The hidden abode of artificial intelligence production: Stretching the limits of artificial intelligence ethics and critique

Introduction: What do we mean when we talk about the relationship between artificial intelligence and labour?

The literature on the relationship between artificial intelligence (AI) and labour is extensive, with a significant portion focusing on the ethical implications of AI's impact on labour markets, recruitment and management practices (Bakiner, 2023). In this sense, the discussion of this relationship usually follows a more general bifurcation of the current debate in AI ethics into two main areas: the ethics of AI research and the ethics of AI consumption. The former deals with the principles guiding the creation and design of AI systems, while the latter focuses on the ethical applications of existing AI technologies.

Luciano Floridi ¹ (2022), an influential voice in the contemporary debate, explicitly draws this distinction, highlighting the need to address both research and consumption ethics jointly. This is done in his discussion of the risks connected with digital ethical dumping and the problem of distinguishing between unethical treatments of practices and of the results of those practices. For instance, he writes, a facial recognition algorithm could be trained on personal data from countries with weak ethical and legal protections and then reimported into the European Union, where this practice would be illegal and not recognised as ethical based on the General Data Protection Regulation (GDPR); however, the resulting algorithm could still be imported and implemented without any issues (Floridi, 2022: 116–117). Furthermore, a consensus exists within the debate about the necessity to consider different ethical issues implied in AI development based on the same, shared set of principles, clearly even beyond such a distinction, although there is no consensus, instead, about what principles should actually be assumed (Hagendorff, 2020).

However, it is usually in terms of an ethics of consumption that the impact of AI on labour is discussed within the debate: in other words, the potential ethical implications of its possible uses as a product take centre stage. Such a tendency can be detected even in officially adopted sets of principles. In the Recommendations on the Ethics of Artificial Intelligence (United Nations Educational, Scientific and Cultural Organization (UNESCO), 2022: 10, 26, 36), for instance, the ethical issues arising from the relationship between AI and labour are primarily discussed as concerning the impact of AI products on the labour market (e.g. the problem of technological unemployment) and on the organisation of labour at large.

What is apparently lacking in this framework is the consideration of the AI production process in its full materiality; that is, taking into account the entire set of people whose labour makes the realisation of an AI product possible. In other words, between an ethics of research standing before the production process and an ethics of consumption standing after the production process, an ‘ethics of AI production’ is lacking. As a matter of fact, the production process of each AI product is long and complex, and implies the performance of various tasks made possible by the participation of many different workers variously skilled, paid, treated and situated. These people are affected by the impact of AI production processes well before being affected by the impact of AI products. Therefore, any ethical discourse will be incomplete – other than, at the very least, morally inconsistent and uneven – if the AI production process is not assumed as a field of application of AI ethics, on par with research and consumption.

The present article thus aims to discuss the possibility of including the sphere of AI production within the domain of AI ethics and investigate its moral implications. For this purpose, the first section considers the role of human labour in the AI production processes, with particular reference to the distinction between, on the one hand, high-skilled and high-paid jobs and, on the other hand, low-skilled and low-paid jobs, their differential distribution in the production process itself, and the labour conditions of ghost workers, in order to analyse the main ethical issues emerging within the field. In the second section, some aspects of the existing critical literature concerning AI and labour are discussed on the basis of the analysis conducted in the previous section, focusing on Marxist and decolonial scholarship and more precisely on its lack of consideration of the global value chain through which AI production processes are structured. Finally, the possibility and limits of an ethics of AI production are reconsidered by assuming the centrality of workers’ struggles and agency along AI's global value chain.

