1. Democratising AI: an inflated agenda

To begin, this section reviews three major understandings of DAI, and their corresponding strategies. However, it is argued that two of these understandings not only presuppose a controversial outlook of democracy, but they also overload DAI with too many commitments, rendering DAI an empty rather than a realistically action-guiding agenda for individuals. This gives us reason to limit our discussion of DAI to widening participation in corporations’ AI decisions, an interpretation of DAI to be examined further in the following parts.

The term ‘democracy' is usually interpreted in a manner inclusive of numerous ‘good' political values, institutions or practices, such as freedom, autonomy, equality, solidarity, active participation, checks and balances, and so on. The DAI agenda is no less ambiguous, since it is linked by AI scholars and practitioners to a range of trending values or practices for ethical and/or responsible AI. For instance, there are the claims that ‘[democratizing AI] may help mitigate algorithmic injustice' (Lin, Reference Lin2024, p. 102), that it may ‘help address the knowledge gap in the development and deployment of AI technologies' (Banifatemi et al., Reference Banifatemi, Miailhe, Buse Çetin, Cadain, Lannquist, Hodes, Braunschweig and Ghallab2021, p. 231), and that it may serve to ‘mediate the powerlessness experienced by oppressed groups’ (Okidegbe, Reference Okidegbe2023, p. 1688). Sometimes DAI is used by tech firms to underpin their technical or commercial ambitions around AI products. For example, IBM interprets DAI as ‘providing AI access to a wider range of users beyond machine learning (ML) experts. Common means of improving access include reducing AI costs and incorporating AI into tools and platforms people are already using' (Gomstyn & Jonker, Reference Gomstyn and Jonker2024). This interpretation of DAI is clearly underpinning the efforts of tech firms to create more user-friendly AI products.

Despite the many senses of DAI, there are three relatively popular understandings of it. The first understanding of DAI sees it as the endeavours to ensure equitable distributions, from both the public and private sectors, of the economic costs/benefits of AI (Clough & Otterbacher, Reference Clough, Otterbacher, Carayannis and Grigoroudis2023; Coeckelbergh, Reference Coeckelbergh2024; Gomstyn & Jonker, Reference Gomstyn and Jonker2024; Randieri, Reference Randieri2024; Rubeis, Dubbala & Metzler, Reference Rubeis, Dubbala and Metzler2022). This echoes the familiar concerns over AI’s potential to exacerbate job displacement, income inequality, and the power disparities between individuals and firms, as noted by Randieri (Reference Randieri2024). Concerns have been raised over AI as an exploitative industry, the quick advancement of which is often predicated on tech firms’ unfair extraction of the natural resources, environment and labour of the Global South (Crawford, Reference Crawford2021). To mitigate these issues, scholars have proposed various initiatives. One example is universal basic income (UBI), to shield individuals against the risks of job displacement driven by automation, or to leverage AI’s power to administer the complex financial frameworks required to make UBI possible (Allegri & Foschi, Reference Allegri and Foschi2021; James, Reference James and Liao2020; Jozaghi, Reference Jozaghi2024; Marr, Reference Marr2024). Another example is the Windfall Clause, according to which there ought to be some profits ‘a firm could not earn without achieving fundamental, economically transformative breakthroughs in AI capabilities' (O’Keefe et al., Reference O’Keefe, Cihon, Garfinkel, Flynn, Leung and Dafoe2020, p. 327). Moreover, those profits should be ex ante donated to initiatives aimed at benefiting humanity widely. There are also proposals to introduce a tax on the use or development of AI (e.g. robots), in order to mitigate the economic disadvantage that individuals may suffer from emerging AI technologies (Abbott & Bogenschneider, Reference Abbott and Bogenschneider2018; Mazur, Reference Mazur2018; Oberson, Reference Oberson2025). Some propose digital commons (i.e. publicly owned technologies and resources, such as data and software) as a way of widening the beneficiaries of AI (Coeckelbergh, Reference Coeckelbergh2024, Reference Coeckelbergh2025; Furendal, Reference Furendal2025; Muldoon, Reference Muldoon2025). This sense of ‘democratising' AI, in short, is about allocating the costs and benefits of AI in ways that avoid disadvantaging certain social groups unfairly.

