---
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: 110
range_count: 10
html_url: "https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02?start=110&limit=10"
markdown_url: "https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02.md?start=110&limit=10"
previous_url: "https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02.md?start=100&limit=10"
next_url: "https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02.md?start=120&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=110 | PRINTED_PAGE=102
102
inhuman power
## PDF_PHYSICAL_PAGE_INDEX=111 | PRINTED_PAGE=103
2) Protests against military and paramilitary AI applications. Over the summer of 2018, a remarkable wave of technology-worker resistance to military and militarized policing projects swept through Silicon Valley (Dyer-Witheford and Matviyenko 2019). At Google, workers organized to shut down Project Maven, a Pentagon project that uses ML to improve targeting for drone strikes, and succeeded. Following this, Google’s CEO, Sundar Pichai, published a statement of principles on AI development, saying the company would not work on AI weapon or surveillance contracts, although reserving the right to pursue other military projects. The company also withdrew its bid on a \$10 billion Pentagon contract, the Joint Enterprise Defense Infrastructure (JEDI) cloud computing project (BBC 2018a). Meanwhile, at Amazon, workers petitioned Bezos to stop selling the corporation’s Rekognition facial identification software to US police departments and the Immigration and Customs Enforcement (ICE) agency, notorious for its ‘zero-tolerance’ enforcement policies. At Microsoft, workers similarly demanded the termination of a \$19.4 million cloud computing contract with ICE (Frenkel 2018), and at Salesforce, workers tried to block the company’s involvement with Customs and Border Protection (CBP). Alongside these events, and belonging to issues of sexism and discrimination discussed in point five below, we can include the worldwide walkout of Google employees against the corporation’s handling of sexual harassment issues. Dissent against the security state is nothing new, but the Silicon Valley revolt was exceptional. It mobilized an elite, high-technology workforce centrally involved in the production of digital weaponry that had for decades been considered immune to serious politicization by virtue of its wealth, status and libertarian mind-set. Such events do not come out of nowhere. In part they indirectly reflect an increasing restlessness among high-tech workers whose wages, good as they are, have not kept pace with the obscene expanding wealth of their bosses. Behind the outbreak, however, lay organizing by groups such as the Tech Workers Coalition, which for years had worked with labour organizations to unionize contracted programming work and patiently campaigned in Silicon Valley and Seattle to develop a critique of corporate power, opposition to racism and sexism, and solidarity between programmers and low-wage support workers (Tech Workers Coalition 2018).
automating the social factory
103
## PDF_PHYSICAL_PAGE_INDEX=112 | PRINTED_PAGE=104
3) Anti-surveillance movements. Since Edward Snowden’s revelations, awareness of the scope of surveillance in capitalist liberal democracies has escalated (Gallagher 2018) (though of course such surveillance is no news to populations elsewhere – for example in China). This is an AI issue, not just because the NSA uses ML in its scans, but because surveillance is crucial in the development of AI. As we have described, large data sets are necessary to train ML systems. Not all of these data sets necessarily scrutinize human behaviour, and there are some technica attempts to mitigate the invasiveness of such observation (Knight 2016b; Marr 2018), but AI development is likely to remain significantly reliant on large-scale people watching. In North America, anti-surveillance self-defence manuals and information sheets have proliferated across social movements. ‘User’s guides for privacy and protest’ summarize practices of ‘obfuscation – the deliberate addition of ambiguous, confusing, or misleading information to interfere with surveillance and data collection’ (Brunton and Nissenbaum 2015). Facial recognition AI, and the possibility of baffling it with ‘adversarial images’, is a recent hot-spot. There are, however, serious limitations of time, expense and expertise on individualized anti-surveillance techno-measures. Politica and legal limits to surveillance have better chances of mass effect. In North America, some of the most important initiatives are from