人工智能和资本主义的未来

The cover of an Economist (2018) report on ‘Artificial Intelligence in the Workplace’ depicts a Brobdingnagian office lamp illuminating Lilliputian workers, under the title ‘AI-spy’. Digital technologies have brought a ‘new Taylorism’ (Salame 2018) to workplaces, in which job deskilling is fused with algorithmic management and accompanied by workplace surveillance capacities far greater than those of any clipboard carrying superintendent. Phoebe Moore and her colleagues (2018) detail how sociometric badges, keyboard counters, email scanning, location tracking, motion sensors, and voice and facial recognition technologies detecting shifts in efficiency and mood minute to minute are now normal in workplaces. ML-driven ‘People Analytics’ is the new obsession of corporate HR departments that draw on AI tools for employee screening and selection based on social media traces, and delegate interviews to chatbots (Buranyi 2018). Fed with such data, AI becomes a way of assessing workers’ ‘productivity and fitness to execute particular tasks (De Stefano 2018: 7). This creation of a digitally ‘quantified worker’ is advanced under the guise of providing employees with beneficial information about their performance, health and state of mind for the purposes of voluntary self-improvement and ‘wellness’; the corporate providers of wearable workplace devices market them as ‘humanizing’ technologies. But this barely masks their obvious enhancement of managerial power by such technologies of ‘limitless worker surveillance’ (Ajunwa, Crawford and Schultz 2017).

Primarily concerned with the stress, indignity and intrusiveness such surveillance inflicts on employees, Moore et al. (2018) are sceptical about ‘deep automation’ and emphasize the economic, technological and social barriers to its abrupt arrival. But such monitoring is a moment within AI development. By the same logic that AlphaGo’s defeat of grandmasters at the ancient game of Go depended on digital digestion of thousands of hours of previous human play, so recording worker performance, even if initiated for disciplinary reasons, opens the way to new transformations of labour into machine form. Autonomous vehicles depend on digitally recording hundreds of thousands of hours of human driving. The recordings of call centre staff become the basis for voice analytic software that first supplements their work and then supplants it. The robotization of Amazon fulfilment centres emerged from electronic tagging of human ‘ambots’: in 2018 the company had two patents approved for a wristband tracker emitting ultrasonic sound pulses and radio transmissions to monitor a picker’s hands in relation to inventory, providing ‘haptic feedback’ to ‘nudge’ the worker towards the correct object (Yeginsu 2018). Reports of these patents sparked a public outcry about worker privacy and speed-up, but it is also clear that the wristband, or even more extensive monitoring devices, could pave the way for the robotic replacement of pickers. The organic link between workplace monitoring and worker replacement, and hence between AI development and surveillance, led The Economist (2018) to conclude that while AI workplaces may not be immediately depopulated, they will become ‘creepier’.

92

inhuman power

Precarious employment is one of the most discussed features of work in twenty-first-century capitalism. Definitions of ‘precarity’ vary, but often include part-time, temporary and self-employment, widely thought to be on the rise since the 2008 crash (Standing 2011). Particular attention has been given to the growth of a ‘gig economy’, organized by online platforms such as Uber, Lyft, TaskRabbit or UpWork, Clickwork and Mechanical Turk by workers designated as independent contractors, hence working without pensions, benefits and other regulatory protections, and often for low wages (Hill 2015; Kessler 2018). Some analysts of labour politics argue that the scale of ‘gigging’ is exaggerated (Henwood 2018b; Moody 2018b). In this view the real problem for workers today is not so much precarity, but rather stagnating low wages, rising costs and poor conditions in the permanent jobs that still account for the largest part of employment. These two perspectives are not entirely contradictory, as workers may supplement low-wage permanent jobs with precarious employment.