Human labour and labour conditions within AI production processes: Some ethical issues

In order to argue that the sphere of AI production should be included as such as an object of AI ethics, it is first of all necessary to enquire whether the production process at issue and the ethical problems it generates are specific to the AI industry. From this point of view, it would be unreasonable to claim that existing forms of poor labour, hyper-exploitation and toxic labour environments along the AI global value chain did not exist before AI production and have been directly and exclusively generated by it. Nonetheless, it is not necessary to assume such a strong and distinctive connection in order to prove the specificity of labour conditions in the AI industry and the specificity of the ethical issues it poses. As a matter of fact, other forms of poor labour, hyper-exploitation and toxic labour environments already existed, at the very least, within the whole digital economy; consider, for example, the work of content moderators for social networks (McIntyre et al., 2022; Newton, 2019; Perrigo, 2022, 2023b), the quality control for research engines (Alphabet Workers Union, 2023; Kerr, 2023), and more generally the wide world of ‘ghost workers’ (Ghost Work Project, 2023; Gray and Siddharth, 2019).

It is thus clear that a certain continuity exists between these sectors of the digital economy: the same applies, under many aspects, to many sectors where forms of algorithmic management have been implemented in the organisation of work (Ajunwa and Schlund, 2020; Tan et al., 2021). Nevertheless, far from being an argument against the inclusion of the sphere of production within the field of AI ethics, this is, if at all, a strong argument for extending ethical concerns regarding AI to other sectors of the digital economy. Furthermore, other industries do not exhibit the same commitment to ethically regulate themselves: the ethical dimension has rarely been so ubiquitous, from both inside and outside, in any industry before the AI case. As remarked by Timnit Gebru (2020: 262), ‘ethics has become the language du jour’ in this context. This peculiarity of the AI industry makes it a particularly adequate point of departure for ethical reflections connected with this field, provided that the industry cannot set the boundaries of its own ethical scrutiny; for example, excluding from its domain the AI production process. Otherwise, rather than AI ethics, we would simply have a particularly complex form of ‘bluewashing’ (Floridi, 2022: 111–113).

It is, however, possible to argue that a strong connection exists between the objective structure of the AI production process and the specific labour conditions existing along its global value chain. James Steinhoff (2021: 173–177) has identified three stages within the production process of an AI product: data processing, model building and deployment. He furthermore identifies five categories of workers within the industry: data scientists, data engineers, data analysts, service workers and ghost workers (Steinhoff, 2021: 151–155).

The role of low-skilled and low-paid workers – to whom the already-evoked ethical concerns about labour conditions primarily apply – is mainly concentrated in the first stage of the process. Far from making their role marginal or limited, it should be remembered that collecting, formatting, labelling and cleaning data consume about 80% of the total production time; this is true even for data scientists, who are instead among the most high-skilled and high-paid workers of the industry and consider these tasks as the least enjoyable part of their work (Crowdflower, 2016: 7; Steinhoff, 2021: 173). Among these tasks, the most mechanical and repetitive is considered to be data labelling, which is also why it is frequently outsourced to ghost workers, both through specialised companies – often in their non-western hubs (Gray and Siddarth, 2019; Perrigo, 2023a) – and platforms such as Amazon Mechanical Turk. Despite being very often invisibilised, ghost workers perform essential tasks – such as, as said, data labelling, content moderation and other forms of micro-work – supporting the production and functionality of AI and digital platforms. Also due to the fact that they are usually employed through third-party companies and platforms, not only do they usually receive low wages, but also suffer from scarce job security and high levels of precariousness (Gray and Siddarth, 2019).

Concerning the AI industry, ghost workers are fundamental both from a technical and an ethical point of view: on the one hand, as seen, data collecting and processing is the first stage in any AI production process, and without labelling and cleaning tasks it would be impossible for data to be used for the training of AI models. On the other hand, AI products devoid of filters impeding the generation of offensive and unethical content are very hard to sell, as the case of GPT-3, the predecessor of ChatGPT, well demonstrates. However, the existence of such filters is only made possible by extensive processes of data labelling, allowing the model to distinguish between what is acceptable and what is not for the final user.