The second understanding centres DAI on the accessibility of AI systems. It refers broadly to the measures aimed at making it easier for individuals to use AI (Clough & Otterbacher, Reference Clough, Otterbacher, Carayannis and Grigoroudis2023; Coeckelbergh, Reference Coeckelbergh2024; Gomstyn & Jonker, Reference Gomstyn and Jonker2024; Lin, Reference Lin2024; Sclove, Reference Sclove1995). An individual’s access to AI technologies can be limited by a number of factors. One factor, as Gomstyn and Jonker (Reference Gomstyn and Jonker2024) pointed out, is people’s general knowledge of the essential technologies (e.g. machine learning) in relation to AI. Without such knowledge, people may find it challenging to use AI. Another factor is the opacity of AI systems, which ‘might deliver the right results, but they do not provide users or affected parties any insight as to how they came to produce those results' (Vaassen, Reference Vaassen2022, p. 2). This opacity problem often contributes to people’s hesitancy to trust AI-produced decisions, especially in contexts (e.g. healthcare, admissions, hiring) where allocation of responsibility is crucial (Lebovitz, Lifshitz-Assaf & Levina, Reference Lebovitz, Lifshitz-Assaf and Levina2022; Smith, Reference Smith2021; Zeide, Reference Zeide2022). In addition, it was found that people have differential access to information communication technology, depending on their demographics, socio-economic circumstances, education, digital infrastructure and so on, whereas this phenomenon is often described as the digital divide (Lythreatis, Singh & El-Kassar, Reference Lythreatis, Singh and El-Kassar2022). To address these issues, some have advocated for initiatives focused on establishing AI’s trustworthiness, namely a set of properties or standards AI systems must exhibit to demonstrate ‘that they are deserving of our trust' (Durán & Pozzi, Reference Durán and Pozzi2025, p. 16). This approach is partly underscored by the emergence of AI ethical guidelines, developed by both public and private organisations, that specify the ‘best practices' for AI development and deployment, such as the EU’s Ethics Guidelines for Trustworthy Artificial Intelligence, IBM’s and NVIDIA’s principles for trustworthy AI, (Díaz-Rodríguez et al., Reference Díaz-Rodríguez, Del Ser, Coeckelbergh, López de Prado, Herrera-Viedma and Herrera2023; EU, 2019; Gomstyn, Jonker & McGrath, Reference Gomstyn, Jonker and McGrath2024; Lu, Zhu, Whittle & Xu, Reference Lu, Zhu, Whittle and Xu2023; Pope, Reference Pope2024). In response to the digital divide, international organisations, companies and scholars have also been calling for widening computer and internet access, for instance (Mathiesen, Reference Mathiesen2012; Reglitz, Reference Reglitz2020, Reference Reglitz2024).

While these two understandings of DAI highlight a range of goals and values that merit our attention, they will not be examined any further, for two reasons. Regarding ‘equitable distribution of AI’s cost and benefits' as an instance of ‘AI democratisation' is potentially misleading, or at least rests on a controversial view of the defining purposes of democracy. Here, a distinction must be drawn between distributive justice and democracy. Although democracy is a form of rule partly characterised by its commitment to equalising the influence and participation of the relevant members, it rarely promises to warrant distributive justice, that is, the just allocation (by whatever standards) of resources, burdens and opportunities among individuals. To clarify, this does not mean that advocates for democracy or democratisation are never concerned about distributive justice, since certain issues in distributive justice (e.g. economic inequality) have implications for the extent to which democratic ideals (e.g. citizens’ equal influence on politics) are realised. But distributive justice is not a defining ideal of democracy that is widely shared.Footnote ¹ Likewise, the distributive justice of AI does matter for all sorts of reasons, but putting it under the umbrella of AI democratisation risks diluting the values that have traditionally been attached to the term ‘democracy', and therefore blurring the boundaries between DAI and other normative agendas. Without establishing what makes DAI distinctive, the agenda is far less likely to provide any added value to other attempts to make AI more equitable or user-friendly.

In the rest of this article, therefore, I shall only focus on the third understanding DAI: widening participation (especially from the general public) in the decisions around the design, development and/or governance of AI systems (Coeckelbergh, Reference Coeckelbergh2024; Gomstyn & Jonker, Reference Gomstyn and Jonker2024; Lin, Reference Lin2024; Noorman & Swierstra, Reference Noorman and Swierstra2023; Zimmermann, Rosa & Kim, Reference Zimmermann, Rosa and Kim2020). This is also the understanding of DAI to which the BP is more relevant. According to this understanding, there should be measures aimed at including, as many stakeholders as possible, in the making, application and regulation of AI technologies. As Noorman and Swierstra argued, ‘all those impacted by these [AI] technologies should have a say over the values and political dispositions built into these technologies…democratic processes should also, self-reflectively, be directed at diagnosing and guiding the ways technologies affect democratic governance' (Noorman & Swierstra, Reference Noorman and Swierstra2023, p. 570). The idea is not just that more people, rather than a few economic and technical elites, should have an opportunity to shape how AI is built or used. It also expresses the thought that our political processes should govern AI systems in such a way that upholds democratic values.

