Introduction

AI has become pervasive in our society and central to business models (Haenlein & Kaplan, 2019; Martin, 2022). However, recent scandals, such as crashing autopilots or AI facial recognition systems that underperformed with dark skins, have caused a widespread technological backlash (Blauth et al., 2022; Floridi, 2019; Flowerman, 2023; Hagendorff, 2020; Hao, 2019; McGuffie & Newhouse, 2020; Obermeyer et al., 2019; Truby, 2020; Yeung et al., 2019; Zimmer, 2010). This “techlash” (Smith, 2018) and the possibilities of abuse or misuse made the public question the AI sector’s legitimacy, doubt its morality (Bryson, 2020; Martin & Waldman, 2022; Truby, 2020; Wischmeyer & Rademacher, 2020), and call for an urgent and fundamental review of its ethical principles. The industry has met such a request with self-regulation in the form of new guidelines, lists of ethical principles, and the establishment of AI ethics boards to oversee its development and use (Hao, 2019). Governments responded by advancing national strategies (Casiraghi, 2023; de Almeida et al., 2021; Smuha, 2019), while the European Commission established a high-expert group on AI (HLEG AI) to delineate an innovative human-centred “trustworthy AI” strategy.

Consequently, in this fast-paced, self-regulatory, multi-stakeholder “ethification” (van Dijk et al., 2021) of the AI sector, ethics boards are becoming fundamental entities (Blackman, 2022). However, current set-ups are far from optimal, efforts far from sufficient, and ethics boards far from flawless: recent mappings of the AI ethics guidelines’ development (Jobin et al., 2019) praised these efforts as the first step in the right direction (Floridi, 2019; Leufer & Hidvegi, 2019). Nevertheless, they also underlined how this process is creating spaces and opportunities for lobbying (Goujard, 2022; Schyns, 2023), “ethicswashing” (Metzinger, 2019), “machinewashing” (Seele & Schultz, 2022), and fostering an overall trivialisation of the ethical debate (Bietti, 2020) that is transforming AI ethics into a mere “public relations campaign [aimed at] creating the surface illusion of positive change without the veritable reality” (Obradovich et al., 2019). Among the most cited examples are the criticisms addressed against the HLEG AI, which has been accused of having been lobbied by the sector’s giants and steered (because of its unbalanced composition) to water down principles and guidelines towards a de- or non-regulation of the sector; noteworthy, in this case, are the sector’s efforts to “exclude the newly introduced concept of ‘general purpose’ AI systems from regulations” (Schyns, 2023).

As such, the sector’s legitimacy is in jeopardy, and reform is always more indispensable. Nevertheless, literature contributions to correct this dysfunctional system seem to have focused solely on temporary and cosmetic fixes that do not solve the systemic loss of public trust. These range from increasing representation via “hand-selecting” (Stix, 2021) experts outside public calls and holding public consultations to represent specific interests or incorporating external consultants with cross-sectoral expertise in the decision-making process (Stix, 2021; Yeung et al., 2019). We deem such stratagems insufficient for two reasons. Firstly, because they are built on the outdated governance model of the “wise king” where “ a ‘few’ experts tak[e] decisions in a top-down way” (Helbing, 2021) instead of a legitimate, genuine, and democratic “socio-political deliberation and consensus” (Floridi, 2019; Noveck, 2017). Secondly, because current approaches failed at conceiving and adapting the “governance mechanisms” and implementing the “proactive measures […] [to] enhance (cyber)resilience” (Blauth et al., 2022) necessary for securing a fair and acceptable development of AI—i.e., the goals pursued by most AI ethics boards (SAP AI Ethics Steering Committee, 2021; Sony Group, 2018).

To counter this insufficiency, we offer an innovative, non-trivialised moral reasoning that “front-load[s] ethics” and fully and genuinely considers “human values, norms, [and] moral consideration” (van der Hoven, 2007) by harnessing the collective intelligence of cognitive diverse groups (Bonabeau, 2009; Landemore, 2013). We propose a model that could undermine the structural issues plaguing the sector, namely: (i) the industry’s rampant (covert) lobbying to impose its deregulation and its over-influence in establishing ethical guidelines; (ii) the instrumentalisation of ethics boards to later ‘shop’ for principles and values; (iii) the corporations’ manipulative behaviour within institutional ethics boards; (iv) and the system’s secretive and untransparent negotiation and decision-making process (Bank et al., 2021; Floridi, 2019; Hagendorff, 2020; Peukert & Kloker, 2020; Schyns, 2023; Whittlestone et al., 2019). Our proposal consists of a two-step model that transposes the documented benefits of the ancient practice of sortition into the selection of AI ethics boards’ members and combines them with the advantages of a truly pluralistic, participative, transparent, and deliberative bottom-up process typical of Citizens’ Assemblies (Courant, 2021): qualified informed lotteries.