The essentiality of ghost workers for the AI production process in combination with the labour conditions they endure clearly poses dramatic ethical issues for any AI ethics framework. In this sense, the most evidently paradoxical aspect is that the systematic violation of ghost workers’ rights largely appears to arise from the attempt to implement ethical principles in AI products. This means that AI industries have two moral alternatives in this sense: either they take a utilitarian stance, claiming that the suffering of some workers is necessary in order to attain a superior good for the rest of humanity – that would be, in this case, ethical AI; or they acknowledge the moral inconsistency underlying the very idea of the implementation of ethical principles into AI products through the systematic violation of the same ethical principles.

The former is hardly a viable path: most ethical principles promoted in the AI ethics debate and in official documents are characteristic of a strictly deontological model of ethics which makes it virtually impossible to adopt a utilitarian position in this sense. Consider, for example, the following passage taken from the section ‘Respect, protection and promotion of human rights and fundamental freedoms and human dignity’ of the UNESCO's Recommendations:

No human being or human community should be harmed or subordinated, whether physically, economically, socially, politically, culturally, or mentally during any phase of the life cycle of AI systems. Throughout the life cycle of AI systems, the quality of life of human beings should be enhanced. (2022: 18)

It would be quite difficult to justify the exclusion of the AI production process from the ‘life cycle of AI systems’ here mentioned; the same applies considering the principles proposed by Floridi (2022), with particular reference to those of beneficence, non-maleficence and justice.

As for the latter, such acknowledgement should be followed by attempts to solve the related ethical issues. The already-in-process automation of the processes of data gathering, formatting, labelling, cleaning and processing, especially through the increasing development of automatic machine learning and the gradual shift from the use of historical data to the use of synthetic data (Floridi, 2022: 69–77; Steinhoff, 2021: 193–200) for machine learning itself, is far from being a solution in this sense.

This is true not only because it would imply, in turn, the posing of the further problem of technological unemployment, which is explicitly addressed by the UNESCO’s (2022: 36) Recommendations. The issue appears to be rather more complex. First of all, attempts to fully replace the labour of ghost workers through automation have been unsuccessful, and, apparently, the situation is not going to change in the short term (Gray and Siddharth, 2019; Steinhoff, 2021: 189–193). Furthermore, according to recent research, intrinsic computational limits would affect the present form of machine learning, especially concerning error-connection techniques of deep learning: the costs of energy would make the continuous growth and perfection of the sector unfeasible, bearing relevant consequences even for the full automation of the machine learning process (Pasquinelli, 2023: 248–249; Thompson et al., 2021a, 2021b). Nevertheless, while these could still be considered contingent limits from a techno-optimist perspective, Pasquinelli's (2023: 8, 21–22) investigation into the social history of computation and AI seems to suggest that processes of automation in this field tend to multiply ghost workers rather than reducing their number. This feature would characterise such processes since their very inception, with the projects of Gaspard de Prony (in the late 18th century) and Charles Babbage (in the first half of the 19th century), to the point that Pasquinelli identifies them as structural to the field:

The replacement of traditional jobs by AI systems should be studied together with the displacement and multiplication of precarious, underpaid, and marginalised jobs across a global economy. AI and ghost work appear to be, in this view, the two sides of the one and same mechanism of labour automation and social psychometrics. (2023: 21–22)

Analogously, according to the same research, the very separation between manual and mental, and skilled and unskilled labour would be produced and reproduced by the same tendency towards automation, provoking a polarisation between a ‘labour aristocracy’ – in this case, data scientists – and the precarious mass of low-skilled workers. Pasquinelli (2023) proposes this thesis in light of the assumption, already discussed during the Industrial Revolution in England in the 19th century, of the inseparability of ‘mind’ and ‘hand’ in labour, which would imply that the distinction between manual and mental, and skilled and unskilled labour would be a constructed social hierarchy within the division of labour rather than the result of an objective difference. Both the theses seem to be reinforced by the fact that, pace all the announcements of a near automation of machine learning, the labour market of ghost work continues to grow (Gray and Siddharth, 2019)

If this is true, a technological solution for the poor salaries and labour conditions of the ghost workers is not realistically achievable, least of all in the short term: on the contrary, such conditions appear to be deeply imbricated with the very structure of the AI production process. Far from constituting a factual justification for these conditions, this strict entanglement casts a shadow on the entire sector and on its moral claims, making one wonder whether an ethical regulation of the field is possible at all.

