1 Introduction

Humans dislike making choices from large sets of options: compared to when choosing from a small choice set, they need more cognitive resources and time to compare the alternatives (Boatwright & Nunes, 2001; Salgano, 2006), feel more regret for the options they did not choose, and are less satisfied with the option they chose (Iyengar et al., 2004; Iyengar & Lepper, 2000). In choosing among large choice sets, today humans can be supported by recommendation or recommender systems (RSs), which filter options on the basis of users’ expected preferences and thus reduce information overload. RSs have become widespread in many aspects of our lives. For instance, RSs suggest to us what to eat, what music to listen to, what online news to read, what movie to watch, what job to apply for, and what person to date (Criado & Such, 2019; Yam & Skorburg, 2021).

The way in which RSs were initially designed was based on content similarity and past behavior of users and prioritized accuracy in predicting one’s choice. Such a traditional design had a main negative effect, the overspecialization problem: RSs suggested items similar to each other and the past choices of users, with the consequence of low user satisfaction. One of the main solutions to overcome this problem was introducing diversity in RSs and thus developing recommendation algorithms that provide the user with diverse items without compromising accuracy (in other words, the diverse items should be likely to be relevant to the user).

In the computer science literature, the justification for implementing diversity in RSs is mainly economic: diversity in RSs increases, first, the user’s satisfaction (Castells et al., 2015; Karakaya & Aytekin, 2018; Kotkov et al., 2016; Symeonidis et al., 2019; Zhao & Lee, 2016; Ziarani & Ravanmehr, 2021), which in turn increases the profit of companies selling items suggested by RSs; second, the profits from long-tail items (i.e., the ones that are less often sold and recommended) (Castells et al., 2015; Kotkov et al., 2016; Symeonidis et al., 2019). Such a justification for implementing diversity is mainly economic, even though customer satisfaction can be interpreted partly as an ethical commitment of a company. Nonetheless, this justification is never connected to further moral principles such as the rights of the user in the digital world, and cannot count as a self-standing moral defense of diversity in RSs. Moreover, justifying diversity in RSs as just a matter of customer preference risks reducing it to only a factor contributing to the economic utility of RSs. Diversity in RSs has been justified by computer scientists also with a technical reason, which in the end has repercussions on user satisfaction: diversity is a solution to the overfitting problem, consisting in the low performance of RSs in recommending new items that were not part of the training data.

From the perspective of the type of item suggested, there is only one RS type that so far was provided with an ethical argument justifying the introduction of diversity: the RSs of information, namely news RSs and RSs suggesting scholarly content, i.e., scientific RSs. The introduction of more diversity in such RSs is proposed to expose users to a larger range of viewpoints. Exposure to different viewpoints, in the case of news recommendation, promotes informed citizenship (Bernstein et al., 2020; Elahi et al., 2022; Heitz et al., 2022; Helberger et al., 2018; High Level Expert Group on Media Freedom and Pluralism, 2013) and in case of scientific RSs, positively affects scientific progress (Monteiro-Krebs et al., 2021; Polonioli, 2021).

Therefore, to my knowledge, the RSs literature lacks a moral justification for the introduction of diversity in RSs or, at most, has provided a local one that applies only to recommendations of information mostly from a societal—not individual—perspective. In this contribution, I provide a two-part moral defense of diversity in all RSs regardless of the type of item suggested. First, I contend that lack of diversity in RSs violates two autonomy rights of the user and has also detrimental effects on the user’s good life. Second, I argue that diversity in RSs enables the latter to respect those two autonomy rights. My moral defense of diversity in RSs focuses on a key stakeholder of RSs, the user, and justifies diversity as an instrumental value of autonomy, namely as a means to support the user’s autonomy.

The current lack of moral defense of diversity in RSs has two detrimental effects on RSs users. First, if diversity is reduced to user satisfaction, it is considered only an empirical issue about how much a user appreciates diversity and it can be overlooked in case other factors are more effective in increasing user satisfaction. Second, RSs involve several stakeholders with conflicting interests, values, and rights (Milano et al., 2021); if the user’s ones are not clearly stated, the user cannot claim their rights, nor can the latter be weighed against those of the other parties involved. This affects the quality of the RSs as the latter cannot structure the interactions among stakeholders in an efficient way (Milano et al., 2021). Therefore, the absence of a moral justification for diversity in RSs brings about a vulnerability in the RSs users that needs to be addressed.