Finally, our model’s pluralistic approach—epitomised in the collective intelligence of the randomly selected members—may also provide the sought-after sound base for constructing a decentralised, bottom-up, participative digital democracy and a “new social contract” (Helbing, 2017; Helbing et al., 2019; Prins et al., 2017).

To this goal, first, we frame sortition within legitimacy theory and as the only option available among the three canonical strategies to regain the public’s trust. Second, willing to fill the operationalisation gap that emerges from the literature on how to implement a moral reasoning strategy, we introduce the ancient practice of sortition and then present our innovative way to apply it to choosing ethics board members: qualified informed lotteries, underlying how it may channel the AI sector—and, more generally, digital societies—away from dystopias.

The operational framework gap

Legitimacy amounts to society’s acceptance of a business’s activities and behaviour (Ashforth & Gibbs, 1990). Such abidance is paramount for the business to obtain or maintain resources and gain constituencies’ support (Scherer, Palazzo, et al., 2013a) and may be undermined in cases where society considers the business’ actions inappropriate to the context or undesirable. We reckon that this is the case with the AI sector since the recent scandals and accusations of lobbying (Bank et al., 2021; Goujard, 2022; Schyns, 2023), “ethicswashing” (Metzinger, 2019), ethics “trivialisation” (Bietti, 2020) or “machinewashing” (Seele & Schultz, 2022) have profoundly questioned the public’s perception of the industry, its behaviours, and its proposed solutions (for example, the institution of ethics boards to oversee the development of AI and related products and services) (Martin & Waldman, 2022). Such mistrust, however, has not remained confined to the private sector but has spilt over to the intergovernmental level with accusations against the European High-Level Expert Group on Artificial Intelligence (HLEG AI) of an institutionalisation of ethicswashing (Metzinger, 2019). Critics condemn the power imbalances present in the group due to an imbalanced composition, which—paralleled to external lobbying efforts—led to the promotion of framing principles favouring the industry with possible downfalls on the citizens’ fundamental rights (Muller & Dignum, 2022). These elements suggest a systemic loss of legitimacy towards decision-making bodies, such as AI ethics boards (institutional and corporate), and their processes and deliverables.

In cases where the public’s trust is jeopardised or lost, legitimacy theorists provide three strategies to secure or regain it: strategic manipulation strategy, isomorphic adaptation strategy, and moral reasoning strategy (Scherer et al., 2013a). A strategic manipulation strategy involves a business organisation actively and willingly manipulating society’s expectations or the policy-makers’ perception of (Barley, 2010; Child & Rodrigues, 2011; Oliver, 1991; Scherer et al., 2013a). The isomorphic adaptation strategy has the entity, whose legitimacy is at risk, change its practices and organisational structure to adapt to society’s expectations (Deephouse, 1996). Finally, the moral reasoning strategy proposes that legitimacy might be secured through a deliberative process where the organisation engages with the relevant stakeholders in a reasoned debate to find a new equilibrium, thus re-establishing legitimacy (Palazzo & Scherer, 2006). Unlike the other strategies, moral reasoning shares the locus of control between the organisation and society (Scherer et al., 2013a).

According to this tripartition, the tech industry has been deploying a strategic strategy. They have put forward a political strategy by “active[ly] attept[ing] to alter the content of institutional requirements and to influence their promoters” (Pache & Santos, 2021), and a decoupling one (Boxenbaum & Jonsson, 2008) by promoting an ethical-compliance image of the sector while avoiding stricter regulations and continuing with business-as-usual (Pieters, 2011). These were carried out through typical methodologies (Barley, 2010; Fombrun, 2005; Oliver, 1991), such as (covert) lobbying (Bank et al., 2021; Goujard, 2022) and exploitation of the public legitimacy of European expert groups (HLEG AI) and their public relations power. However, since recent “techlash[es]” (Smith, 2018) and the following accusations of ethicswashing (Metzinger, 2019)—and related ones—this method is no longer viable. And, given that isomorphic strategies have never been a realistic solution—because of the high cost of organisational change entailed in adapting to stakeholders’ heterogeneous interests (Scherer et al., 2013b)—we reckon that the moral reasoning strategy is the only practicable solution remaining (Fig. 1). Nevertheless, despite general statements on the sharing of the locus of control, to our knowledge, the literature lacks any reference to the operationalisation of this dialectical exchange. We aim to fill such shortcomings by proposing an operational reform that counters the sector’s structural problems and aligns with participative bottom-up, transparent governance required by digital democracies.