There is an affinity between AI and precarious labour. As we have seen, AI depends on online microwork. In addition, AI is important for the algorithmic management used by platform capitalism to organize contingent workers, scaling their numbers, earnings and schedules to demand, and thus ‘contributing to a casualization of work patterns and job and income instability’ (De Stefano 2018). Uber is a company self-described as having ‘machine learning in its DNA’ (Reese 2016a).

automating the social factory

93

When it was rapidly expanding in 2014 and 2015, Uber was adding up to 50,000 drivers every month, a number impossible for human managers to schedule and supervise, ‘so some system of virtual management had to be put in place’ (Caddell 2017). The company filled the gap with a variety of automated management tools providing a ‘rating system, performance targets and policies, algorithmic surge pricing [and] insistent messaging and behavioral nudges’, that ‘steer drivers to work at particular places and at particular times’ – a system that, as Alex Rosenblat (2016) remarks ‘complicate[s] claims that drivers are independent workers whose employment opportunities are made possible through a neutral, intermediary software platform’. From 2016 Uber consolidated these algorithmic management tools within its own internally produced ‘Michelangelo’ ML system, which was ‘designed to cover the end-to-end ML workflow: manage data, train, evaluate, and deploy models, make predictions, and monitor predictions’ (Hermann and Del Baso 2017). The logical culmination of Michelangelo’s trajectory seems to await the fulfilment of Uber’s ambitions to supply the software for self-driving cars, in which case the data culled from hundreds of thousands of precarious self-employed drivers would be absorbed by AI-directed autonomous vehicles.

A third current tendency AI is likely to amplify is the polarization between ‘high-end’ and ‘low-end’ jobs. This trajectory, apparent in the US since the 1950s, has intensified as digital automation and outsourcing hollowed out ‘middle-level routine’ jobs in favour of either high-paying cognitive labours or low-paying manual work, albeit with downward pressures on both (Autor and Dor 2013; Beaudry, Green and Sand 2013; OECD 2016; Elliott 2017). The US Department of Labour’s (2015) occupational projections to 2024 lists ‘computer and mathematical occupations’ as one of the fastest growing, and best paid, areas of employment. AI-capital will (at least until its self-automating capacities kick in) contribute to this growth, as it requires not just direct programming of ML, but also coding of the social media, apps and sensors that provide the Big Data on which it depends. The US national average salary for IT jobs is about \$81,000 (more than double the national average for all jobs), and the field is ‘set to expand by 12 percent from 2014 to 2024, faster than most other occupations’ (US Department of Labour 2015). But coding labour is dwarfed in absolute numbers and rate of growth by jobs in areas such as personal care work, nursing and medical assistance, retail sales and food service, many with wages in the \$20,000 a year range (US Department of Labour 2015; Henwood 2015). This suggests a future labour force divided between programmers coding other mid-level jobs out of existence, leaving tasks such as the care of an aging population to workers whose jobs remain at once too cheap and too humanly complex to automate – though this does not inhibit rampant fantasies about the imminent possibility of sex-bots and robot teachers. Such dreams and nightmares may be finding realization in the elder-care facilities of Japan, where the demographic problems of an aging population are particularly acute, and automata find relatively easier cultural acceptance. The development of ‘social robots’ as potential aides to caregivers has been a focus of government-supported industrial research (Kyung-Hoon 2018), and the Japanese public is exposed to the constant promotion of ‘exoskeletons, internet-of-things gizmos, humanoid or animal-shaped robots ... destined for a career of endlessly cheerful toil in the country’s growing network of nursing homes’ (Lewis 2017). Elsewhere, such trajectories will probably be slow-moving, at least while wages for personal care workers remain low.