Further ethical concern derives from a connected feature that seems to structurally characterise the condition of ghost workers as well; that is, their ‘invisibilisation’ (Gray and Siddarth, 2019; Irani, 2013; Irani and Silberman, 2013). In Pasquinelli's (2023: 8) words, ghost workers ‘are removed from sight to let the show of machine autonomy go on’. In this sense, ghost workers are not simply denied visibility in the public sphere: it is actually pretended that they do not exist (Gray and Suri, 2017; Nakashima, 2018; Yu, 2017). From this point of view, a recent case that has caused some scandal concerns Amazon's grocery stores ‘Just Walk Out’, allegedly managed automatically through facial-recognition cameras, sensors on the shelves and AI. It was instead found that about 1000 Indian workers were actually ‘watching the cameras and labelling footage of shoppers’, to the extent that they ‘reviewed about 70% of sales made in the “cashier-less” shops as of mid-2022’ (Bridle, 2024). Although Amazon commented that their role concerned the training of the AI model, the invisibilisation of such workers remains a blatant phenomenon. This is seemingly due to the tendency within the AI sector, and, to a certain extent, in the digital economy at large, to downplay the still important role played by human labour, thus producing a sort of ‘magical aura’ around the research and products of the industry (Gray and Siddharth, 2019), in fact creating the impression that the level and ubiquity of AI technology development are more advanced than they already are; this achieves the effect of enhancing the industry's public image, promoting the spread of a certain technological imaginary and perhaps bolstering the confidence of financial investors. As is well known, at least since the famous case of the ‘Mechanical Turk’, this tendency appears to be extremely recurrent in any imaginary concerning AI.

However, the invisibilisation of ghost workers also facilitates, other than their social under-representation, their moral misrecognition, their low wages and in general their representation as a ‘residual’ group, destined to disappear at any moment due to automation. Nevertheless, even if this were actually the case, it would remain difficult to argue that prospective automation is a sufficient reason to continue ignoring their current labour conditions and violating ethical principles that are normally deemed fundamental in the spheres of AI research and consumption.

Finally, as will be better seen in the next paragraph, a further ethical issue arises when considering the spatial distribution of the AI production process; that is, its nature of a global value chain: as a matter of fact, ghost workers are very often operating from non-western countries, through specialised companies and platforms. Even in this case, existing global inequalities are taken very seriously in the debate within AI ethics concerning the spheres of research and consumption, while leaving aside the sphere of production (Floridi, 2022: 115–118; UNESCO, 2022: 23, 33–35). The UNESCO's Recommendations, for example, clearly acknowledge the risk for AI technology to reproduce and widen existing global inequalities, thus requiring Member States to take appropriate action in order to reduce them; for example, counteracting the digital divide and putting in place

mechanisms to require AI actors to disclose and combat any kind of stereotyping in the outcomes of AI systems and data, whether by design or by negligence, and to ensure that training data sets for AI systems do not foster cultural, economic or social inequalities, prejudice, the spreading of disinformation and misinformation, and disruption of freedom of expression and access to information. (2022: 28)

It is tragically ironic how the very attempt to ‘ensure that training data sets for AI systems do not foster cultural, economic or social inequalities’ can pass through the reproduction of inequalities along the existing borders of the international division of labour, with the clear risk for new and deeper inequalities to develop along the AI global value chain, if the global market for AI is going to grow as expected in the next years (Global Industry Analysts, Inc., 2023).