Fig. 1
figure 1

Process Flow Diagram of organisations’ strategy responses to the techlash

Full size image

Borrowing an ancient solution: the historical case for sortition

To this goal, we propose to make use of the ancient practice of sortation—today relegated mostly to jury selections, statistical surveys, or random controls at borders, but once used worldwide until the French Revolution (Frey et al., 2022). Sortition consists of assigning posts by selecting individuals from a larger population through aleatory procedures such as lotteries or draws (Frey et al., 2022). As such, it is the application of random choice on the mode of selection of decision-makers and not on the content to be chosen—as the latter option would be considered by many as irrational or even arbitrary (Frey, 2020).

The first archaeological pieces of evidence suggest that the Sumerians already used random selection processes in the third millennium BC and that it became a governmental practice with the Assyrians in the eighth century BC (Buchstein, 2020). Religious texts, such as the Old Testament and Solomon’s proverbs, confirm such findings and further support this practice by describing its use in assigning priests’ positions and pointing to its various benefits (Buchstein, 2020). Interestingly, this aleatoric selection process has never been peculiar to a specific geographical area but has been used worldwide. For example, in the Tamil Nadu region (India) during the Chola period, the names of the political candidates were written on palm leaves, then extracted by a kid, while in 1100 BC, the North American confederation of the Haudenosaunee employed a sortition-based election and a series of discussion moments called caucus—with the latter still used in today’s American elections—to equally represent the different clans (Bridle, 2022).

Nevertheless, its lay use dates to the fifth and fourth century BC in Athens, where Athenians used to assign civil posts and filled juries or councils by random selection. Described as a “significant political decision-making device” (Duxbury, 1999), sortition has been praised by Aristotle in his The Politics [1303a15] for fostering equality, avoiding corruption, and constraining the consolidation of power in the hands of few (Buchstein, 2015)—features that promoted its diffusion in the heterogenous political environment of the Middle Ages.

Because of its features that weakened the conditions for bribing, sortition systems became common models to halt costly political feuds among stocks (Duxbury, 1999). As such, sortition systems were in use by the Venetians, who used a system of wooden balls called ‘ballotte’ (hence, the word “ballot”) to elect the doge (Molinari, 2018), by the Florentines, who used the “tratta”, viz., a draw to allocate offices that decreased factionalism (Caserta et al., 2021), but also in other city-states throughout Northern Italy (Reybrouck, 2019) and in numerous Spanish cities of the Murcia region (Bridle, 2022).

Although some thinkers, among which Montesquieu, Rousseau and Bentham, still defended its use (Buchstein, 2020), the choosing by lot fell into oblivion with the advent of the Enlightenment since it was seen as “irrational and illegitimate” (Frey et al., 2022) that outsourced man’s agency to chance and ignored qualities (Buchstein, 2020). As such, sortition systems were abandoned all over Europe—with the only exception of the University of Basel (Rost & Doehne, 2020).

The practice resurfaced in recent years with numerous Citizens’ Assemblies. Launched in Canada in 2004, these randomly-selected assemblies were tasked with discussing critical social and ethical issues and ever since have been introduced to other countries such as the Netherlands (Fournier et al., 2011), Australia (Carson, 2013), Belgium, the United Kingdom (Flinders et al., 2016), Iceland, and Ireland (Courant, 2021). The most famous one among these is the “Irish Model” (Courant, 2021; Gastil & Wright, 2019; Harris, 2019; Reuchamps & Suiter, 2016; Setälä, 2017).

In 2016, Ireland attempted to address its most thorny issues through an assembly of randomly sorted citizens, the Irish Citizens’ Assembly. The Assembly was tasked to deal with five divisive “open” problems, i.e., issues which presented “no clear-cut solutions” but required “new ideas”, and “closed problems”, namely, matters that require “the search for a compromise between several known solutions, but incompatible and antagonistic” (Dienel, 2010). The contentious themes were: abortion, the ageing population, climate change, referendums’ rules, and parliament’s fixed term—with the first one being, by far, the most divisive one (Courant, 2021) because of Ireland’s deeply Catholic history.