My moral defense of diversity in RSs provides a three-fold contribution to the study of RSs. First, it is a resource for the ethical assessment of RSs from the perspective of the user’s autonomy. Second, it shows that all RSs except the knowledge-based ones can threaten the user’s autonomy and thus need to be ethically assessed, not only the RSs suggesting news or scholarly content. Third, it points out that the worth of diversity cannot be reduced to a subjective preference for x degrees of diversity, rather diversity has a normative force grounded in the user’s agency.

My strategy is as follows: in Section 2, I introduce the taxonomy of RSs, their functioning, and the implementation of diversity in RSs in terms of diversity metrics and user control of diversity. In Section 3, I present the theoretical basis of my moral defense of diversity in RSs, which are two autonomy rights of the user: the rights to an open present and to be treated as an individual. In Section 4, I develop my moral defense. First, I contend that the various types of RSs, once deprived of diversity, differently limit the user’s autonomy and negatively impact their good life. Second, I show how diversity in RSs supports the rights to an open present and to be treated as an individual, which, when taken together and applied to RSs, constitute the right to be an exception to predictions. The latter is the moral principle requiring respecting the users’ capacity to be the authors of their own actions in predictive decision-making, of which recommendation algorithms are part. In Section 5, I answer the main objections against my moral defense of diversity in RSs and its theoretical basis. In section 6, I give my concluding remarks.

2 Diversity in RSs

As the aim of my contribution is to provide a moral justification of diversity in RSs, in this section, first, I shortly treat the technical aspects of RSs. Then, I deal with the implementation of diversity in RS by presenting some diversity metrics relevant to my discussion and the ways in which the RSs providers enable the user to affect the diversity of the recommendations.

2.1 RSs

RSs “are software tools and techniques that provide suggestions for items that are most likely of interest to a particular user” (Ricci et al., 2015, cit., p. 1). A RS predicts the items that the user finds most relevant, on the basis of some information related to them: their preferences, interests, characteristics, groups they are part of, and/or observed behavior. In RSs, relevance is the interest of a user for an item (Vargas & Castells, 2011) and is also indicated as the utility of the item (Silveira et al., 2019). The relevance of recommended items can be measured with classification metrics such as accuracy, precision, and recall (Avazpour et al., 2014; Fayyaz et al., 2020; Kaminskas & Bridge, 2014; Shani & Gunawardana, 2011; Silveira et al., 2019; Vargas & Castells, 2011). RSs relieve the user from the burden of choosing among many options, thus making them save time and cognitive resources and avoid regret due to the unchosen options.

RSs are a subclass of predictive algorithms as their goal is to predict the rating a user gives to an item. Their ability to predict the user’s ratings is usually measured in terms of accuracy-based metrics. Predictive systems are software tools and techniques using machine learning and data mining to make predictions, namely educated guesses about the likelihood of future outcomes, based on historical data (Criado & Such, 2019).

Burke (2002) provided a taxonomy of RSs based on the type of data they use and how they use it. Content-based RSs suggest items that are similar to the ones the user positively rated in the past; collaborative-filtering RSs suggest items that are similar to the ones that users with similar preferences liked (the user is associated with these users because of similarity in preferences with them); demographic RSs suggest items that are positively rated by groups with whom the user is associated because of demographic features such as age or gender; knowledge-based RSs suggest items on the basis of explicit knowledge about items; utility-based RSs suggest items that have the highest utility computed on the basis of a utility function; and hybrid RSs combine the techniques of two or more types of RSs. In section 4, I will contend, first, that content- and utility-based systems deprived of diversity provide the user only with content that is similar to past choices and, for this reason, violate the user’s right to an open present. Second, I will argue that demographic and collaborative-filtering systems deprived of diversity provide the user only with content similar to that which other people with whom the user is associated liked and, for this reason, violate the user’s right to be treated as an individual.

2.2 Diversity Implemented in RSs

As seen, RSs that only achieve accuracy propose homogeneous content that may bore or disappoint the user (overspecialization problem). To avoid such an effect and to increase the selling of niche products, diversity was introduced in the design of RSs.

In general, diversity is understood as “the internal difference within parts of an experience” (Castells et al., 2015, cit., p. 882). In RSs, a diverse recommended item is different from what the user previously experienced (Castells et al., 2015, p. 885). Diversity in RSs is a design principle that implies a comparison between either sets of recommendations of one user or of all users of a RS (Ricci et al., 2015, p. 25). Thus, I conceive diversity in RSs as a property of recommendation. This property is scalar and therefore can be measured, as we will see below.