The limits of critique: Reassessing existing Marxist and decolonial approaches to AI

In his recent overview concerning the debate within AI ethics, Onur Bakiner (2023: 522) has highlighted that

[t]he field […] appears united in what it excludes: even though capitalism and gender are dominant frameworks to discuss AI ethics, labour exploitation, poverty, global inequality, and gender inequality are not prominently mentioned as problems. […] [T]he scholarship is attentive to bias and discrimination on the basis of (mis) recognition, but does not pay as much attention to the possibility of deepening structural inequality.

Perhaps surprisingly, such evaluation seems to partially apply even to Marxist and decolonial scholars participating in the debate.

A significant limitation of the critical discourse concerning AI ethics is immediately visible within decolonial scholarship, in light of the last problem discussed in the previous section. While the points discussed within this part of the literature are undoubtedly important and should have more space in the wider debate on AI ethics, they seem somehow to lose sight of the possibly most evident issue from a postcolonial and decolonial point of view. For example, critiques addressing the potential and actual forms of Eurocentrism inherent to the allegedly universal concepts underlying the definition of many postulates and principles within AI ethics (Adams, 2021), the risk for AI products to marginalise subaltern voices and perpetuate gender and racial biases (Gebru, 2020) and pre-existing discriminations (Arun, 2020) are all fundamental issues when discussing the relationship between AI and oppressed groups and between AI and the Global South (the same applies to risks connected with ‘digital ethical dumping’, as mentioned in the introduction (Floridi, 2022: 116–117)). It is equally true that these and similar problems clearly depend upon the persistence of colonial relations of power and of inequalities inherited from the colonial era which continue to be reproduced in the postcolonial world, still shaping our global present.

Nonetheless, it is the current structuration of the AI production process (and of many other sectors of the digital economy as well) as a global value chain that primarily shows the persistence of such a colonial legacy in the present: from the point of view of a postcolonial critical approach to AI ethics, it is unavoidable to confront how such persistence structures and shapes both the AI production process itself and the labour conditions of the workers within the AI industry. In this sense, the fact that low-skilled and low-paid jobs are outsourced in the Global South while high-skilled and high-paid jobs remain mainly reserved for the Global North is perhaps the most apparent form assumed by such continuity between the colonial age and the postcolonial present. The practices of data extraction for the AI industry, which are primarily conducted in the Global South (Adams, 2021; Couldry and Mejias, 2019), should be in turn read in this framework, and thus in continuity with the forms of extraction and dispossession of resources that characterised the relationship between the West and the rest of the world during the colonial era, but also in connection with the continuous history of colonial labour exploitation in the Global South.

The case which emerged at the beginning of 2023, concerning low salaries and bad labour conditions in the Kenyan delivery centre of the training-data company Sama, which has notoriously played a significant role in the realisation of ChatGpt, appears particularly significant from this point of view (Perrigo, 2023a). According to the investigation, published by Time, the data labellers employed by Sama (headquartered in San Francisco but, importantly, operating in Kenya, Uganda and India) on behalf of OpenAI were receiving a salary ranging from $1.32 to $2 per hour, while the minimum wage of a receptionist in Nairobi was $1.52 per hour in the same period (it should furthermore be remembered that, at that time, OpenAI was allegedly valued around $29 billion (Jin and Kruppa, 2023)).

However, the extremely low wage paid to these outsourced workers is just the tip of the iceberg. As confirmed by OpenAI, ‘Sama employees in Kenya contributed to a tool […] to detect toxic content, which was eventually built into ChatGPT’ (Perrigo, 2023a), in order to prevent it from generating, as its predecessor GPT-3 was prone to do, ‘violent, sexist, and racist remarks’ (Perrigo, 2023a). The work consisted of reading and labelling 150–250 text passages per day (9 hours a day), such texts including descriptions of ‘child sexual abuse, bestiality, murder, suicide, torture, self harm, and incest’ (Perrigo, 2023a). One of the workers interviewed by Time reportedly said that

he suffered from recurring visions after reading a graphic description of a man having sex with a dog in the presence of a young child. ‘That was torture,’ he said. ‘You will read a number of statements like that all through the week. By the time it gets to Friday, you are disturbed from thinking through that picture.’ (Perrigo, 2023a)