minorities that know only too well how profiling systems at once constitute and control suspect social groups (Dyer-Witheford and Matviyenko 2019): African Americans, opposing the digital tracking both of entire neighbourhoods and Black Lives Matter leaders; First Nations fighting the digita scrutiny of land and pipeline protests; and Muslim Arabs, opposing wholesale databank inventories of their existence. The major politica steps in surveillance restriction have, however, come from Europe, where memories of Nazism and Stalinism are still alive. Anti-corporate surveillance movements such as Max Schrem’s ‘None of Your Business’, and governing elites’ opposition to national sovereignty infringements give the issue real political weight. The European Union’s General Data Protection Regulation (GDPR), which came into force in 2018, limits the data gathering of digital capital, including the great AI developers. Its requirements for active user opt-in (rather than default acceptance) of digital tracking, corporate transparency about the use of personal data, users’ right to leave platforms without losing stored data, explanations of automated decisions, and for a ‘right to be forgotten’ all crimp the use of population data as extractive raw material for the construction of machine intelligence (Meyer 2018; Kaput 2018; Srnicek 2018). The negative response of ML developers and computer scientists, and their critical comparison of Europe’s regulations with the corporate freedoms of the United States and the powers of China’s surveillance state (K-F. Lee 2018), indicate how big a problem anti-surveillance movements pose to AI-capital, but also the challenges such movements face in an AI-competitive world-market.
104
inhuman power
## PDF_PHYSICAL_PAGE_INDEX=113 | PRINTED_PAGE=105
4) Social media defection. More diffuse and invisible than anti-surveillance activism is dissent by subtraction. In the wake of the Cambridge Analytica scandal, leaving social media paradoxically became a social media meme, driven by an amalgam of concerns over digital surveillance and manipulation, the psychological effects of techno-addictions, and revulsion at the extreme commodification of online environments. A ‘Delete Facebook’ campaign sprung up, then subsided. There is, however, evidence that people in the 18 to 24 age range are either curtailing or, occasionally, abandoning social media use, even while their parents become increasingly immersed (Kale 2018). In so far as ML is a big data undertaking, some AI-capital may be impeded by a slow exodus of data-subjects ‘logging off’ social media.
5) Algorithmic bias busting. We have seen the tendency of AIs to amplify pre-existing sexist and racist discrimination. Businesses owned and dominated by white men, and governance systems administered by the same, are producing AI systems that discriminate against women and minorities because they are trained on data sets that reflect the historical levels of hiring, wealth, clean records and career success enjoyed by white men (Noble 2018). Such discriminatory AI affects those who encounter it on the screens of police, border guards and welfare services with particular severity. That algorithmic bias has even been named is, however, evidence of struggles against it. Groups addressing the issue range from the Algorithmic Justice League, fighting ‘the coded gaze (Buolamwini 2018) at the Massachusetts Institute of Technology, to NGOs like the San Francisco-based, non-profit Human Rights Data Analysis Group (HRDAG), which revealed algorithmic bias in US predictive policing software, to community organizations such as Silicon Valley De-Bug, which is involved in immigrant and prisoner rights and economic justice in the San Jose area, or Philadelphia’s Media Mobilizing Project, which is organizing low-wage workers and poor communities (Wykstra 2018). As correcting algorithmic bias increases the accuracy of AI predictions, it can, in theory, be serviceable to capital to improve its niche targeting of commodities and hiring of talented labour-power, even if this requires acknowledging embarrassing and expensive displays of sexism, racism and homophobia: these, at least, are the terms progressive neoliberalism sets for such initiatives.2 However, in so far as algorithmic bias discriminates against people who are both poor and female, queer or racialized – hence the most vulnerable of the vulnerable – algorithmic justice movements hamper AI-capital’s disciplining and discarding of its so-called surplus populations, challenging the ‘matrix of domination’ that has informed its development (Costanza-Chock 2018).
