toil overseas rise, call centres are being AI automated. Customers (and workers) have long been familiar with the recorded message that ‘your call may be recorded for training and quality purposes’; that ‘training’ has come to include training ML systems (Marr 2016). AIs capable of applying natural language processing draw on digital banks of frequently asked questions, as well as knowledge of customers’ equipment, competence and past calls to automate answers to common questions, shuttling the more esoteric ones to the remaining human operators, sometimes while simultaneously analysing the callers ‘tone, vocabulary, sentiment, and even silences to gauge emotion and satisfaction’ (Agrawal, Gans and Goldfarb 2018: 90–1).

The market consultancy company Gartner declared in 2017 that ‘customer service spaces’ were ‘the number one case for applying AI’, accounting for ‘about 70% of all use cases in AI’ (Baraniuk 2018). In 2018, the British retail giant Marks & Spencer replaced 100 switchboard staff with chatbots using Twilio’s speech recognition software and Google’s Dialogflow AI tool to interpret and route customers’ verbal requests. The large US insurer Allstate adopted the digital assistant Amelia – billed as ‘the most human AI’ – to supplement information staff, reducing call durations from an average of 4.6 to 4.2 minutes – an apparently small saving that across millions of calls adds up to huge cuts in labour costs. Call companies like to term such changes ‘augmentation’ rather than ‘automation’ of human staff. For the moment, its main effect may be to intensify the work of human operators, but many observers believe automation will prevail in the long term (Wood 2018). Similar capacities are being extended to other corporate communications. In 2018, Google demonstrated how its Duplex system combines calendar search and natural speech patterns to schedule appointments by phone, with voice simulation difficult to distinguish from a human; it was promoted as ‘useful for the 60% of small businesses in the US that don’t already have an online booking system’ (Solon 2018).

automating the social factory

83

Automated transport will move goods to and from warehouses that are themselves already automated by AI. Amazon’s ‘fulfilment centres’ are the paradigm case. Amazon is a company that has made ML its ‘flywheel’ (S. Levy 2018) and has deployed it in the corporation’s algorithmic recommendation system, online purchases, smart home devices and cloud computing empire. It is also in the cavernous fulfilment centres where orders are received, picked from shelves, packed, and loaded to trucks, by notoriously poorly paid, speeded-up, stressed-out and electronically monitored workers. Demonstrating Marx’s thesis that capital’s machinery intensifies rather than diminishes exploitation, the heightened pressure on human ‘ambots’ has accompanied the deployment of real robots (Greenmeier 2008; Abdelrahman 2017). Amazon acquired robotics company Kiva Systems (renamed Amazon Robotics) in 2012 for \$775 million, inspired by the discovery that ‘fulfillment center workers were spending more than half their time walking around the warehouse to find items and put them in their tote’ (Agrawal, Gans and Goldfarb 2018: 145). Large-scale deployment of its round, low automata followed a Christmas crisis during which the corporation failed to meet promised delivery dates. By 2017 Amazon had more than 100,000 Kiva bots in select fulfilment centres worldwide (Wingfield 2017). Where these bots are used, they are directed by a constantly updated computerized inventory system and guided by motion sensors to bring goods to pickers stationary on a work platform. The robots have reduced the time to complete an order by a fifth. They save Amazon space – their squat profile allows fitting 50 per cent more inventory in a given area – and power bills; Kivas work in the dark and do not need air conditioning to prevent the heatstroke that regularly brought ambulances racing to fulfilment centres (Tam 2014).

Amazon’s fulfilment centre automation still has major gaps. One is the ‘picking’ of items from the robot-propelled mobile shelves. This may seem a routine task easily mechanized, but an Amazon warehouse, unlike an auto assembly line, deals with an ‘almost infinite variety of shapes, sizes, weights and firmness of items’ (Agrawal, Gans and Goldfarb 2018: 144). This has made picking an intractable automation problem. For several years an Amazon Picking Challenge invited major robotics companies to find a solution, without success. One recent line of research uses a ‘a mix of automated software and a human controller’, with the robot automatically navigating to a shelf where the human – wearing a virtual reality helmet, possibly in a remote location – guides its arm to grip and move the item; the long-term aim is to use a machine-learning artificial intelligence ‘trained on many observations of human grasping via teleoperation to teach the robot to do that part itself’ (Agrawal, Gans and Goldfarb 2018: 145).