Apparently, the traumatic nature of the tasks caused Sama to interrupt the contract with OpenAI in advance; that is, in February 2022. ²

The concentration of both the exploitation of cheap labour and the processes of data extraction in the Global South thus appears to reproduce the old colonial relationships along the AI global value chain. The same can be said about the inequalities in accessing AI products (together with more general issues such as the global digital divide). As a matter of fact, while this would be an ethical issue per se, implying the risk for AI to reproduce and further existing socioeconomic inequalities in the local, national and international dimensions, the problem is even more serious once the role of the Global South in the AI global value chain is recognised. If people of the Global South fundamentally contribute to the development of ‘western’ AI products with their data and their labour, being then denied access to those products, both for infrastructural reasons (e.g. the digital divide) and for economic and legal reasons (e.g. the patent system and other access limitations to AI products in the framework of the ‘race to AI’ (Floridi, 2022; Smuha, 2021)), the injustice they endure is even more serious. In this sense, although there is no space here to discuss the question at length, it should be considered how intellectual property rights and the patent system deeply contribute to the reproduction of such global inequalities, shaping the fluxes of data extraction from the Global South to the Global North and their employment in training processes, as well as limiting the global diffusion of AI models due to the protection of them by big digital corporations through patents. From this point of view, it could be said that intellectual property rights problematically affect each step of the AI production process, from neural network architectures to the models, from training data sets to specific configurations of training parameters (NXP, 2023).

However, the structuration of the AI process production as a global value chain seems to remain ignored or seriously under-analysed within existing Marxist scholarship as well. For example, as seen, James Steinhoff (2021) has conducted an important investigation concerning the AI production process, also interviewing workers within the industry. While he devotes great attention to the dimension of AI production, nonetheless, as he himself recognises, a representation of ghost workers in the AI industry is lacking in his interviews, due to the fact that at the time he conducted them he ‘was unaware of the significance of their work’ (Steinhoff, 2021: 8). While this does not prevent Steinhoff (2021: 154–155) from conducting an interesting theoretical analysis of their role in the industry, it limits the view of their global class composition and spatial distribution.

Furthermore, the absence of direct consideration of the ghost workers in Steinhoff's account is not without consequences for his theoretical conclusions. As a matter of fact, apart from his discussion concerning the category of ‘immaterial labour’ (Hardt and Negri, 2000: 289–294), which could be said to follow common misunderstandings concerning it (Zaru, 2016) and falls short of historicising it (Mezzadra and Neilson, 2013: 124–125), it is hard not to agree with the critiques he addresses to the post-operaist perspective on ‘cognitive capitalism’ (Steinhoff, 2021: 75–98), also extending it to left accelerationism (Steinhoff, 2021: 229–239). It is certainly true that reading post-Fordist developments of the economy as inherently producing a new autonomy of labour from the capital and thus as intrinsically emancipatory is hardly arguable and overly optimistic, especially taking into account the actual economic developments that occurred between the formulation of the hypothesis of cognitive capitalism (Vercellone, 2005) and today.

Nonetheless, the combination of the Bravermanian approach he counterposes to post-operaism and, as said, the lack of direct consideration of ghost workers within the AI industry (whereas the latter appears to reinforce the former) has possibly led him to fall into the opposite excess. In this regard, his defence of Braverman's ‘deskilling thesis’ (Steinhoff, 2021: 51–55) risks preventing him from acknowledging, in fact, the implications in this context of the simultaneous existence of tendencies and countertendencies in the historical development of the relationship between labour and capital, and more specifically between labour and technology. In the case of the AI industry, in fact, he seems to end up embracing the thesis – widely held in the field – according to which the automation of the AI production process is near, combining it with Braverman's ‘deskilling thesis’: the theoretical conclusion he derives from this is the reversal of post-operaist ones. The advancement of the digital economy, and more specifically of AI development, is thus seen not only as entailing ‘deskilling and fragmentation’ of human labour as a univocal tendency resulting from the impending automation, but also the unavoidable loss of power of workers in the production process itself: after all, he defines his position in terms of ‘communist pessimism’ (Steinhoff, 2021: 236).