94

inhuman power

This polarization has gendered and racialized dimensions. Early waves of digital automation hit men particularly hard because of their effects on manufacturing employment: the ‘information revolution’ is often cited as a contributory cause to the gradual advance of women in many fields of work. Many of the jobs likely to be AI automated are, however, typically held by women in administrative, secretarial, sales and customer service work, while software engineering is masculinized (Howcroft and Rubery 2018). In North America, the number of women studying computer science is static or falling, while in the UK only one in five of computing students is female. It seems likely that ‘men will disproportionately benefit from the greater demand for STEM skills’ (Boston Consulting Group– Sutton Trust 2017: 21–2). Silicon Valley companies have demonstrated an incapacity to move beyond a ‘brogramming’ culture, and also display ethnic segregation: a report from the US Equal Employment Opportunity Commission found that just 8 per cent of tech sector jobs were held by Hispanics, 7.4 per cent by African Americans, and 36 per cent by women (2016). Such patterns may be intensified by algorithmic hiring tools. In 2018 it was reported that an Amazon AI recruiting technology trained predominantly on men’s résumés had unsurprisingly developed an intrinsic bias against women, having ‘effectively taught itself that male candidates were preferable’ (BBC 2018b). That the system was discarded reflects a growing awareness of and resistance to high-tech sexism, but it is not an isolated example; analysis of automated hiring processes by both LinkedIn (Reese 2016b) and Google (Gibbs 2015) suggest similar sexist biases. At the other end of the spectrum, many of the fastest growing low-wage service jobs are ones in which women and minorities are strongly represented. AI-capital may therefore mark a new round of gendered and racialized work segregation.

automating the social factory

95

The route to survival in the face of AI automation is universally held to be education. Promises of ‘reskilling’ for new AI-related work can be glibly made at the macro-level of the entire workforce, but the micro-fate of individual workers will not be so easy. A capitalist nostrum is that the technologically unemployed should upgrade their programming skills and computer literacies. In the recessionary, post-2008 crash context of abundant low-cost labour, corporations widely abandoned training: it was cheaper just to hire someone new with the relevant skills. In such a labour market, workers and the unemployed have to conduct costly upgrading efforts at the very time when neoliberal policies have made access to education sharply stratified. As Dan Shewan (2017) writes of the US:

Private schools such as Carnegie Mellon University ... may be able to offer state-of-the-art robotics laboratories to students, but the same cannot be said for community colleges and vocational schools that offer the kind of training programs that workers displaced by robots would be forced to rely upon. In light of staggering student debt and an increasingly precarious job market, many young people are reconsidering their options. To most workers in their 40s and 50s, the idea of taking on tens of thousands of dollars of debt to attend a traditional four-year degree program at a private university is unthinkable.

Notional remedies for this problem include ‘massive open online courses, career-oriented nanodegrees ... that provide industry-recognized credentials, and coding boot camps that teach career-ready IT skills in a few months’ (Kinder 2018). At present, however, ‘these programs are used predominantly by those who are already highly educated or digitally savvy and looking to further enhance their employability by mastering cutting-edge technology. Few workers at the lower end of the labor market are taking advantage of such programs’ (Kinder 2018). These problems will be intensified because in an AI-suffused context ‘the speed of technological change will require these skills to be acquired rapidly, but they will also become obsolete faster with a shorter “half-life”’, so that ‘the need to continuously re-skill and up-skill oneself will raise the cost and time required for education and individual development’ (Boston Consulting Group–Sutton Trust 2017). Regardless of the net loss of jobs associated with AI, the concatenation of problems it brings around issues of precarity, labour market polarization and job restructuring threaten greater difficulties in workplace organizing, intensified divisions between the new elites of high-technology labour and low-wage or unemployed workers, and for most people compounding stresses and costs around education and training. It will also more widely affect how people think, communicate and interact around political issues, a point we address in the following section.

96

inhuman power

GRAMMATIZING THE NETWORKS

It is intrinsic to the concept of the social factory that, though composition and decomposition of the working class may start in the workplace, it extends beyond it (Notes from Below 2018a). Within the factory gates, the working class was decomposed by Taylorism and by increasingly advanced automation technology. This process of deskilling continues with AI, but increasingly extends into the spheres of circulation, consumption and social reproduction. To explain the effects of the AI-automated social factory we borrow the concept of ‘grammatization’ from Bernard Stiegler (2010, 2015, 2018), precisely because he uses it to refer to a deskilling that affects workers’ knowledge not just of how to work but also of how to live and think, a process he relates directly to AI.