84

inhuman power

Other companies may overtake Amazon. The UK online-supermarket Ocado is a leader in designing highly automated systems it sells to other grocery chains. Its warehouse system looks like a ‘a huge chessboard, populated entirely by robots’ that engage in tasks of ‘lifting’, ‘moving’ and ‘sorting’ in a ‘hive-grid-machine’; each bot has a central cavity and claws to grab and pull crates inside ‘like an alien abduction in a supermarket aisle’ (Vincent 2018). The bots are interchangeable, switching from one task to another; production can be scaled up or down by the addition or subtraction of bots, and their algorithmic programming allows novel solutions to warehouse organization. China’s warehousing firm JD.com, working in collaboration with Google, claims to have achieved even higher levels of automation (Paquette 2018). There are still plenty of workers in Ocado and Amazon facilities, but as bot systems deepen and spread, the pace of work increases and its security declines.

Circulation culminates in sales. Retail jobs too are registering the tremors of AI’s introduction. Again, ML builds on previous digital changes; online sales have transformed retail business, self-checkout lines are common, and both made deep cuts to in-store employment. AI extends this logic to the creation of staff-less automated retail outlets run almost entirely on downloaded apps and scanned products. These technologies were pioneered in Sweden, initially in remote-area grocery stores (Prindle 2016). They were then deployed by Swedish companies in China (Sun 2017), where mobile phone payment for daily shopping has been popularized through services such as WeChat, and no-staff stores, some operating on a relatively low-tech self-checkout basis, are common. The North American version debuted in Amazon’s Go stores which, as we have discussed, combine mobile phone apps, automatic scanning and ‘artificial intelligence and computer vision to match the face of the buyer with the items in one’s bag, to eliminate checkout altogether’ (Horwitz 2018). Such systems have the potential to include chatbots providing shoppers with ‘customized suggestions’, and IoT delivery orders from automatically self-stocking refrigerators and other smart home devices (Sun 2017; Pirrone 2018). Walmart, already famous for its massive system of computerized inventory control, is researching similar systems in partnership with Microsoft, a development that, given Walmart’s status as the largest single corporate employer in the US, has major labour market implications.

automating the social factory

85

Shopping requires money, and in an era characterized by the expansion of both debt and speculation, financial institutions have been amongst the earliest adopters of AI. ML methods are now widely used ‘to assess credit quality, for fraud detection to price and market insurance contracts, and to automate client interaction’ (FSB 2017). But it is the highest levels of the finance sector that provide the most striking examples of AI automation. At the time of the Wall Street meltdown, stock traders were already being decimated by the algorithmic agents of high-frequency trading. Since then, stock exchanges in Tokyo, Singapore, London and Hong Kong have eliminated their trading floors. By 2018 only some 10 per cent of stock market trades were actively conducted by humans; 40 per cent are ‘passive’ trades scheduled by mutual funds; the remaining 50 per cent are executed by algorithms that are increasingly informed by ML (Pearlstein 2018; Coles 2018). In 2000, Goldman Sachs equity trading employed some 600 people: by 2016, all but two had been replaced by ‘complex trading algorithms, some with machine-learning capabilities’ (Byrnes 2017). J.P. Morgan revealed plans for an AI focus (Terekhova 2017), and poached an ML expert from Microsoft to oversee this reorganization (Kolanovic and Krishnamachari 2017). The one-time ‘masters of the universe’ were being eaten by the tools they had helped create (Byrnes 2017).