There are three more widely discussed approaches aligned with this third understanding of DAI. The first approach, R&D Inclusion, is to widen participation in organisational efforts in researching, developing and designing new AI products. Participation, however, should not only come from members within an organisation. It should also, as Sclove phrased it, allows ‘communities, groups, and citizens – including those today least empowered – to help directly initiate some R&D programs and design technologies responsive to their needs' (Reference Sclove1995, p. 207). This echoes the calls for participatory design of AI, which place an emphasis on including a wider range of stakeholders in developing new AI systems (Birhane et al., Reference Birhane, Isaac, Prabhakaran, Diaz, Elish, Gabriel and Mohamed2022; Buddemeyer et al., Reference Buddemeyer, Nwogu, Solyst, Walker, Nkrumah, Ogan and Stewart2022; Gill, Reference Gill1989; Schmager, Pappas & Vassilakopoulou, Reference Schmager, Pappas and Vassilakopoulou2025; Zytko, Wisniewski, Guha, Baumer & Lee, Reference Zytko, Wisniewski, Guha, Baumer and Lee2022).

The second approach, Inclusive Intra-Organisational Oversight, is to reform or establish intra-organisation bodies that review or oversee an organisation’s AI practices. In particular, those bodies should represent, as much as possible, the voices of different stakeholders, including those outside of the organisation (Blackman, Reference Blackman2022). A range of tech firms have already attempted to set up internal teams to oversee their AI practices, such as IBM’s AI Ethics Board, Google’s short-lived Advanced Technology External Advisory Council (ATEAC), and OpenAI’s Safety and Security Committee (OpenAI, 2024; Rossi, Reference Rossi2020; Wakefield, Reference Wakefield2019). However, these initiatives, are often criticised on the grounds that they are instances of ‘ethics washing', that corporates’ commercial interests often trump the priorities and decisions of such teams, that such teams often play an advisory role with no substantive influence on a corporation’s AI decisions, and so on (Gilbert, Reference Gilbert2020; Himmelreich, Reference Himmelreich2023; Lee & Getahun, Reference Lee and Getahun2024). Therefore, there is the argument that such AI oversight bodies within organisations should have greater power and include a wider range of representatives, such as employees at different levels of the organisational hierarchy, experts external to the relevant organisation, affected citizens, and the like (Himmelreich, Reference Himmelreich2023; Schuett, Reuel & Carlier, Reference Schuett, Reuel and Carlier2025).

The final approach, Democratic Control of Tech Firms, is to subject tech firms to greater control by the democratic state. Risse, for instance, argues that the power of big tech firms ‘needs to be harnessed for democratic purposes. It is hard to see how that can be done unless the Big Tech companies are either dismantled entirely … or treated and thus regulated as public utilities alongside enterprises like phone companies' (Reference Risse2023, p. 47). A key implication of this view is that aligning AI companies with democratic values will require greater state control of such companies, although the state must be democratic itself. Similarly, Coeckelbergh maintained that ‘one could consider establishing a new institution that better links technology-development contexts to democratically elected bodies and does so in ways that redistribute power towards the elected bodies' (Reference Coeckelbergh2024, p. 75). In a like manner, Aytac has proposed the establishment of Citizen Boards of Governance (CBGs), which are randomly selected citizens ‘authorized to govern the algorithmic infrastructure of Big Tech platforms' (Reference Aytac2024, p. 1431). Overall, the aim of these proposals is to hold tech firms accountable to social or political bodies where the core members are selected by democratic means (e.g. elections or sortition).

2. Democratising AI and the boundary problem

The previous section has shown that DAI, understood as widening participation in the decisions around AI’s design, development and governance, typically entails three approaches: (1) R&D Inclusion, (2) Inclusive Intra-Organisational Oversight and (3) Democratic Control of Tech Firms. This section turns to the implications of the BP for these approaches, by revisiting several major principles and considerations from which political theorists have attempted to diffuse the BP of democratic citizenship. More importantly, this section proposes a solution to the boundary problem of DAI, arguing that corporations have legitimate reasons to exclude non-employees from their efforts in fostering R&D Inclusion as well as Inclusive Intra-Organisational Oversight. I shall also highlight the immense complexities involved, as we seek to address the BP in the context of Democratic Control of Tech Firms.