The Assembly consisted of ninety-nine randomly selected citizens, strangers to one another, screened according to criteria of gender, age, location, and social class. To provide their insight on the first of the five contentious matters, its members met several times over the months in a castle on the outskirts of Dublin. Here, they received a state-of-the-art overview of the issue through experts’ lectures, advocacy group representatives’ presentations and individual testimonies, all notions that could be later elaborated through Q&A sessions, roundtables facilitated by professionals (Courant, 2021), and plenary discussion moments. Noteworthy is that all sessions were broadcasted to increase transparency over the deliberative process and boost the general public inclusion (Bridle, 2022). These debates were formalised via a final secret vote: a majority of sixty-four per cent expressed support for legalising abortion. Such a result was later worded into recommendations for the parliament, which, after careful examination, accepted to hold a general referendum. The results paralleled those obtained by the Irish Citizens Assembly, with sixty-six point four per cent of Irish voters—and an exceptional turnout of sixty-five per cent (Elkink et al., 2020)—expressed a clear favour to enforcing a right to abortion. This right was turned into legislation 2 years later (2018).

Ever since the resurfacing of sortition as a governance system, also the scope of sortition-selected panels has since been broadened from social issues to tackling technological problems; one example in this area is the Consensus Conferences, viz., a randomly-selected panel established in the 1980s by the Danish Board of Technology (later also introduced in other EU members states and Switzerland) tasked to impartially evaluating contested technologies (Joss & Bellucci, 2002).

We insert this paper in this expansion of applications of sortition advocating that such a system might provide a fair, unbiased, and democratic solution to selecting ethics board members to achieve a genuine moral debate on AI. We aim at transposing sortition’s known benefits to make the whole process more transparent, thus avoiding lobbying and ethics-related accusations that now jeopardise the sector.

Implementing sortition

We propose using an aleatoric process for choosing ethics board members within institutional and private ethics boards through qualified informed lotteries: an innovative two-step model (Fig. 2) in line with the concept of digital democracy (Hague & Loader, 1999) that draws on sortition literature but also on the strengths of successful case studies, such as the Irish Citizens’ Assembly (Courant, 2021).

Fig. 2
figure 2

Process Flow Diagram of Qualified Informed Lotteries

Full size image

Step 1

The first phase is a modified version of qualified lotteries proposed by Frey et al. (2022), where panellists are first preselected and then randomly chosen. We divided such a process into two sub-steps (Fig. 2).

Step 1.1: preselection

First (step 1.1), a canonical preselection based on conventional desirable criteria is carried out before drafting the panellists according to the board’s size. This qualification control increases the process’s legitimacy (Suchman, 1995) since it tames the feeling of a process that is inconsiderate of “any reasoning, merit, or human will” (Frey, 2020). However, unlike Frey et al. (2022), we suggest keeping the desirable criteria as broad and basic as possible and omitting expertise as a selection criterion. This choice aims, first, to avoid the introduction of criteria that might discriminate the entrance into the decision-making process to most of civil society and, second, to minimise any possibility of manipulation by the industry of the initial poll from which the panellists are randomly selected (this is because the stricter the criteria, the likelier it is to bias the panel’s composition towards one’s interests). We suggest using as a desirable criterion—and consequently discriminatory parameter—a stakeholder approach. Accordingly, “any group or individual that can affect or be affected by the realisation of an organisation’s purpose” (Freeman et al., 2010) should be eligible to be part of the panel. As such, for corporate’s AI ethics boards, the stakeholders would be identified through canonical primary and secondary stakeholder theory and would comprise: customers, employees, financiers, communities, suppliers, government, competitors, consumer advocacy groups, special interest groups, media (Freeman et al., 2010). The same process could be applied to institutional AI ethic boards, such as the European HLEG AI.

Step 1.2: sortition

The preselection process is followed by the random selection of the ethics board’s panellists (step 1.2). This aleatoric process permits countering one of the underlying causes identified as favouring ethicswashing practices: biased compositions of ethics boards. It has been remarked that in the case of the HLEG AI, for example, the panel’s composition heavily favoured the industry. Algorithmic Watch underlined how the industry “heavily dominate[s]” the HLEG’s composition with twenty-four business representatives (Klöver & Fanta, 2019). Others showed in detail how such influence by examining the composition of the subgroup C of the HLEG AI: the industry had thirteen representatives and five trade and business representatives, while to counteract this force, there were only representatives distributed among academics (of which nine had financial relations with the tech industry (Schyns et al., 2021)), research institute representatives, and think tank members, representatives of NGOs, professional association’s representatives and the representative of a law firm (Nicklas & Dencik, 2020). These data suggest an under-representation of society’s needs and an over-representation of (mainly) US companies’ interests (Hickok, 2021)—a conjecture strongly reinforced by the lack in the final document of a declaration of no conflict of interest (Charlesworth, 2021). This configuration facilitated and promoted the industry’s influence on the decision-making process and its results, and made it possible to disregard civil society’s needs and interests (Bartoletti & Faccioli, 2016; Leufer & Hidvegi, 2019; Schyns, 2023; Stix, 2021; Ulnicane et al., 2020).