automating the social factory
105
## PDF_PHYSICAL_PAGE_INDEX=114 | PRINTED_PAGE=106
6) Digital city disturbances. As giant AI developers impress an ever stronger footprint on the urban landscape with their headquarters, campuses and experimental sites, social conflicts explode in the city-incubators of AI. Nowhere is this more evident than around Silicon Valley, where billionaires in luxury mansions, millionaires working in high-tech campuses, low-paid service workers and a large homeless population coexist in constant tension. As software production expanded north from the Valley of San Francisco and the Bay area, so too did these social fault-lines. In 2013–14, the private buses that transport high-tech workers to the ‘Googleplex’ in Mountain View became a flashpoint for protests against gentrification, eviction, displacement and congestion (Goode and Miller 2013). Demonstrators blockaded the bus routes. Signs read ‘Gentrification and Eviction Technologies: Integrated Displacement and Cultural Erasure’ and ‘Fuck Off Google’. Leaflets accused Google of ‘building an unconscionable world of surveillance, control and automation’ (Streitfeld and Wollan 2014). The protests flared up, went out, and then re-appeared in 2018. ‘Techsploitation is Toxic’ militants blocked a dozen Google buses by piling in front of them the electric scooters used by an Uber-style company, and denounced plans to clear the streets of homeless encampments with slogans such as ‘Sweep tech not tents’ and ‘They call it “Disruption.” We call it displacement’ (Streitfeld 2018). These urban disturbances also include contestation of AI-based ‘smart cities’. In 2017, Sidewalks Labs, a subsidiary of Google’s holding company, Alphabet, negotiated with the city of Toronto a ‘public-private’ partnership for the development of the waterfront neighbourhood Quayside. Under this pact, the details of which were not disclosed, Quayside would become a showcase for AI-saturated urban innovations; streets designed for autonomous vehicles, robot garbage collection, ubiquitous sensors and security cameras, and a centralized information management system reading residents’ ID from library and health cards. Canada’s Prime Minister Justin Trudeau praised the project as a harbinger of clean, green urbanity. Eric Schmidt, Google’s former CEO, explained that the idea came from Google’s founders getting excited thinking of ‘all the things you could do if someone would just give us a city and put us in charge’ (Kofman 2018). Apparently unmentioned in the initial agreement were the ownership and terms of collection of digital data to be gathered in Quayside. This issue remained obscure until forced into the public eye by a campaign led by TechReset Canada (Barth 2018), a group opposed to unrestrained digital commodification that criticized the opacity of the deal, insisted the contract be made public, and demanded provisions to ensure that ‘the data and data infrastructure of this project are the property of the City of Toronto and its residents’. This mobilized a coalition whose participants ranged from housing activists to disgruntled city planners; their cause won support even from leading Canadian high-tech entrepreneurs, objecting to intellectual property surrender to foreign corporations. Prominent members of the board overseeing the Quayside agreement resigned. As of early 2019, the issue remains unresolved. Sidewalk Labs may concede to objections, but also dodge the protest, by agreeing to collect only aggregated, anonymized data – a ploy that, while it addresses the limited issue of personal privacy, leaves untouched the larger problems of corporate command over population-level information and urban design (Greenfield 2013).
106
inhuman power
## PDF_PHYSICAL_PAGE_INDEX=115 | PRINTED_PAGE=107
7) Anti-corporate techlash. Many of these concerns have converged in a nebulous ‘techlash’ against the power of the large digital corporations. Facebook, Cambridge Analytica and the Russian cyber-war scandals in the 2016 US election catalysed a revival of long-dormant discussions about concentration of ownership in digital oligopolies (Foer 2017; Mosco 2017). Even within a US high-technology community famously opposed to state regulation, voices expressed scepticism as to the capacities of Google and Facebook to police themselves and called for enforced transparency and even financial liability for the negative externalities of their algorithmic processes. Other criticisms went deeper – asking, for example, about the degree to which capital’s media corporations are deeply and basically dependent on the ‘fake news’ of advertising and on highly engineered attention manipulation. In Europe, litigation was launched against the giants of digital capitalism over violations of public regulation and labour law, tax evasion and anticompetitive abuse of monopolistic powers. Calls for the disaggregation and regulation of some of the largest AI-developing corporations appeared in the platforms of Bernie Sanders and Jeremy Corbyn (and even, erratically, in the ruminations of right-wing authoritarian populists such as Trump). Antitrust politics are by no means the same as anti-capitalism, but one effect of the ‘mainstreaming’ of such concerns has been to open up a space for discussion of more radical propositions for a ‘socialization of the data banks’ (Morozov 2015). These extend beyond the regulation or breakup of oligopolies to the establishment of ‘commons’ institutions separate from both state and market, and the formation of public computing utilities kept at arm’s length from governmental and corporate power. We return to these issues in the Conclusion. By 2018 even business commentators were speaking of changed attitudes towards digital capital as one of an array of factors, alongside gadget-saturated markets and international trade wars, that were giving high-tech a ‘hammering’ on that infallible index of human values, the stock market (Rajan 2018).