Initially, the BP is a question more relevant to issues in democratic citizenship: who should constitute the ‘demos', the citizens who share equal rights and responsibilities (notably, to vote) in a democratic society (Abizadeh, Reference Abizadeh2012; Miller, Reference Miller2020; Owen, Reference Owen2012; Simmons, Reference Simmons2013; Song, Reference Song2012; Whelan, Reference Whelan1983)? Here, an important point should be flagged: citizens within the same ‘demos' are supposed to share equal rights and responsibilities. Citizens and non-citizens, even when they reside in the same territory and are subject to the authority of the same government, do not have equal rights and responsibilities. Expanding the list of rights or responsibilities of non-citizens does not immediately turn them into democratic citizens, the full members of the ‘demos.' Non-citizens become full members of the ‘demos' only when they have rights and responsibilities that are equal, or at least comparable to the citizens. This is a crucial feature of the BP to be observed, as we consider whether it is indeed challenging for DAI.

Who should be included in the decision-making processes on AI, a question in the spirit of the BP, has often been considered a potential challenge for the DAI agenda as well (Han, Reference Han2025; Himmelreich, Reference Himmelreich2023; Murphy & Taylor, Reference Murphy and Taylor2024; Noorman & Swierstra, Reference Noorman and Swierstra2023). This concern is most clearly expressed by Himmelreich, who mentions how the global impact of AI troubles any attempts to ‘include' more people in the decisions around AI: ‘As typical for software and the digital economy generally, AI, data or access to APIs can easily be transferred and marketed across the globe. The society that lives under AI is the global society' (Reference Himmelreich2023, p. 1335). Before specifying the senses in which the BP may challenge DAI initiatives, I shall consider several major principles aimed at addressing the BP, in normative democratic theory.

First, the all-affected interests principle is probably one of the most celebrated approaches to the BP (Dahl, Reference Dahl1990; Fung, Reference Fung, Nagel and Smith2013; Goodin, Reference Goodin2007; Owen, Reference Owen2012; Rosenberg, Reference Rosenberg2020; Wilson, Reference Wilson2022). The principle, in its most basic formulation, states that ‘All whose interests are actually affected by a decision should be able to participate as equals in the democratic decision-making process' (Owen, Reference Owen2012, p. 131). The intuitive appeal of this principle is widely recognised, since a lot of citizens and even non-citizens (e.g. immigrants) are subject to the impact of the decisions made by their democratic government. Likewise, individuals, perhaps all of us around the world, are strongly impacted by corporates’ decisions around AI, which essentially sustains a number of digital tools. However, if the all-affected interests principle holds, then DAI would seem to require including everyone affected in corporates’ decision-making processes over AI. This is practically impossible.

Second, solidary has also been cited as a crucial consideration in resolving the BP (Miller, Reference Miller, Hutchings and Dannreuther1999, Reference Miller2009, Reference Miller2020; Song, Reference Song2009, Reference Song2012). In Song’s terms, ‘Solidarity might be forged through a shared history, shared culture, and/or shared values' (Song, Reference Song2012, p. 47). On one view, solidarity motivates people to ‘listen to and understand one another' rather than make decisions based on self-interests (Song, Reference Song2012, p. 47). If the ‘demos' lacks solidarity, its characteristic activity (i.e. democratic participation) will produce outcomes based on self-interests rather than mutual understanding. To ensure the solidarity of participants in any democratic initiatives (including those related to AI), therefore, there is good reason to include only people with a relatively solid collective identity, marked by shared values and culture. Clearly, this criterion has implications for whether participants in DAI endeavours (e.g. intra-organisation oversight bodies of AI) ought to include only people sharing a similar identity.

Third, the all-subjected principle has also been proposed as a solution to the BP (Abizadeh, Reference Abizadeh2021; Goodin, Reference Goodin2016; Goodin & Arrhenius, Reference Goodin and Arrhenius2024; Honohan, Reference Honohan and Bauböck2018; Inoue, Reference Inoue2024). The essence of the principle is to include only ‘subjects' within the demos. For individuals to be ‘subjects' of a democratic body, they must be coerced into following the decisions of that body (Abizadeh, Reference Abizadeh2012; Goodin, Reference Goodin2016). The all-subjected and all-affected interest principles are different, however. The all-affected interest principle is much broader in scope: it implies that the relevant democratic decision-making body on AI should include individuals affected by its outcomes, even if those outcomes are enforced in non-coercive manner. For instance, the all-affected principle would imply that I have a say over a company’s decisions over an AI tool it developed, even if I use that tool on an entirely voluntary basis. But the all-subjected principle will urge the body to include only those who will be forced or coerced into following its decisions. Imagine, for instance, the state requires citizens to record their health conditions through a government-designed AI tool every week during a pandemic; anyone failing to do so may be prosecuted. According to the all-subjected principle, this will be a particularly concerning scenario that calls for widening citizens’ participation in the decisions around that tool, with the reason being that they are coerced into using that app.