Descending from the heights of financial manipulation to the depths of everyday scraping-by, we arrive at the role of AI in social reproduction. Algorithmic profiling is ubiquitous in corporate and state decisions on loans, insurance, medical claims and job applications. It is applied with special ferocity on those who fall on the margins or outside of waged labour, i.e. the reserve army of the unemployed and the surplus populations likely to be enlarged by AI automation. As Cathy O’Neill demonstrates (2016), such machinic decisions, by intention or oversight, persistently discriminate against all indications of penury and precarity, whether by postal code, ethnicity, gender or credit history. Virginia Eubanks (2017) documents how the use of automated decision-making tools to administer welfare benefits, social housing and family care interventions constructs an AI managed ‘digital poorhouse’ that denies or deters applications, subjects recipients to police scrutiny, and stigmatizes them on the basis of behavioural patterns and past associations. A corollary of such control is the well-documented (Larson et al. 2016; Lum and Isaac 2016; Ensign et al. 2017) deployment of AIs for predictive policing and sentencing, promoted as reducing human bias, but in fact algorithmically continuing and amplifying previous patterns of judicial and police racism in a ‘runaway feedback loop’ (Reynolds 2018). This is the sharp end of the more general process by which ML, but specifically deep learning (Impett 2018), becomes in the hands of capital the means for an ever more rigorous, comprehensive and recursively self-confirming sifting, segregating and disciplining of various grades of permanent, precarious and surplus labour-power – becomes, that is, an instrument of proletarian decomposition.

86

inhuman power

APOCALYPSE NOW OR BUSINESS AS USUAL?

Marx, as we have seen, understood technological unemployment as a weapon in the conflict between capital and labour. The issue of AI automation’s effects on employment is central to the analysis of class composition. Controversy about this topic has raged amongst bourgeois economists and technologists for the last decade. But unlike Marx, none of the participants in this mainstream debate want capitalism destroyed. On the contrary, all are all anxious to preserve it. The advocates of what we will call the ‘AI Apocalypse Now’ position believe an imminent jobs crisis calls for emergency measures to save the system, while ‘Business-as-Usual’ proponents point to historical precedent to claim that labour markets will adjust to AI with only relatively minor employment shocks.

The AI Apocalypse version is often presented by futurists with a computer science background, such as Martin Ford (2009, 2016). He acknowledges that capitalism has sustained employment through successive waves of technological change, from steam engines to electrical power, and early computerization. Nevertheless, Ford asserts that AI and other fourth industrial revolution technologies are different because of their exponential, Moore’s-Law-driven speed of improvements and cross-sectoral applicability. Previous technological transformations moved slowly from one industry to another. AI automation, Ford says, will have a near simultaneous adoption; workers displaced from, say, ‘Industry $4 { \cdot } 0 ^ { \ ' }$ manufacturing will be unable to find work in call centres because these too will be AI automated. MIT computer scientists Erik

automating the social factory

87

Brynjolfsson and Andrew McAfee (2014) make a similar argument, emphasizing how digital automation now extends beyond routine, manual ‘blue collar’ work to ‘white collar’ occupations such as journalism, advertising, lawyering, medicine and other middle-class professions. For capital to survive, radically new social policies would be required, including the introduction of a universal basic income or citizen’s wage (an idea we discuss in our conclusion to this book). Such prophecies seemed corroborated by a 2013 study from Oxford economists Carl Frey and Michael Osborne, suggesting that 47 per cent of the 702 occupations into which US jobs are conventionally sorted are ‘likely to be substituted by computer capital’ within two decades. Coming amidst AI exploits like Watson’s Jeopardy victory and the defeat of grandmaster Lee Sedol by Google’s AlphaGo, this study helped make the hailing of future robot overlords a virulent cultural meme.