automating the social factory
107
## PDF_PHYSICAL_PAGE_INDEX=116 | PRINTED_PAGE=108
CONCLUSION: TOWARDS THE EVENT HORIZON
Some working-class recomposition is occurring in digital capital, with the powerful presence of AI as an invisible attractor around which various rebellions and dissents array themselves. We do not want to overstate this process. Anti-corporate techlash can be politically shallow; smart city opponents are still marginal; algorithmic justice both hinders and helps capital; defections from social media are slow and partial; anti-surveillance policies could be rolled back by competitive pressure; despite the success of the Project Maven revolt, other victories against militarized AI are scant; above all, over the uptick in strikes against AI-capital hangs the shadow that resistance may spur further automation. While we have suggested that, regardless of job-loss issues, AI-capital poses many other threats to the well-being of its populations, it is also the case that the future of struggles within it depends heavily on how deeply the fourth industrial revolution undermines the basis of wage-labour.
As we have seen, early predictions of an imminent AI-caused employment implosion have been significantly watered-down by more recent studies, and have lost their edge as the post-recession US economy returned to what is officially regarded as full employment. Yet some observers suggest even limited, sectoral job loss to automation could figure as largely in upcoming US elections as the jobs-gone-abroad issue did in 2016 (F. Levy 2018). Longer term, estimates of AI employment effects continue to vary wildly, and even some Business-as-Usual theorists have changed their models to admit the possibility that capital’s job-creating and job-destroying functions might go seriously out of sync (Acemoglu and Resteropo 2018). Some maintain quite catastrophic positions; the widely read treatise on AI by Kai-Fu Lee (2018), former head of Google China, insists on a loss of up to 40 per cent of US jobs by 2030, and similar numbers for China – in other words, numbers of world-historical turbulence. Such runaway AI, massively increasing the reserve army of the unemployed, would further weaken unions and other worker organizations, but the extremity of immiseration could generate new forms of proletarian revolt (even, perhaps, as bloody and determined as those of Westworld’s androids). In such speculations, we arrive at the event-horizon foreseen by the ‘Fragment on Machines’, where capital’s own machinic drive explodes it from within: terminally decomposing its working class, capital decomposes itself. It is the apparently irrefutable logic of this futurist proposition that perhaps gives leftists a certain last-ditch confidence, complacency or even enthusiasm as they face AI-capitalism. ‘Probably’, one says, ‘the tech won’t really work – and, anyway, if it does, it only perfects the system’s self-destruction.’ But how warranted is this assurance?
108
inhuman power
## PDF_PHYSICAL_PAGE_INDEX=117 | PRINTED_PAGE=109
automating the social factory
109
## PDF_PHYSICAL_PAGE_INDEX=118 | PRINTED_PAGE=3
Perfect Machines, Inhuman Labour
Narrow AI is a special case of capitalist machinery that challenges Marxist thought from numerous directions. This chapter, however, turns to artificial general intelligence (AGI), or AI with capacities for reasoning with general knowledge and doing a variety of tasks in diverse and unfamiliar domains (Gubrud 1997; Wang and Goertzel 2007: 5; Goertzel 2014: 2; Baum 2017: 3). AGIs are thus machines with a flexibility or generality of intelligence that is similar in scope, though not necessarily in functioning, to that of humans. As Pei Wang and Ben Goertzel explain, an AGI could be used ‘in situations where ready-made solutions are not available, due to the dynamic nature of the environment or the insufficiency of knowledge about the problem … what we expect from an AGI system are not optimal solutions … but flexibility [and] creativity’ (2007: 5).