Fourth, the effects of inclusion also matter for what and how people should be included by the ‘demos.' For instance, in liberal political theory as well as epistemic theories of democracy, a democratic procedure that includes more voices is valuable, because the increased diversity of inputs will ultimately enhance the quality of the procedure’s outcomes (Anderson, Reference Anderson2006; Cohen, Reference Cohen and Estlund2002; Estlund, Reference Estlund2008; Goodin & Spiekermann, Reference Goodin and Spiekermann2018; Landemore, Reference Landemore2017; Mill, 1859 [Reference Mill, Philp and Rosen2015], 1861 [Reference Mill, Philp and Rosen1861 [2015]]). There may be reason to include more individuals, whoever they are, insofar as this improves corporations’ decisions around AI design, development and governance. Meanwhile, there is the view that one’s understanding of politics is enhanced through participation in democratic activities (Aristotle, 1995; Landemore, Reference Landemore2017; Mill, 1859 [Reference Mill, Philp and Rosen2015], 1861 [Reference Mill, Philp and Rosen1861 [2015]]). To understand how to vote well, one must have the right to vote; likewise, one may need an opportunity to participate in the decision-making activities on AI, to gradually build the literacy essential for informed decisions over AI-related issues. The upshot is that DAI initiatives can be justified on educational grounds, building individuals’ ethical and AI literacy. Finally, solidarity is not necessarily a precondition for democratic membership; rather, the latter might be a crucial condition for solidarity instead. In democratic theory, for example, some have stressed how equal democratic participation serves to foster the collective identity and reciprocal relationships between participants (Christiano, Reference Christiano2004, Reference Christiano2008; Rawls, Reference Rawls1999, Reference Rawls2001, Reference Rawls2005). Seen in this light, certain DAI proposals can be conceived as endeavours to build solidary among those who are included.

Having clarified multiple responses to the BP, let us now turn to their implications for the following three DAI initiatives: (1) R&D Inclusion, widening participation in organisational efforts in researching, developing and designing new tech products; (2) Inclusive Intra-Organisational Oversight, reforming or establishing intra-organisation bodies that oversee an organisation’s AI practices, whereas those bodies should represent as far as possible the voices of different stakeholders; and (3) Democratic Control of Tech Firms, subjecting tech forms to greater control by the democratic state. Clearly, each of these approaches invoke a different sense of ‘demos’.

For instance, the ‘demos' of (1) refers to the participants in corporations’ R&D processes; the ‘demos' of (2) refers to participants in the relevant AI oversight bodies within an organisation; the ‘demos' of (3) refers much more broadly to the citizenry of a democratic society. Such differential constitution of the ‘demos' makes the BP challenging for these approaches in different ways.

I shall first comment on (1) and (2). In my view, while there may be some value of widening participation, there is no compelling reason to equalise participation in the R&D of AI products, as well as their intra-organisational AI oversight bodies, to those outside of the same organisational ecosystem. To clarify, the view does not say that the R&D and oversight teams of an organisation’s AI practices should never invite the inputs of non-members (e.g. communities, citizens) of the organisation. The BP may not be concerning at all, if what is being demanded is merely that inputs of non-members should have some role to play. The reason, as I noted earlier, is that the BP is concerned with the criteria for which someone should be granted rights and responsibilities, on a par with other members within an organisation or society. For example, it is common for corporates to invite non-employees to beta-test their AI products before rolling them out on the market, to improve the user experience of such products. But this does not mean that the inputs of employees or non-employees share an equal degree of authority over how AI products ought to be developed and governed. Yet, if what DAI is demanding is merely that non-employees get to shape a corporation’s strategies for AI development or governance in some moderate ways, regardless of how much influence these non-employees can effectively exercise, it is not obvious what makes the DAI agenda ambitious enough to differentiate itself from other proposals of participatory AI design and/or governance.

