Pieter Verdegem – AI for Everyone?: Critical Perspectives

~ Sample quotes ~

CHAPTER 5 – Post-Humanism, Mutual Aid – Dan McQuillan

...particular technologies appear to have an inherent compatibility with particular socio-political systems (Winner 2020), so it’s fair to ask what feedback loops connect AI and the societies into which it has emerged.

~

When we deal with social classification we can’t escape questions of power. The distribution of power in society may be complex and multivalent but it is also highly asymmetric. This not a problem created by machine learning, of course, but machine learning was produced within these structures of power and it is acting back on them. While the mathematics of AI may be expressed as matrices, it is a human activity that is inescapably immersed in history and culture. AI acts as an activation function for specific social tendencies. As an idea, or ideology, AI seeks to escape association with these worldly concerns by identifying with a pure abstraction and a neoplatonic purity of forms (McQuillan 2017), but this is only a plausible cover story to those who occupy an already privileged standpoint. The optimisation functions themselves will in general also be defined from these positions of social privilege. When AI talks in terms of ‘models’ it means the learned weights in a neural network, not the classic idea of a model that describes inner workings; as the goal is prediction not explanation, it is not supplying insights that could be used for causal interventions.

~

In more familiar political terms, it means acting as if solidarity were not only a stance but a core facet of being, as if mutual aid was not simply a choice made after social reality was sedimented out but a driving element in the iterative reproduction of the world. What we are currently experiencing instead is not an established order but the entropic disorder established by apparatuses that utterly lack the balance necessary to sustain us or our world.

~

People’s councils, based on solidarity and mutual aid, are an attempt to inoculate our meaning-making structures against fascism. The operations of fascism past and present show the ability to embrace technology as technique while replacing modernism with a cult of authoritarian traditionalism, a disturbing tendency already visible in the ‘dark enlightenment’ narratives of neoreaction circulating in Silicon Valley (Haider 2017). AI as we know it forms a harmonic with neoreaction’s ‘near-sociopathic lack of emotional attachment’ and ‘pure incentive-based functionalism’ (MacDougald 2015). People’s councils are a diffraction of AI, introducing the difference of care as a mode of interference and superposition.


CHAPTER 8 – The Social Reconfiguration of Artificial Intelligence: Utility and Feasibility – James Steinhoff

Open sourcing can generate a community around the software which entails skilled developers (and potential future employees) for the company who produces the software. It can also create a software ecosystem based on those tools, which a company can retain control over through a variety of mechanisms from mandatory lock-in agreements to closed source variants of programs. Google used (and uses) such strategies to make Android the most popular mobile operating system in the world (Amadeo 2018). While Google’s TensorFlow can currently be run on competing clouds, there are indications that the tech giants are aiming towards fully siloed AI ecosystems. Google is not alone in developing proprietary hardware specially designed for AI. Google’s Tensor Processing Unit (TPU) provides a ‘performance boost’ over traditional hardware, but ‘only if you use the right kind of machine-learning framework with it … Google’s own TensorFlow’ (Yegulalp 2017).

~

...Bernes discusses the non-modularity of certain technologies. By this he means technologies that ‘fit together into technical ensembles that exhibit a strong degree of path-dependency, meaning historical implementation strongly influences future development, precluding or making difficult many configurations we may find desirable’ (Bernes 2018, 334). He singles out energy infrastructure as particularly non-modular and argues that hopes of simply substituting clean energy sources, even if all political opposition were removed, is wishful thinking because the ‘technology [we] would inherit works with and only with fossil fuels’ (Bernes 2018, 334).

~

Data itself also raises several questions of feasibility. The first pertains to data collection. Many AI systems are trained on publicly available datasets in early stages of development, but usually, proprietary datasets are necessary to complete a project (Polovets 2015). The preparation and labelling of these is a labour-intensive and time-consuming process (Wu 2018). Creating a dataset also requires a venue for data collection in the first place. Companies such as Amazon and Google harvest reams of data from the interactions of users of their applications, even when they claim not to be, as smart home devices have shown (Fingas 2019). One business analysis of IBM suggests that because the company lacks a data collection venue, it will face difficulties developing its AI endeavours (Kisner, Wishnow and Ivannikov 2017, 19–20). This perhaps indicates why, in 2020, IBM entered into partnership with data-rich enterprise software company Salesforce. AI entails a capitalism built around surveillance, enabling ‘data extractivism’ (Zuboff 2019). How desirable is pervasive, multimodal surveillance for a socially reconfigured AI?

Machine learning’s reliance on data also necessitates a unique form of maintenance. A model which functioned well when it was deployed will no longer do so if the domain it is applied to changes such that the data it was trained on no longer accurately reflects that domain (Schmitz 2017). Imagine a hypothetical model trained to recognise traffic signs. If overnight the red octagons reading STOP were replaced with purple triangles reading HALT, the model would no longer function and would require maintenance. A social reconfiguration of AI will presumably be one component of a larger democratic restructuring of society with substantial changes to the normal routines of social life. A preview of this sort of disruption for AI has been provided by the COVID-19 pandemic, ‘models trained on normal human behavior are now finding that normal has changed, and some are no longer working as they should’ (Heaven 2020). When substantial shifts in human behaviour occur, models no longer map onto reality. It is reasonable to assume that models trained on data produced by life under capital may not function in a society striving to fundamentally change its basic axioms.