For Stiegler, grammatization is ‘the process through which the flows and continuities which weave our existence are discretized’ (2010: 31). This concept refers to the development of technologies that record, separate and codify aspects of human activity. The translation of speech into alphabetic writing, with its grammatical rules, is an early and preeminent example of this process (hence ‘grammatization’). But grammatization, in Stiegler’s usage, also involves sense organs (vision and hearing), the movements and gestures of the body (at work, at home or while consuming), and entire patterns of social interaction (Stiegler 2010: 10; 2018: 8). The deskilling of workers by the transfer of their knowledge to machines in a way that subordinates them to capital is but an initial moment of a widening process of the grammatization of proletarians (which we gestured to in the previous chapter). This process later spreads to the automation of consumption – people are trained as consumers by the audio-visual technology of the culture and advertising industries – and then to human self-understanding, a knowledge that is automated by, for example, social media and big data analytics.

automating the social factory

97

Grammatization, we argue, should be understood as an extension of the real subsumption of labour to what Marx in the 1844 Manuscripts termed ‘life-activity’ (2007: 75). What ML subsumes is not just labour but more general human behaviour, communication, knowledge and skills that relate to how we live – including purchasing and consumption, but also the social relations of friendship, conviviality and love that are, from capital’s point of view, engines to be harnessed for its accelerated circulation. As Stiegler (2010: 45) puts it, using three useful French phrases, savoir-faire (knowing how to make and do), savoir-vivre (knowing how to live) and savoir-être (knowing how to be) are stripped from humans across the entire span of the social or planetary factory. In a recent work he explicitly relates this to AI and big data, suggesting that in their current forms these create disaffected, confused and incapacitated subjectivities, in a profound sense ‘stupefied’ (Stiegler 2015). For Stiegler, machine learning ‘calculate[s] correlations … in order to automatically anticipate individual and collective behaviour, which they also provoke and “auto-realize” by short-cutting and bypassing any deliberation’ (Stiegler 2017: 231). In our terms, they create an intensified state of decomposition.

It is through social media that capital is exercising some of its most powerful AI interventions, which at once restructure everyday social relations and have major implications for class composition. Google’s search-ranking algorithm and Facebook’s social graph of networked personal relationships have for over a decade been transforming the activities of these platforms’ users into digital profiles for targeting content and sale to advertisers. Where this crosses the boundary into the realm of AI is debatable; algorithmic processing is subject to the AI effect. Nevertheless, across two significant frontiers – the capacity to improve their own performance, and the escalating degree of predictive confidence that results – both corporations claim to be taking giant strides.

Facebook in 2016 unveiled its ‘self-improving, artificial intelligencepowered prediction engine’, dubbed ‘FBLearner Flow (FBL)’ (Biddle 2018). This uses ‘machine learning expertise’ to draw on ‘location, device information, Wi-Fi network details, video usage, affinities, and details of friendships, including how similar a user is to their friends’, to address corporate ‘core business challenges’. FBL is not merely offering advertisers the ability to target people based on demographics and consumer preferences, it also provides a ‘loyalty prediction’ service that searches a user base of over 2 billion individuals to detect millions ‘at risk’ of jumping ship from one brand to a competitor, subjects primed to be ‘targeted aggressively with advertising that could pre-empt and change their decision entirely’.

98

inhuman power

Another example of capital’s AI-powered grammatization of personal communication is Google’s Federated Learning project. This was announced in 2017 as a ‘decentralized’ edge AI, whereby ‘machine learning models are trained directly on smartphones of users’. Google could now delegate AI training to Android phones via an app, which would read users’ files, improve the personalization of their digital profiles, and relay this information to advertising clients, all the while leaving the original data ‘intact’ on the phone rather than lodging it on Google’s servers. Although Google chose to feature the user privacy benefits of this approach, Federated Learning brought the corporation a number of other, more self-interested, advantages. ‘Decentralized’ AI that functions ‘locally’ on users’ devices with no dependence on a network connection, ‘means less power consumption, ... minimal latency and faster machine learning processing’, and actually gives access to ‘more user data’ than could practically be harvested to a remote server in the cloud (Kulian 2017, emphasis added).