Against this, the Business-as-Usual position is championed by professional economists (Autor 2015). It usually suggests that the achievements of narrow AI are mainly in formalized and/or highly predictable situations. These limitations, alongside the regulatory problems confronting technologies such as self-driving cars, will, Business-as-Usual theorists argue, slow AI adoption, making a sudden-onset, across-the-board crisis unlikely. More importantly, while automation destroys some jobs, it creates others. Some emerge in the very industries that build automating technologies, such as the data scientists and microworkers who make AIs. But other new employment opportunities emerge even in the industries AI automates. So it is suggested that autonomous vehicles may make truckers and taxi-drivers obsolete, but will create jobs for engineers designing hardware and software, as well as for workers ‘taking customer calls, cleaning and repairing cars, and updating ... high-definition maps’ (T. Lee 2018). To stay competitive, firms will have to pass some benefit on to consumers in the form of lower prices; this will increase real income levels, enhance consumption and raise employment. The ‘displacement effect’ of automation – workers being thrown out of jobs by AIs – will be offset by the ‘income effect’ of heightened demand for labour (Hawkesworth, Berriman and Goel 2018). Business-as-Usual advocates admit the possibility of a ‘difficult transition’ and reiterate that ‘there will be winners and losers’, but the overall message is that AI-driven job loss will be limited, and remediable by skills upgrading: capitalism should keep calm and carry on.

88

inhuman power

The argument, raging for several years, has become wellchoreographed. In 2018, the BBC reported on a study conducted for the City of London that suggested a third of jobs in the UK capital would be performed by machines within the next 20 years; wholesale and retail sales, transportation, storage, and food industries, employing a million people, would be hard hit (Kulka and Brown 2018). The Bank of England’s chief economist, Andy Haldane, interviewed on the topic, declared that the ‘Fourth Industrial Revolution’ would be of a ‘much greater scale’ than changes wrought by steam power, electricity and early computerization. Though these had ‘a wrenching and lengthy impact ... leaving many out of work for long periods struggling to make a living’, such effects would be yet more severe ‘when we have machines both thinking and doing – replacing both the cognitive and the technical skills of humans’ (Morrison 2018). Immediately, Simon Jenkins (2017), a reporter with The Guardian, responded with an article: ‘Worrying About Robots Stealing Our Jobs? How Silly’. Noting a chronic labour shortage in British social services such as health and education, Jenkins declared fear of AI mass unemployment absurd. New technology would mechanize some activities, but, ‘as throughout history’, labour markets would shift and ‘people – or their offspring – retrain’. ‘Short term disruption’ was ‘the reality of economic history’, yet ‘we survived and prospered as new needs, and jobs, emerged’. Another industrial revolution would release time for work in ‘service industries [that] essentially involve human relationships’ and ‘cannot be done by robots or machines’. Economists would, Jenkins concluded, do better to ‘welcome AI as releasing workers into the experience economy’.

Since the two contending positions were first outlined there have been a spate of predictive studies. Some are produced, commissioned or published by elite advisory organs of capital, such as the Organization for Economic Cooperation and Development (OECD 2016; Arntz, Gregory and Zierahn 2016; Nedelkoska and Quintini 2018) or the World Economic Forum (WEF 2016, 2017); others by private business consultancies with an interest in AI-related opportunities, like the McKinsey Institute (Manyika et al. 2017), PricewaterhouseCoopers (PwC 2017; Rao and Verweij 2018; Hawksworth, Berriman and Goel 2018) and Gartner (2017). Others come from legal associations dealing in labour law (Wisskirchen et al. 2017); the International Labour Organization (Chang and Huynh 2016) and anti-poverty organizations (Boston Consulting Group–Sutton Trust 2017). There have also been many independent academic studies.¹ This outpouring has in some respects refined analysis and in others rendered it more confusing.

automating the social factory

89

In general, studies that focus on AI replacing entire jobs – such as that of Frey and Osborne (2013) – have been succeeded by more fine-grained analyses of tasks within any one job that are susceptible to automation. This allows for possibilities of AI complementing or intensifying human labour rather than entirely substituting for it. In the preferred business jargon: ‘augmenting’ rather than ‘automating’. Subsequent OECD studies have dialled back the percentage of employment at imminent risk of automation from the apocalyptic to the merely alarming: 14 per cent for all its member countries, equivalent to over 66 million jobs, with another 33 per cent ‘deeply restructured’ by automation. In the US a mere 9 per cent or some 13 million jobs are tagged as AI-vulnerable, although the authors remark ‘this would amount to several times the disruption in local economies caused by the 1950s decline of the car industry in Detroit’ (Nedelkoska and Quintini 2018). The picture is further revised when various estimates of employment opportunities created by AI are reckoned in. A few have gone so far as to promise overall employment increases consequent on technologies of the fourth industrial revolution. Yet they often still admit the possibility of a net job loss (Arntz, Gregory and Zierahn 2016; Economist 2016a; Nedelkoska and Quintini 2018; Acemoglu and Resteropo 2018).