Furthermore, there is a compelling case for the employees within a corporation to remain the key decision-makers of R&D and oversight of AI, while corporations are justified in assigning to non-members less influence over these matters. The main ground for this view is that non-members of organisations do not plausibly fulfil three out of four criteria for democratic inclusion above. Although the interests of individuals are often affected by organisations’ AI practices, of which they have no membership, such individuals lack the kind of solidary and subjection to an organisation’s authority that supports granting those individuals equal rights and responsibilities, within the relevant decision-making processes. But employees of an organisation are different. Compared to non-employees, employees are marked by a higher level of shared interests, culture and values, that constitute their collective identity. Because they belong to the hierarchical structure of the organisation, employees are also often coerced into prioritising the commands and interests of those from senior levels of management; employees have relatively little leeway to disobey, unless they are prepared to be marginalised or even leave their organisation. The mere fact that individuals’ interests are being affected by companies invested in AI development and application may only justify that their voices be included by organisations to a modest degree, but this does not warrant equal participation of members and non-members of an organisation.

In particular, it is unclear how the positive effects of inclusion must be achieved by providing both employees and non-employees with an equal say over R&D and organisational AI governance. As mentioned earlier, if collective decision-making is constituted by participants with a relatively loose shared identity, they are likely to deliberate and make decisions based on self-interests rather than mutual understanding. Also, citizens’ skills and knowledge of AI research, development and oversight are not acquired exclusively through participation, on an equal footing, with intra-organisational professionals working on these domains. It has become a lot more common today for individuals to have access to the tools or software to develop AI systems, outside of an organisational setting. Given the availability of resources and opportunities for individuals to educate themselves about AI-related issues, it is also hard to justify giving non-employees substantive power, comparable to employees, to contribute to decisions around the R&D and governance of AI. Finally, the view that inclusive decision-making enhances the quality of decisions being made, as proposed by epistemic theorists of democracy, often assumes that the participants, on average, should pass some modest thresholds of competency, or that there are expert moderators who serve to organise the wisdom of everyone in ways contributing to high-quality decision-making (Brennan, Reference Brennan2011, Reference Brennan2016; Somin, Reference Somin2014). Yet, if the participants in the R&D and governance of AI are supposed to meet these requirements, this does not seem to deviate meaningfully from existing common practices by corporations to include outsiders or laymen to improve their research, development and oversight of AI, while allowing expert employees to moderate the inputs of such individuals and transform them into actionable items for the corporate.

In short, these facts reveal a broader challenge of widening participation in the development and/or governance of AI products. If widening participation merely means non-employees have some say over AI development and/or governance of corporations, this does not seem highly controversial and effectively trivialises the DAI agenda. But if widening participation means equalising the say and responsibilities of employees and non-employees with respect to AI development and/or governance, this requires a much stronger justification as it involves revisiting the criteria for membership within the ‘demos', which refers to the corporation in this context.

Let us turn to Democratic Control of Tech Firms. This proposal is by nature different from the previous two, since it is not primarily concerned with providing employees and non-employees with an equal or at least comparable say over AI-related corporate decisions. It is essentially about expanding the state’s control of (tech) firms, in such a way that their AI practices take democratic values more seriously. Here, the ‘demos' refers broadly to the entire citizenry of democratic society. However, the tendency of tech firms to disregard democratic values in their AI practices does not seem clearly owed to the fact that our ‘demos' is not sufficiently inclusive. Rather, that issue is more obviously linked to two general facts: (a) tech firms have tremendous power, compared to other citizens, in shaping the agenda and outcomes of democratic processes, and (b) existing democratic institutions do not have adequate mechanisms to hold corporations accountable to AI practices that undermine democratic politics. These facts, however, are not clearly addressable by granting more people with the rights and responsibilities of citizens, at least not without some further justifications. Also, resolving the BP of who should quality as full members of a democratic society is a much broader question, an informed response to which requires more careful consideration of the moral rights, interests and responsibilities of individuals, across a wide spectrum of political issues beyond AI governance. The BP, therefore, can hardly be framed as a suitably specific threat to Democratic Control of Tech Firms.

3. Insights from workplace democracy

Section 2 has examined the senses in which the BP may or may not be challenging for DAI. Specifically, it has shown that either the ‘widening participation' proposals associated with DAI are not ambitious enough to render themselves vulnerable to the BP, or such proposals disregard the legitimate interests of corporations in prioritising employees over non-employees, when it comes to AI development and governance. But the BP is no more threatening to Democratic Control of Tech Firms, as the problem does not respond aptly to the distinctive concerns such expanded control is meant to address.