In the wake of the 2008 financial crash, it seemed that the growing reach of corporate-owned social media might actually expand possibilities for social dissent. The wave of occupations, assemblies and riots that burst out in 2011 were widely dubbed ‘Facebook (or Twitter, or YouTube) revolutions’. Although this appellation has been widely and properly criticized for its absurd techno-determinism (e.g. Tawil-Souri 2012), it remains true that a feature of these uprisings was the extensive use of social media platforms and mobile phone networks by protestors (albeit alongside good old fashioned on-the-ground organizing and face-to-face communication). This contributed to the rapid mobilization and contagious spread of protests, though probably also to their later problems of political coherence and short staying power (Dyer-Witheford 2015).

In the aftermath of these unrests, however, two AI-related developments posed major problems for networked anti-capitalist movements.

automating the social factory

99

The first was the discovery of the scope of nation states’ digital surveillance of populations, and in particular Edward Snowden’s disclosures about the US National Security Agency (NSA). These revealed that surveillance programmes set up in the wake of 9/11 piggy-backed on the algorithmic data collection of US social media and search engine corporations, all of whom collaborated with the NSA’s PRISM programme. What Snowden also showed was that the NSA processed the information gathered from PRISM and other projects by means of its own advanced AI programmes, collecting data and metadata, storing it on NSA cloud servers, and applying ML to identify suspect activities and associations (Gallagher 2013a, 2013b; Grothoth and Porup 2016). While the ostensible target of such AI-driven surveillance was terrorism, its technologies could unquestionably be applied to other purposes. Protestors already subject to ‘cyber-crackdowns’ via CCTV, mobile phone and social media surveillance were now potentially subject to a new extreme of scrutiny, if not necessarily from the NSA itself, then by the downward creep of AI-driven surveillance from elite intelligence agencies such as the NSA and GCHQ to police forces, private security firms and corporate labour spies (Dyer-Witheford and Matviyenko 2019).

The second pulse of AI-driven decomposition into internet politics can be summed up in the name ‘Cambridge Analytica’, if we take this as a shorthand for the convergence of alt-right networks, electoral digital strategy and Russian cyber-war interventions around Donald Trump’s election campaign. The contribution of AI to this scandal, which is still ramifying, has several levels. First, while the details of the voter profiling and targeting programs used by billionaire Robert Mercer’s company to assist Trump remain opaque, his background as an AI pioneer make it a virtual certainty these involved ML. Second, the Trump campaign relied on Facebook, an ML pioneer, both for the data obtained by Cambridge Analytica, and for the overt assistance Zuckerberg’s company provided Trump’s campaign in tailoring its digital advertising. Third, Trump’s electoral strategists, his alt-right supporters and friendly Russian operatives all relied for the circulation of their messages on social media algorithms promoting provocative, attention-grabbing, and hence advertising-lucrative content. By 2016, these algorithms were engineered with the assistance of ML. It is not for nothing that Robert Mercer’s daughter, Rebekah, devised a parlour game based on Trump’s campaign called ‘The Machine Learning President’ (Mayer 2018).