There have also been reversals over the types of jobs most liable to replacement. One of the main claims of AI Apocalyptics (and, one might cynically suggest, a reason for the prominence given to the issue) is that not only manual but also mental ‘middle-class’ jobs are threatened. More recent studies nevertheless stress that it is routinized work most immediately at risk. It is generally agreed that transport drivers, administrative assistants, cashiers, counter and rental clerks, telemarketers and accountants are the most endangered, and of those middle-class jobs, the most likely to be automated are ‘paraprofessional’ jobs, such as ‘paralegals, payroll managers, and bookkeepers’ and ‘semi-administrative jobs’ that provided a way into professional industries for those without advanced educational qualifications (Vincent 2017).

The geographical distribution of AI-induced automation is similarly contested. Many studies assume that developed economies are likely to be hit hardest, as ‘higher average wages incentivizes automation’ (Manyika et al. 2017). Others suggest AI may have the most dramatic effects in less advanced areas of the planet factory, precisely because there are more routine industrial jobs there (Kinder 2018). A study by the International Labour Organization (Chang and Huynh 2016) estimates that more than 137 million workers in Cambodia, Indonesia, the Philippines, Thailand and Vietnam are at high risk of replacement by machines such as ‘sewbots’ that combine automated sewing with computer vision (Larson 2018). Such ‘premature deindustrialization’ (Rodrik 2015) jeopardizes models of capitalist development, threatening to ‘slash millions of jobs and create an upswing in trafficking and slavery across south-east Asia’ in a ‘race to the bottom’ between the costs of machines and of humans (Kelly 2018). The geographical distribution of job losses in the centres of advanced capitalism will also be uneven. A study by Frank et al. (2018), focusing on the US, argues that job losses to AI automation will affect small cities worst because large cities have a disproportionate number of occupations that are of a cognitive or analytical nature at less risk of automation, while small city jobs are disproportionately routine, such as clerical work in retail and food service.

90

inhuman power

The great speculative AI and jobs debate has been coloured by changing economic conditions. AI Apocalyptics had a major impact in the high unemployment aftermath of the 2008 crash. When, a decade later, recovery was declared and employment levels restored (at least in the US), Business-as-Usual pundits regained their confidence. The tone of most recent managerial reports and policy advisories on the traumatic effects of the fourth industrial revolution is neither ‘denial’ nor ‘doomsday’ but ‘daunting’, with discussion of major Marshall Plan-sized projects for retraining workers. Great uncertainty thus continues to dog the AI and jobs debate. Behind the euphemisms and promises, however, there is visible the likelihood of a world of proletarian woes, with large sectoral job losses and chronic insecurity. And even in the most optimistic estimates of new AI-related job creation, there remain other important questions about its consequences for class relations, to which we now turn.

PANOPTICS, PRECARITY, POLARIZATION AND PROLETARIAN SCHOOLING

As Valerio de Stefano (2018) has pointed out, discussions about AI-related job loss and creation focus on quantitative rather than qualitative employment issues. While the size of the reserve army of the employed affects class decomposition and recomposition, so too do management techniques, the space and temporalities of labour, divisions within the workforce, and the forms of training and education through which that workforce must reproduce itself. As one commentator remarks, it might be true that ‘robots will create more jobs, ... but what if these jobs are less good and less well paid than the jobs that automation kills off?’ (Elliott 2017). We consider four points: panoptics, precarity, polarization and proletarian schooling.

automating the social factory

91