There is, however, at least one reason for optimism concerning data. A promising alternative to mass surveillance and siloed data ecosystems comes from the notion of data commons, in which individuals and institutions share data willingly, with controls over anonymity and a goal to make data valuable not only to tech companies, but also to its producers. The DECODE projects in Barcelona and Amsterdam have piloted aspects of a data commons successfully and are planning to scale up in the future (Bass and Old 2020). An interesting aspect of these projects is their use of other relatively new technologies, such as smart contracts (Alharby and Van Moorsel 2017), to aggregate and analyse sensitive data in ways which preserve privacy and retain user control. These projects provide a concrete demonstration of the feasibility of reconfiguring some aspects of data ecosystems. That they draw on novel smart contracts should remind us that assessments of feasibility are necessarily contextual; they are constrained by the knowledge of the assessor and the current technological milieu.


CHAPTER 9 – Creating the Technological Saviour: Discourses on AI in Europe and the Legitimation of Super Capitalism – Benedetta Brevini

Scholars in political economy of communication have shown how discourses around digital technologies have historically been constructed around modern myths (Mosco 2004) with major references to utopian worlds and possibilities. Myths, conceived as the dominant ideologies of our time (Barthes 1993) become powerful devices that normalise conventional wisdom into ‘common sense’ (Gramsci 1971), thus making the conception of alternatives virtually impossible. As a result, digital developments and policies are adopted without the benefit of an informed debate (Brevini 2020).

~

There are three crucial ways in which myths are used in the context of legitimising the status quo. Firstly, they are used as a weapon to control political debates. Secondly, they are used to depoliticise discourses that would otherwise show their contested political character. Thirdly, they are a crucial component of hegemonies, thus making it difficult for a counter-hegemonic discourse to arise.

~

In exploring public discourse shaping the popular imagination around possible AI futures, Goode (2018) observes that contemporary discourse is:
skewed heavily towards specific voices – predominantly male science fiction authors and techno-centric scientists, futurists and entrepreneurs – and the field of AI and robotics is all too easily presented as a kind of sublime spectacle of inevitability (…) that does little to offer lay citizens the sense that they can be actively involved in shaping its future. (Goode 2018, 204)

~

Furthermore, the latest study on media coverage of AI in the UK conducted by the Reuters Institute (Brennen, Howard and Nielsen 2018) showed that the UK media coverage of AI was overwhelmingly influenced by industry concerns, products and initiatives.

~

In pure enlightenment fashion, this absolute faith in technology, embraced and supported by cybertarians’ Silicon Valley circles (Dyer-Witheford 1999; Brevini 2020) turns into a powerful apology for the status quo and the current structure of capitalism, without any real space for critique.

~

... it is impossible not to recall Williams’ analysis in Towards 2000 where he stated that ‘The sense of some new technology as inevitable or unstoppable is a product of the overt and covert marketing of the relevant interests’ (Williams 1985, 133). In reality technological development is not predetermined, and alternative paths to a market-led development that reinforces the current neoliberal status quo are always a possibility (Brevini 2020).

~

The recurrent myths that are omnipresent in the European Framework for AI have two major consequences. Firstly, they structure a hegemonic discourse that makes it impossible to think of alternative paths, framing resistance as futile because technological development is predestined. Accordingly, they legitimise a neoliberal ideology that pushes consumerism and productivity above all values and strips technology from the social relations that are at the basis of technology development (Williams 1985; Brevini 2020). Secondly, they redirect public discourse, by obfuscating and inhibiting a serious debate on the structural foundations of AI, its progressively concentrated ownership and the materiality of its infrastructures. Taken together, these myths of AI, construct a type of discourse that frames the problem of AI in a way that excludes any emphasis on crucial questions of ownership, control and the public interest. It also diverts attention from known problems of inequality, discrimination and bias of data analysed algorithmically that lies at the heart AI systems (Brevini and Pasquale 2020). When these crucial questions are asked, they are only addressed through the ‘AI ethical’ framework that has little to say about the structural inequalities on which AI is built (Wagner 2018).

~

Major lobby groups go in to bat for their vested interests in the policy arena, armed with funded academic research on the benefits of AI and efficiency. For example, a report published in 2019 by the New Statesman revealed that in five years Google has spent millions of pounds funding research at British universities including the Oxford Internet Institute (Williams 2019), while DeepMind, Alphabet’s own AI company, has specifically supported studies on the ethics of AI and automated decision-making. Correspondingly, Facebook donated US $7.5m to the Technical University of Munich, to fund new AI ethics research centres. Another troubling case is the US-based National Science Foundation program for research into ‘Fairness in Artificial Intelligence’, co-funded by Amazon (Benkler 2019). As scholar Yochai Benkler explained, the digital giant has ‘the technical, the contractual, technical and organizational means to promote the projects that suit its goals’ (ibid. 2019). Hence, ‘Industry has mobilized to shape the science, morality and laws of Artificial Intelligence’ (ibid. 2019).


CHAPTER 12 – Algorithmic Logic in Digital Capitalism – Jernej A. Prodnik

It is essential to underscore that data is not simply one of the resources in what Srnicek (2017) calls platform capitalism or what Fuchs (2019) defines as Big Data capitalism. It has become the resource for major companies, especially in the case of machine learning (Coeckelbergh 2020). This is why datafication – and correspondingly Big Data and mass surveillance – is not simply an optional thing. If you block surveillance the effectiveness of algorithms plummets and many of the existing business models start to collapse. Surveillance and privacy breaches are therefore a necessary part of the algorithmic logic in digital capitalism. They are not a bug but a constituent feature that powers its development.
A continuous push for datafication also brings about a highly unequal concentration of the ownership of the data, which is syphoned off using digital surveillance (cf. Mosco 2014). These information inequalities are even more intensive than in the past, when Perelman (2002, 5) pointed out that ‘intellectual property rights have contributed to one of the most massive redistributions of wealth that has ever occurred’. He based this assessment on the fact they were owned almost exclusively by the rich and the powerful. Processes occurring with algorithmic datafication, however, are accentuating and intensifying this problem even further.


Edit
Pub: 11 Aug 2023 12:48 UTC
Views: 110