Sortition’s correction of mis- or under-representations of specific segments of society stems from the capacity of providing a “descriptive representation” (Pitkin, 1972) of the reference group. To explain: the basic functioning of this property is that if in an initial community of reference R, there is a certain percentage P of people that have a feature F (e.g., a specific characteristic, belief or interest), then, by randomly selecting a sufficiently large subgroup S from the initial community, there will be P members of S that possess F. Thus, S resembles R (Stone, 2009). As such, all people would be represented since S would be “an exact portrait, in miniature, of the people at large” and, as such, it could “think, feel, reason and act like them” (J. Adams, 1851). Because of this fair representation (no reasonable claim can be raised against it), the members of R are likely to identify with the randomly selected panellists (viz., S) and thus ascribe legitimacy to the decision-making process (Fishkin, 2018; Hardin, 2009; Parker, 2011). Moreover, the descriptive representation permits having F (which could be, for example, ‘promoting a deregulation of the AI market’ or ‘advocating for the development of certain types of AI’ defended by the industry’s lobbyists or ‘promoting an ethical human-centred development of AI’ advocated by society) defended in the decision-making panel only in as much as it is present in the population’s interests. Such an approach should heavily curb the current industry’s lobbying.

To sum up, the random selection excludes a priori any accusations of biased compositions of AI ethics boards, decreases possibilities for lobbying, increases legitimacy while promoting a fair, equal and diverse representation of interests in line with the principles of (digital) democratic societies (Hague & Loader, 1999; Helbing, 2021). Therefore, this sortition-based approach circumvents the complex compensatory measures proposed insofar to retrofit the current systemic failures (Stix, 2021; Yeung et al., 2019).

Step 2

This additional step to the model of Frey et al. (2022) is borrowed from the successful case of the Irish Citizens’ Assembly (Courant, 2021) and is fundamental to making the decision-making process deliberative and in line with a “democratisation” of technologies (Ienca, 2019). In fact, the power of a cognitively diverse group (Bonabeau, 2009; Landemore, 2013) is harnessed to correct the omission of expertise from the preselection criteria, while the discussion moments foster civil society inclusion in the decision-making process to obtain a genuine “interactive rationality” (Benhabib, 2004). As for the previous step, the second one consists of two sub-parts (Fig. 2): information and discussion.

Step 2.1: Information

First (step 2.1), all randomly selected panellists must follow meetings and lectures by all relevant stakeholders to receive a full-fledged, science- and data-based, multi-perspective overview to develop an informed view of the topic. Such informative moments are to correct the omission of expertise from the desirable criteria because if it were to be set as one, as in canonical qualified lotteries (Frey et al., 2022), then most people would be excluded from the decision-making process since they would lack the required skills to pass the preselection (Ahlstrom-Vij, 2012). Consequently, the panel would be composed of AI experts or industry representatives, seriously undermining civil participation—thus causing the resurfacing of criticisms of underrepresentation and unbalanced composition. However, by omitting expertise (as suggested in Step 1.1) and providing an issue-specific education after the panel is randomly selected, the panel would benefit both from having a descriptive representation of interests and having informed members.

Although the board members are instructed on the different specifics of the issue, it could be pointed out that, given its complexity, expertise remains necessary to take a reasoned stance. We concede that it might be the case that a board member may not have, even after being properly informed, all the competencies and specific knowledge to fully understand the problem under examination; however, the ignorant member is part of a group where the notion of cognitive diversity applies. According to theories of “multiple intelligence” (Gardner, 2011; Salovey & Meyer, 1998; Sternberg, 1985), we approach the world differently and are “equipped with different toolboxes” (Landemore, 2013). This plurality of approaches, known as cognitive diversity, transposes into various ways of tackling the same issue (Landemore, 2013; Page, 2007), which are of fundamental importance for complex problem-solving or decision-making since the more people with different cognitive toolboxes participate in the process, the larger the spectrum of possible solutions considered: the higher the cognitive diversity, the more innovative the solutions. Hence, increasing the participation of different stakeholders could bring better results than the expert-based panel since, as the “Diversity Trumps Ability Theorem” points out, a large enough group has a cognitive diversity that can compensate for its members’ specific lack of competencies (Page, 2007). In simpler words: “when complex problems must be solved, diversity wins, not the best” (Helbing, 2021). As such, the sortition selection process permits reaching the highest cognitive diversity since it descriptively represents the points of view and moral perspectives present in the larger group.