These references to flexibility and creativity bear a striking resemblance to Marx’s concept of labour. Marx, however, presupposed that ‘labour … is an exclusively human characteristic’ (1990: 283–4) and that the capacity to labour exists only ‘in the physical form, the living personality, of a human being’ (1990: 270). Against this anthropocentrism, we argue that there is an isomorphism between the theoretical notion of AGI and Marx’s concept of labour and labour-power; AGI, therefore, profoundly challenges Marx’s labour theory of value; in particular the axioms that only human beings can labour and create value, and that machines categorically cannot. AGI is, in other words, such a special case of capitalist machinery that its status as machinery must be questioned. We argue that AGI could not only potentially labour, but under certain social conditions also create value. The potential ramifications of AGI for the capitalist mode of production and humankind are drastic. Were AGI to be created, many believe it would be the ‘technological singularity’, that is, an outcome of ‘exponential technological progress’ resulting in ‘such dramatic change that human affairs as we understand them today c[o]me to an end’ (Shanahan 2015: xv).1 While the further research and development of narrow AI will likely see the pool of people superfluous to capital accumulation grow, the development of AGI would go beyond this already miserable vision as discussed by Marx to an even worse one, that of Homo sapiens becoming a surplus species. AGI thus suggests the possibility of a capitalism without human beings.2
3
## PDF_PHYSICAL_PAGE_INDEX=119 | PRINTED_PAGE=111
To make this argument, we discuss why Marx assumed both labour and value creation to be inherently human, zeroing in on the sharp distinction he makes between humans and animals. We consider why he grounds this distinction in various capabilities like creativity and imagination, and how some AI projects, as of 2019, are trying to automate such capabilities. While the previous chapters have engaged in some necessary speculation, this chapter envisages a distant future that is more science fiction than science fact. Thus, before we delve into the specifics of Marx’s value theory and his arguments about human uniqueness, we first justify why we are devoting a chapter to a technology that has yet to, and might never, see the light of day.
ARTIFICIAL GENERAL INTELLIGENCE
No scientific or philosophical consensus exists on the definition of intelligence, but most perspectives agree that true intelligence must be to some degree flexible or general (Kaplan 2016; Lake et al. 2017: 9–19). Existing ML and GOFAI are distinguished from human intelligence by their narrowness and inflexibility. Research is, however, underway to produce AI which can be generally applied – AGI. The term AGI was first introduced by Mark Avrum Gubrud (1997) to describe ‘AI systems that rival or surpass the human brain in complexity and speed, that can acquire, manipulate and reason with general knowledge, and that are usable in essentially any phase of industrial or military operations where a human intelligence would otherwise be needed.’ The idea of AGI, however, actually dates back to the original goal of the 1956 Dartmouth workshop (Wang and Goertzel 2007: 6). But the goal of emulating human intelligence was quickly abandoned in favour of more manageable practical problems that could be solved by narrow AI, and until the early 2000s AGI was considered so unrealistic that people working on it were seen as crackpots (Pennachin and Goertzel 2007: 1; Wang and Goertzel 2007: 4). While there were research projects that were focused on the goal of building AGI, such as the Soar cognitive architecture that John Laird, Allen Newell and Paul Rosenbloom started in 1983 (Laird
perfect machines, inhuman labour
111
下一段：[继续读取](https://workshop.closeclaw.top/api/ai/read/a2eab101-adfc-4a0d-8142-63a4d11cee02.md?start=120&limit=10)