Nevertheless, as seen in the previous section, at least two tendencies within the development of the AI industry appear to contradict Steinhoff's analysis: on the one hand, the tendency to automation tends to multiply the number of ghost workers rather than substituting them technologically. In this sense, their growing conflictuality as a class on a global scale – exemplified by the multiplication of their unions – appears to contradict not only their ‘residuality’, but also their loss of power as a result of deskilling processes (Alba, 2023; Alphabet Workers Union, 2023; Gray and Siddharth, 2019; Kerr, 2023; Novak, 2021; Perrigo, 2023a, 2023c). On the other hand, such deskilling processes appear to be far from univocal: as seen, according to Pasquinelli (2023), the tendency to automation in the field of computation seems to produce and reinforce a polarised distinction between skilled and unskilled labour. ‘Deskilling processes’ can thus be understood only in light of their countertendency, which is after all made evident by Steinhoff's own research: it is sufficient, from this point of view, to consider how data scientists and data engineers have been constituted as a ‘labour aristocracy’, high-skilled and high-paid.

The fact is, if at all, that such tendencies and countertendencies affecting labour do not occur in a social vacuum, but in a world where wealth, power and resources are already unevenly distributed: a world, as highlighted, which is still shaped by past colonial relations of power. As far as we can see regarding the global present of work in relation to the development of the AI sector, what clearly appears is that technological unemployment, low-skilled and low-paid jobs, toxic labour conditions, etc., in the industry are unevenly affecting the different hubs of the AI global value chain.

It is therefore necessary to question the causal connection asserted by Steinhoff between deskilling and the loss of power and autonomy on the part of workers within the production process. In this sense, the operaist approach appears much more suitable for understanding, recognising and making visible the autonomy and agency exhibited by AI industry workers, especially ghost workers, compared to the Bravermanian approach advocated by Steinhoff. In his seminal work concerning the relationship between capitalism and machinery, the operaist author Raniero Panzieri (1980:) wrote:

There exists no ʻobjective', occult factor, inherent in the characteristics of technological development or planning in the capitalist society of today, which can guarantee the ʻautomatic' transformation or ʻnecessary' overthrow of existing relations. The new ʻtechnical bases' progressively attained in production provide capitalism with new possibilities for the consolidation of its power. (49; italics in the original)

Post-operaist ‘optimism’ criticised by Steinhoff can thus be equally criticised on the basis of Panzieri's thought: no immediate connection exists between technological development and workers’ emancipation. However, neither is the opposite true:

This does not mean, of course, that the possibilities for overthrowing the system do not increase at the same time. But these possibilities coincide with the wholly subversive character which working-class ʻinsubordination' tends to assume in the face of the increasingly independent ʻobjective framework’ of the capitalist mechanism. (Panzieri, 1980: 49)

Therefore, the agency and autonomy of workers are not only not necessarily diminished by technological development but can potentially grow in parallel with it. The multiplication of labour unions within the AI industry and throughout its entire global value chain; the instruments found by ghost workers to communicate, coordinate and act together; and the ‘rebellion’ of Sama workers in Nairobi seem to confirm this point (Alba, 2023; Alphabet Workers Union, 2023; Gray and Siddharth, 2019; Kerr, 2023; Novak, 2021; Perrigo, 2023a). Recent strikes against the use of AI products in other economic sectors – among which is to be found the important strike of Hollywood writers and actors in 2023 – should be interpreted in the same framework and as a possible basis for a wider politics of alliances centred on the workers’ conflict within the AI industry.