So far, I have defended the thought that the BP for DAI, understood as widening participation in AI development and governance, is largely resolvable. However, underneath the BP-motivated struggles between different options to broaden participation in AI-relevant decision-making processes, is a potentially concerning shift of people’s attention away from a major factor disincentivising corporate leadership to prioritise ethical AI practices: the power disparities among members within the same corporate hierarchy. In other words, the main problem is not whether and how inclusive the process of AI development or governance should be, but the power dynamics that suppress voices around ethical AI. The aim of this section, then, is to explain why workplace democracy, rather than DAI, presents a more promising normative aspiration in response to such power dynamics, although the former needs to be supplemented by Democratic Control of Tech Firms to have a more significant effect.

Workplace democracy refers broadly to the institutional proposals aimed at infusing democratic values into a corporate setting (Foley & Polanyi, Reference Foley and Polanyi2006; Frega, Reference Frega2021; Frega, Herzog & Neuhäuser, Reference Frega, Herzog and Neuhäuser2019; Landemore & Ferreras, Reference Landemore and Ferreras2016; Mayer, Reference Mayer2000; Vrousalis, Reference Vrousalis2019; Yeoman, Reference Yeoman2014). The most notable feature of workplace democracy is that it seeks to turn workplace into ‘systems for the exercise of power by workers or their representatives' (Frega et al., Reference Frega, Herzog and Neuhäuser2019, p. 1). There are a variety of arguments for workplace democracy, appealing to the similarities between the state and firms, the importance of equality, democratic education, meaningful work and so on (Frega et al., Reference Frega, Herzog and Neuhäuser2019). Unlike the DAI agenda, however, expanding organisational participation from non-employees is not often one of the characteristic arguments for workplace democracy. Rather, their focus is usually placed on creating less hierarchical organisational structures, or any other ways to mitigate the power disparities between employees at different organisational levels.

The DAI agenda is often pitched by ethicists and AI practitioners as a way of accelerating responsible or ethical AI decisions. Because many DAI initiatives focus on widening participation in AI design and governance, who should be included in the relevant participatory proposals – motivated by the BP – naturally emerges as a matter of concern. However, an over-emphasis on broadening participation unfortunately shifts our attention away from a main factor driving corporate leadership to make unethical decisions on AI: that is, employees advancing ethical AI decisions are in a fragile position within their organisation. As is observed by a recent report of Ada Lovelace Institute, ‘While calling for democratising AI, many of the biggest technology companies have also made mass redundancies to internal teams that are focused on ‘responsible AI’ and ethics initiatives – the precise teams that would presumably undertake public participation work' (Groves, Reference Groves2023, p. 11). If corporate leaders already had limited incentives to enforce the advice around ethical AI from their employees – who, compared to non-employees, can be reasonably expected to know better the organisational contexts that produce AI decisions – then why would they prioritise the opinions of the outsiders, who may even be less familiar with the culture, interests and priorities of the corporation?

The problem, in short, is not that corporate leaders are unaware of what decisions should be made to ensure ethical AI development and governance, but rather that ethical AI practices do not weigh as much as profitability in the eyes of many corporate leaders. Voices around ethical AI, whether they come from the team responsible for advancing it or from other employees, are usually subordinate to the interests of corporate leadership. This often results from the relationship of domination between corporate leaders and employees. This relationship inevitably exists, however, insofar as employees need to stay in their current organisations for an income, without substantive exit options. But it is possible to democratise the workplace, to moderate this relationship of domination. Ways of democratising the workplace ‘range from rather bland organizational reforms of hierarchical structures to delegation systems based on workers' representation, such as the German codetermination model, to more involved forms of workers’ direct participation through works councils…[as well as] ownership of the firms by workers’ (Frega et al., Reference Frega, Herzog and Neuhäuser2019, p. 1). It is beyond the scope of this article to defend a specific proposal of workplace democracy. However, the normative aspiration of it (i.e. addressing domination in the workplace) responds more directly to a more crucial factor impeding corporates’ interest in taking ethical AI practices seriously, rather than as a symbolic gesture, while this factor rarely invites serious inquiry in the scholarships on AI democratisation.

It is beyond the scope of this article to develop a full defence of workplace democracy, and there are several legitimate concerns over my view that should be acknowledged. My responses to these concerns, though, will shed light on several ways in which workplace democracy and Democratic Control of Tech Firms, as a paradigmatic initiative of the DAI agenda, complement each other. The first concern is that, if some corporations object to DAI on the basis that this impedes innovation and efficiency, why would they opt for a more radical proposal to democratise themselves? Arguably, corporations are unlikely to use workplace democracy to elevate those responsible for ethical AI, at least voluntarily. Corporate leadership is, in many cases, concerned with profitability and generally views ethical AI as a cost that interferes with this goal.Footnote ² To be sure, it will be challenging to enforce any proposal to empower workers in the workplace, whether this takes the form of widening participation in AI governance and development, or limiting the power disparities between employers and employees. Realistically, perhaps the most viable way to incentivise corporations to take ethical AI seriously is to impose greater political and economic costs on corporate leaders when they fail to hold AI practices up to sound ethical standards. This is when Democratic Control of Tech Firms becomes complementary, or even essential. Public institutions authorised through democratic means serve as external forces imposing pressures on corporations, to hold them accountable not just for their ethically concerning decisions on AI, but also for the ways in which corporations moderate the relationship of domination between employers and employees.