100

inhuman power

Trump’s election was an exemplary instance of decompositionary class politics: it aimed precisely at dividing people along lines of race and gender – with special attention to rust-belt populations damaged by globalization – in order to get them to vote for a capitalist plutocrat. Trump is, and will be, far from the only political figure to deploy ML technologies; Obama was, after all, ‘the big data President’ (Scola 2013). His campaign, his supporters and similar ventures such as the UK Brexit campaign have demonstrated how deeply AI-powered interventions can corrupt network communication. They have also shown the affinity between right-wing populism, or, more properly, neo-fascism, and digital exacerbations of the divisions within working classes already fragmented by decades of neoliberal class warfare, and on the verge of fresh AI-driven fractures. At the very moment when recompositionary initiatives around the social factory seem most important for anti-capitalist politics, these have been rendered far more difficult by the attention-shattering impact of algorithmic advertising, the chilling effect of ML-informed mass surveillance, and inflammatory fake news, toxic chatbots, cyber-warfare and other forms of ‘weaponized AI propaganda’ (Anderson and Horvath 2017). AI thus contributes to the transformation of the internet from a potential arena for the ‘circulation of struggles’ (Dyer-Witheford 1999) to one dominated by the circulation of commodities, the surveillance of resistances and the destruction of class solidarities.

A HEPTAGON OF STRUGGLES

Nonetheless, popular apprehension about the scale and direction of capitalism’s AI ambitions is growing. Recent years have seen resistances and protests relating to AI springing up around the social factory. What we emphasize is that these are eruptions of popular sentiment, not for more AI (‘full automation now’), but for its refusal. None are outright anti-AI struggles, but each rejects or contests specific aspects of AI. In their combination and overlap, and despite frequent disconnections and contradictions, they challenge the current trajectory of AI-capital. We list seven of these antagonisms – a heptagon of struggles.

1. Strikes and other workplace actions against AI-driven wage-depression, speed-up, monitoring, precarity and algorithmic management. The preeminent example is the wave of strikes and walkouts across Amazon fulfilment centres in Germany, Italy, Spain, Poland and the UK, demanding wage increases, job security and improvements in worker safety (Boewe and Schulten 2017; Cillo and Pradella 2017; Amazon Workers and Supporters 2018). The October 2018 pledge by Jeff Bezos to raise Amazon workers’ minimum wage (though undermined by claw-backs to a variety of other payments) was a partial victory for this movement, a concession intended to head off any transatlantic migration of struggles from Europe to the US (Reese and Struana 2018). Other strikes have disrupted the algorithmic management of precarious labour. The food courier Deliveroo operates with a ML-driven dispatch engine, which, the company claims, is called ‘Frank’ and is ‘constantly calculating and recalculating the best combination of riders to orders ... rider travel time, food preparation time etc.’ (Pudwell 2017). From 2016 onwards, strikes by Deliveroo drivers against the low pay and high insecurity enabled by ‘Frank’ spread virally from the UK across Europe, reaching as far as Hong Kong (Cant 2017; Zamponi 2018). A related stream of struggles is the rolling and complex series of lega challenges mounted from London to California by drivers for Uber against their algorithmically managed non-worker status (Rosenblat 2018). These have unfolded both in tandem and in contradiction with the legal objections, street blockades and protest suicides of ‘regular’ taxi drivers fighting the undercutting of their work. Intense as these tumults are, they pale in comparison with the scale of driver strikes, protests, algorithm sabotage and system-gaming raging across the vast operations of Didi, China’s Uber equivalent, insurgencies that may prefigure the chaotic future of labour activism under AI-capital (Chen 2018). To the file of worker actions should be added the unexpected North American unionization wave in new digital journalism enterprises such as Gawker, Salon, Huffington Post, Vice Canada and Al Jazeera involving young workers under ‘immense pressure to generate 24/7 online and social media content’ whose earnings and labour conditions are directly shaped by AI-driven advertising and directed news feeds (Cohen and de Peuter 2018: 115). More diffuse forms of worker organization are to be found in the microwork platforms that, as we have seen, are part of the AI production process. Turker Nation, the online forum of labourers on Amazon’s Mechanical Turk, has for years fostered microworker mutua support and knowledge sharing, challenged egregious exploitation by the platform’s corporate clients and protested the policies of Amazon itself (Harris 2014; Katz 2017). The more ML and other forms of AI suffuse capital, the more all worker resistance to corporate power will involve de facto resistance to the algorithmic forms in which that power is instantiated.

automating the social factory

101