In conclusion, although expertise might appear as a sine qua non condition for decision-making processes and the presence of experts necessary, cognitive diversity theory suggests that it does not play a decisive role if the selected poll is sufficiently large to compensate for the single members’ deficiencies. This provides a prima facie good argument in favour of our two-step model and its postponed education.

However, if what we have argued insofar is correct, it could still be argued that the increase in diversity for achieving the greatest cognitive diversity necessary for compensating for the lack of expertise would turn out to be practically impossible and extremely costly: the decision-making panel cannot go over an optimum where cognitive diversity and size are maximised without the blowback of size issues. We concede that there is a threshold over which the sortition system becomes impractical, and the information costs become too high, thus hampering the procedure (van der Hoven, 2005). Still, this unfeasibility claim is why a sortition system should be favoured over competing systems since it all comes down to an allocation problem. Modern deliberative democracies tried to solve this issue through a representational system; however, this approach has been demonstrated to promote specific groups—thus homogenising the representatives’ cognitive diversity and, consequently, decreasing the group’s ability to find innovative solutions (Mueller et al., 2011; O’Leary, 2006). On the other hand, sortition can, as the representational system, overcome the size issue by imposing a cap while still maximising the group’s cognitive diversity thanks to the descriptive representation feature. Therefore, if only a certain number of places are available, a sortition approach would provide, ceteris paribus, a higher cognitive diversity—thus “hold[ing] the promise of an important epistemic improvement for the quality of deliberation among representatives.” (Landemore, 2013)—while still working within certain functional boundaries. This counterargument should suffice to silence any claims on the necessity of expertise.

Step 2.2: Discussion

The discussion phase (step 2.2) consists of a series of small groups followed by plenary consultation moments—which should be accessible to the whole civil society to further increase transparency—where the heterogeneity of perspectives is accommodated into a “rational consensus” (Habermas, 1998; Rawls, 1999). Making small independent groups discuss internally before plenary sessions fosters the experimentation of diverse methodologies and further increases the general epistemological diversity that would have been undermined had “there [been] too much communication at the beginning of this process” (Helbing, 2021).

The discussion process will inevitably highlight divergences and uncover disagreements (Cohen, 1986), but also have the panellists consider others’ perspectives and have a chance to change theirs. This increases the overall epistemic quality of the deliverables and solutions envisaged. Similarly to current proposals (Neufeld et al., 2022), we suggest using experts to supervise the (possibly) tortuous discussion and provide the necessary legal and ethical constraints to frame the solution envisaged by the panel. Such experts do not have the role of taking a stance in the debate but guiding it: they must direct the group in a “pro-social logic” (Suchman, 1995). We suggest using ethicists as external discussion guides experts. As a common denominator for what makes an ethicist without getting into the moral tribes of different academic camps and disciplines, we suggest understanding critical thinking as a common value and approach (Seele, 2018). Critical thinking—at best guided by ethical theory, allows professional ethicists to have a greater “awareness of more effects upon a greater variety of situations and contexts” (Taylor, 1944) and know how similar challenging issues have been considered by previous philosophers; accordingly, they are “specialists in the troublesome process of solution and adjustment of social conflicts” (Hoor, 1947). In this new role, the ethicist cooperates with the civil society to translate the moral wisdom of the panel into practice (Hoor, 1947), thus performing a “creative function” that “consists in choosing and clarifying for social use, those alternative ideas which have the best chance of remaining consistent with present and future experience, and with the theories which may be drawn from them.” (Taylor, 1944). As such, AI ethics experts are maintained in the loop—although not in their function of producing the deliverables but in guiding the democratic process of decision-making within the ethics board.

In conclusion, the second step silences the accusations of lobbying by meaningfully including and actively engaging civil society in decision-making—thus increasing the process’s acceptability and legitimacy.

Output: deliverable

The deliberative process would end when an “informed compromise” (Scherer et al., 2013a, b) in the form of a set of statements representing the new consensus between society and the AI sector would be agreed. We suggest that the agreed points and the drafting of the document are left to the ethicist in charge of guiding the discussion as, for the reasons above-mentioned, this expert should have the capacity to coalesce a heterogeneity of perspectives into a cohesive one that still highlights the. We further suggest that if there is no unanimity on the issue’s solution or its reasons, the document should feature a ‘dissenting opinion’ or, respectively, a ‘concurring opinion’ section in the tradition of a ‘minority report’ to live up to one of the criteria of open democracies, minority protection. The former would highlight an opposing point of view, while the latter—which agrees on the substance but not on the rationale—would present diverging or additional argumentations (Wex Definitions Team, 2021, 2022). Although such outlying positions might be seen as undermining the epistemic power of the decision taken (it would not be seen anymore as “the only conceivable solution and way of argumentation” (Azizi, 2011)), these sections have several positive effects. First, they better the epistemic quality by incorporating additional elements (Galbraith, 1973) and requiring the advocates of the majority opinion to better explain the rationale behind the choice taken (Dooley, 1999). Second, they increase the transparency—an essential element in a democratic society—of the decision-making process by making explicit dissenting views or highlighting concurrent arguments and the internal dynamics of the decision-making panel (Azizi, 2011). Third, they decrease the risky phenomenon of groupthinking by retaining diversity even in the final document (Janis & Mann, 1979).