Conclusions: Stretching AI ethics and recognising AI workers' agency

Most papers on AI ethics conclude by proposing further ethical principles to be added to the AI ethical framework and suggesting how to implement them; that is, usually, through legislation (Bakiner, 2023: 522). In contrast, there are no further ethical principles to be derived from this article: as argued in the introduction, the point, if at all, is to include the sphere of AI production into the field of AI ethics, fully taking it into account when ethical principles for AI regulation are discussed, when a moral judgement is formulated on the basis of the same principles usually applied to AI research and consumption, and when guidelines aimed at implementing ethical principles are proposed. Ethical scopes such as the enhancement of human well-being and autonomy and the reduction of global inequalities cannot be attained through means clearly violating the principles informing those scopes. For example, reproducing colonial relations of power in order to contrast racial and gender biases in AI products is, euphemistically, paradoxical and morally inconsistent.

On a theoretical level, it is not difficult to envisage how existing principles in AI ethics could be extended to the dimension of production, thus constituting an ethics of AI production. For example, the extension of the principle of justice as characterised by Floridi (2022), or the values of peace, inclusivity, justice and equity as characterised by UNESCO (2022), would imply that companies involved in the AI industry should take responsibility for ensuring ethical and fair labour conditions, including adequate wages, physically and psychologically safe labour environments, the right to form unions, and so on, for all their employees, be they directly hired or outsourced, along the entire global value chain. Alternatively, to give another example, the extension of the principle of transparency would imply that companies should be transparent about the AI production process and the labour practices and policies they adopt in it, including information about who the workers are, where they operate, under what conditions they work, and so on.

However, the real challenge is more practical in character. While the debate in AI ethics is very prolific of ethical principles, it is surprisingly uncreative when it comes to actually implementing such principles: as mentioned, most papers simply propose to resort to legislation (or even to self-regulation) (Bakiner, 2023: 522). While the risk of utmost ineffectiveness for such principles clearly concerns also the spheres of research and consumption (Munn, 2023), this is even truer for production. The reason for this was clarified by Karl Marx (1992: 279–280) a long time ago: it is the sphere of circulation, that we are forced to ‘desert’ when entering the ‘hidden abode of production’ – ‘on whose threshold there hangs the notice “No admittance except on business” ’ – that is ‘the very Eden of the innate rights of man’, ‘the exclusive realm of Freedom, Equality, Property and Bentham’, the sphere of rights and morality. The sphere of production lies outside this realm. The history of the working-class movement has further demonstrated how legislation alone appears to be ineffective in regulating the production processes from the above: when such legislation has been successfully implemented, there were workers’ struggles behind it. It is the role of such struggles that is necessary to take into account when it comes to thinking of the implementation of the ethical principles discussed within the debate on AI ethics. If we assumed as true the univocal connection between deskilling and the loss of workers’ power within the AI production process proposed by Steinhoff, in line with Braverman's ideas, this would clearly not be an option. On the contrary, through Panzieri, we have seen that it is at the same time possible and necessary to recognise the actual and potential expression of the agency and autonomy of the workers in the AI industry, other than of the workers outside the industry who are nonetheless affected by AI products, as is increasingly manifesting today.

For this reason, such recognition also implies a further step: it is not sufficient to recognise the necessary role of AI workers in the implementation of principles for an ethical AI. AI workers are usually not even considered among the ‘stakeholders’ who should participate in AI governance, regulation and ethics according to the most important recommendations concerning AI ethics. On the contrary, the recognition of their agency and autonomy as workers of the industry should also imply the possibility of such participation for them, as well as the expression of their needs, interests, positions and decisions within this framework, both as individuals and as groups (i.e. through labour unions). In other words, their voices should be heard and taken into account within the very debate concerning AI ethics, and more generally within decision-making processes concerning the development, production, implementation and use of AI products in our societies.