The second concern is that workplace democracy may erode efficiency and corporate freedom, as two drivers for AI innovations. For example, some argued that monitoring from higher levels of corporate leadership promote employees’ performance at work (Holmstrom & Milgrom, Reference Holmstrom and Milgrom1994). But such monitoring may not be possible in a democratic workplace. Moreover, capital owners have a legitimate interest in entrepreneurship, which implies creating organisational structure and culture in accordance with their own values (Tomasi, Reference Tomasi2012). In response, two points should be flagged. On the one hand, efficiency and corporate freedom are not the only decisive factors in evaluating the moral promise of different governance strategies of AI. An emphasis on efficiency and corporate freedom, in fact, echoes the familiar but contested utilitarian assumption that good AI policies depend enormously, if not fully on the benefits they can produce. But this is not a comfortable assumption to be granted, as the pursuit of AI’s benefits must be weighed against the moral rights and interests of other individuals as well. Efficiency and corporate freedom are desirable, but not always decisive. On the other hand, these concerns are not characteristic of workplace democracy, but of various DAI initiatives as well. Similar arguments, for instance, can be made of employees’ interest in having more ownership of the strategies they impose on their own AI products or governance structures as well, due to employees’ shared culture, interests or identities that make such products or structure possible. It is thus reasonable for corporations to assign less weight to the inputs of non-employees, when making decisions around AI development or governance, and prioritise the voices of their own employees. While this is an implication of my earlier comments, it does not follow that a similar point could be made of workplace democracy. In essence, workplace democracy aims to equalise the power of the individuals within, not those outside of a single organisational structure. Employees have a stronger claim to be given rights and responsibilities comparable to those of their leaders, precisely because their corporate membership is a plausible basis for their shared culture, interests and identity. Thus, even when capital owners have a legitimate interest in entrepreneurship, it should be competed against the interests of employees in having meaningful control of their common workplace. Relatedly, Democratic Control of Tech Firms is valuable, because it honours institutional mechanisms through which citizens express their views on how, if at all, workplace democracy, or other AI democratisation initiatives, should be enforced.

Finally, it might be argued that, although reducing the power inequality between employers and employees provides the latter with greater opportunity to advance ethical AI decisions, this does not necessarily incentivise employees to prioritise ethical AI over the profit. Arguably, employees could have even stronger incentives to prioritise their firms’ profitability over everything else. The reason is that, in a relatively democratic workplace where the firms’ power and benefits are distributed more equitably, employees can be expected to have a greater sense of ownership of their work and solidarity. If, for instance, there is a moderately democratic organisation that is commercially successful mainly because it prioritise innovations above everything else, challenging existing paradigms of what technology should and cannot do, then members of the organisation are likely to share that particular emphasis on innovation rather than safety. In fact, a democratic workplace in which employees unanimously support ethically concerning AI practices, would provide such practices with even greater legitimacy. This is, in my view, the most challenging objection to workplace democracy being the panacea for corporations’ limited incentives to enforce ethical AI practices. After all, organisations known for their flat hierarchy may still actively endorse AI strategies underpinning democratic values. This points out that why accountability burdens on corporations coming from the wider democratic society are crucial. In cases where an organisation is marked by its members’ unanimous support of ethically controversial AI practices, external forces from other democratic citizens must be in place to constrain them.

4. Conclusion

In this article, I have offered a resolution to the boundary problem (BP) as a challenge for the ‘democratising AI' (DAI) agenda. While the BP is less challenging than it first appears, however, many DAI initiatives misleadingly shift our attention away, from the main reason why corporations are demotivated to make ethical decisions around AI – that is, the power disparities within organisations that put employees advancing ethical AI in a vulnerable position – to the inclusion of non-employees in corporations’ processes of the R&D and governance of AI. I argued that workplace democracy provides a more promising normative aspiration than the DAI agenda, although some elements of the latter will complement certain problems that workplace democracy is unable to address, for example, when members of tech firms unanimously prioritise ethically concerning AI practices.