The output could have the form of a set of deliverables consisting of recommendations that, in the case of institutional ethics boards, would be handed to the competent legislative body (as it was done in the Irish case); while for private organisations, they would hand to the CEO and the company board.

Such recommendations would likely be much more legitimate and acceptable (Table 1) than the ones currently proposed through the expert system—since they result from a genuine, transparent, participative, deliberative process. Accordingly, the final deliverable would most likely be a document of the people, by the people, for the people: a starting point—exempted from accusations (or less likely to be criticised) of lobbying (Bank et al., 2021; Goujard, 2022; Schyns, 2023), ethicswashing (Metzinger, 2019), machinewashing (Seele & Schultz, 2022), and ethics bashing (Bietti, 2020). We reckon that the final output produced by the model would represent a more unbiased input for legislators than current ones to discuss the development and constraints of AI.

Table 1 Benefits per step
Full size table

Conclusions

The recent “techlash” (Smith, 2018) has underlined how “it is time the debate about regulating technology reach[es] a more sophisticated and substantial level” (Schaake, 2021) that goes beyond the mere façade set up by the industry that pretends to call for or introduce more regulation while simultaneously lobbying to avoid such stricter regulations (Bank et al., 2021; Goujard, 2022; Schyns, 2023). This urgency is also dictated by the fact that the mistrust towards the alleged solutions to frame the development of AI, i.e., the establishment of ad hoc ethics boards, has spilt over from the private sector to the public sphere with serious accusations to institutional bodies, such as the HLEG AI, of being lobbied by the sector (Bank et al., 2021; Schyns, 2023) and institutionalising ethicswashing (Metzinger, 2019). As such, the legitimacy of AI ethics boards hangs on the balance, and with them, that of the entire sector. A new strategy, different from the current top-down decision-making panels of experts, is necessary to counter this instrumentalisation (de Almeida et al., 2021). Among the various possible approaches proposed, the only viable one appears to be the moral reasoning strategy, i.e., a strategy where the industry shares the locus of control with society to deliberatively find a new comprise and equilibrium that can re-establish the public’s trust (Noveck, 2017). However, no detailed operationalisation has been put forward. In this paper, we proposed just such: a way to fill this gap by deploying qualified informed lotteries, i.e., an innovative two-step model that permits transposing the numerous advantages that made sortition systems so prolific in the past to choosing ethics board members.

Qualified informed lotteries secure several benefits (Table 1). The first step’s preselection criteria seriously curbs lobbying possibilities by increasing public participation through a bottom-up decision-making, while the sortition represents interests descriptively. As such, claims of mis- or under-representation are silenced, and the major possibility for the industry to steer the decision-making process is neutralised. Moreover, sortition’s produced high cognitive diversity makes it an important driver for innovative solutions to the complex issue of framing the development of AI.

The second stage corrects the preselection criteria’s broadness and their omission of expertise as a prerequiste. Such a choice is made to boost civil society inclusion and promote a bottom-up approach to counter the current top-down control that facilitates industry’s manipulations. The correction, obtained through a full-fledged, science-based education process in which all relevant stakeholders present their points of view to the panellists, permits curbing the criticism of the necessity of experts to make reasoned decisions and broadens civil society participation. This promotes high and necessary levels of sociodiversity, self-determination, and participation that foster the creation of innovative solutions (Leighninger, 2011; Prins et al., 2017). Finally, the panellists, who descriptively represent the wider society’s interests and are properly informed on the topic, discuss in small groups and plenary sessions under the scrutiny of the wider society—to increase the transparency of the decision-making process and further contain any possibility of lobbying—and under the guidance of an ethicist. The constructive discussion—fundamentally an “interactive rationality” (Benhabib, 2004) aimed at finding a “rational consensus” (Habermas, 1998) on how to solve best societal issues (Bozdag and Hoven 2015)—permits a genuinely pluralistic, participative, and deliberative process where citizens negotiate as equals free from external forces and undue influences (Cohen, 1989). This process operationalises the “society-in-the-loop” (Rahwan, 2018) idea through a decentralised, bottom-up deliberation that, “in the end, leads to better epistemic justifications, better arguments, and it increases legitimacy and respect towards one other” (Bozdag and Hoven 2015)—while, at the same time, empowering people, supporting “flexibility, local adaptation, diversity, creativity, and exploration” (Helbing, 2021).

In conclusion, our model permits reaching a set of deliverables free from external influences (thanks to the randomness of sortition) with a strong legitimacy and political acceptance (thanks to the high pluralistic civic participation and the use of preselections)—unassailable from claims or accusations of lobbying (Bank et al., 2021; Schyns, 2023), “ethicswashing” (Metzinger, 2019), “machinewashing” (Seele & Schultz, 2022) or “ethics bashing” (Bietti, 2020). Moreover, due to the sharing of the locus of control with society, it represents a possibility for developing the needed “new social contract” (Helbing et al., 2019).

Limitations

The first limitation that emerges from the model is the preselection of desirable criteria that could open the possibility for biases. Since a total exclusion of bias would be only achievable by excluding the preselection altogether and skipping directly to the random selection (a possibility that is untenable because the entire process could become illegitimate and unfair since everyone could participate in it—even those that do not have a stake in it), we maintain that some desirable criteria must be set. We advocate that through our minimal and broad parameters, we minimise (but not exclude entirely) the industry’s possibilities for intromission while securing the civil society’s legitimacy, sociodiversity, and civil society participation necessary for achieving insightful results.

As a second-order limitation, we may also see the risk that the stakeholder orientation and sortition do not guarantee full legitimacy as the pool of candidates may still be biased given the incentives to participate. Therefore, the last limitation is the question of whether to compensate the AI ethics board membership financially. As honorariums would account for the invested time and effort of board members, it would also incentivise some candidates more than others: as such, a meaningful balance has to be found.

Overall, we acknowledge that AI ethics boards are just one piece of the mosaic, and not the panacea, for making digital organisations, cities, and societies more ethical. Nevertheless, we advocate that sortition may help to overcome some of the imbalances created by over-lobbying and smokescreen policies that are now plaguing the sector and undermining its trustworthiness. Despite sortition’s documented benefits and the ones it could have in choosing ethics board members, we recognise the need for a more rigorous and full-scale application of values in organisations, cities, and societies.

Future outlook: building bridges to ESG and reporting guidelines

In the AI and tech sector, a similar trend unfolds in analogy to environmental issues: the growth of corporations is paralleled with a documented diminishing power of nations. As a consequence, corporations become more and more political actors (Wagner & Seele, 2017). Because of it, corporations have been called to adopt governance standards, hence the increasingly more important role of AI ethics boards, and increase transparency (Buchholtz et al., 2008). As such, corporations are adapting their governance structures and communication (Palazzo & Scherer, 2006) by increasing their transparency via the disclosure of information through reporting standards. One of the voluntary and, in many countries, mandatory measures is reporting on non-financial issues, also known as Corporate Social Responsibility (CSR) or the ESG framework (environmental, social, governance). The Global Reporting Initiative (GRI) is the most used reporting guideline among non-financial reporting standards. Although such reports, in the case of environmental standards, do not halt the practice of greenwashing, they certainly help curb it. As such, we suggest incorporating our sortition-based two-step model into such reporting guidelines. This is because it would account for (i) the governance model (Baumann-Pauly & Scherer, 2013)—both because of the required stakeholder engagement dimension (i.e., the process of talking and engaging with stakeholders to understand their point of view towards the company’s business) and the governance disclosure (which permits to the public to assess whether the company is walking the talk and its level of corporate citizenship (Baumann-Pauly & Scherer, 2013); (ii) the legitimacy dimension (namely, the fact that their actions should be subjected to public scrutiny and legitimised by the civil society through a deliberative moral dialogue (Buchholtz et al., 2008)); and (iii) the democratic dimension (viz., the accountability and transparency elements necessary to secure a deliberative democracy (Scherer & Palazzo, 2007)). Therefore, as for greenwashing, this suggested incorporation would not stop practices of lobbying, ethicswashing, ethicsbashing, and machinewashing in toto—but, due to the added transparency, it might help to decrease the possibilities for such deploying them.

Practically, the inclusion of AI ethics boards in private and public organisations may be developed into a reporting indicator analogue to the existence of a stakeholder roundtable in GRI. Next to the existence of an AI ethics supplement to the existing GRI framework, a sub-indicator could also disclose the approach and how the AI ethics board has been composed, e.g., by a stakeholder approach and/or through sortition as proposed here. This could be seen as an institutionalisation process to tackle the over-lobbying and biases reported at the beginning of this article.