Diversity as a design principle in RSs requires making available choice options that are different from the content a user previously chose and thus less likely to be chosen by the user according to her past choices or the choices of people with whom the user is associated. In many cases, the different content is also new to the user, in the sense that the user does not know the item suggested and/or its attributes. In RSs, novelty refers to a property of recommendations and is a further design principle of such systems (Ricci et al., 2015, p. 25; Castells et al., 2015, p. 902). In this contribution, I will not deal with the moral defense of novelty in RSs, but since, as just seen, diverse content partly overlaps with novel content, my moral defense of diversity in RSs is also valid for content that is new and diverse. Proposing content that is novel and/or diverse from previous experiences exposes RSs to the risk that the user does not like the content. It is noteworthy that a diverse item can be relevant or irrelevant to a user, and a relevant item can be diverse or similar to the other ones suggested to the user. In fact, consider a RS whose only goal is achieving accuracy by using similarity metrics to make recommendations. Such a RS suggests only items similar to each other and in the long run risks boring the user, with the consequence that the items suggested become less and less relevant. An ideal RS endowed with novelty and diversity would propose relevant content and thus enhance serendipity, namely the discovery of content that the user does not expect (as it differs from their profile), does not know, and likes (serendipitous content).Footnote 1

As I will show in section 4, content- and utility-based systems, on the one hand, and demographic and collaborative-filtering systems, on the other, do not violate the rights respectively to an open present and to be treated as an individual as long as they introduce diversity in the sense of diversity of items from respectively the user’s past or other people’s choices. Therefore, the diversity metrics introducing diversity from the user's previous choices into content- and utility-based systems enable these RSs to respect the right to an open present. Similarly, the diversity metrics introducing diversity from other people’s choices into demographic and collaborative-filtering systems enable these RSs to respect the right to be treated as an individual.

For the aim of this contribution, which is providing a moral defense of diversity in RSs, what is relevant is endowing RS with diversity from the user’s past choices or similar people’s choices, not the way such diversity is introduced. Therefore, I will present examples of the diversity metrics introducing diversity in the relevant sense I indicated. I refer to Avazpour et al. (2014), Castells et al. (2015), Vargas & Castells (2011) for the reviews of the diversity metrics.

One of the most used diversity metrics is the average intra-list distance or intra-list diversity, which is the average pairwise distance of the items of a set (Smyth & McClave, 2001; Vargas & Castells, 2011). In such a metric, distance is usually conceived as the opposite of similarity measures (Avazpour et al., 2014; Ziegler et al., 2005). While intra-list diversity applies to a set of recommended items, inter-recommendation diversity measures refer to all recommended items that a RS provides to a set of users and they are thus also defined as aggregate diversity measures (Adomavicius & Kwon, 2012, 2014; Zhou et al., 2010). An example of such a metric is the inter-user diversity, which measures the difference of recommendation lists of various users (Zhou et al., 2010). A third approach to diversity in RSs consists in measuring diversity in the evolution of RSs over time. Temporal diversity measures the difference between recommended items suggested to the same user at two consecutive time points and thus takes into account the possibility that the users change their preferences and/or their environment changes (Lathia et al., 2010). In utility-based RSs, diversity is introduced by incorporating reinforcement learning that favors the user’s long-term utility in the form of discovery (Ma et al., 2020).

All the above-mentioned metrics enable the introduction of diversity with respect to items chosen in the past and by other people, but some of these metrics are more efficient in achieving such a goal, as they are specifically designed to introduce diversity in these two respects. As measures of aggregate diversity compare the recommendation lists of all users of a RS, they are suitable for increasing diversity of items with respect to other people’s choices in collaborative-filtering and demographic RSs. Although temporal diversity has been assessed only in collaborative-filtering algorithms so far (Lathia et al., 2010), it is suitable for increasing diversity of items with respect to the user’s past choices in content- and utility-based systems.

On the side of the RS user, there have been various ways to implement diversity by increasing user control, which consists in enabling the user to have direct effects on the recommendations (Jannach et al., 2017). There are three phases of the recommendation process in which the user can introduce diversity by exerting user control. In the input phase, the user can specify their preferences and/or constraints (Jannach et al., 2017) or give some rules (a sort of filter) (Tang & Winoto, 2016). In the process phase, the user can adjust the parameters or weights of the recommendation algorithm (Bostandjiev et al., 2012; Harper et al., 2015) or choose the recommendation algorithm (den Bogaert et al., 2022; Harambam et al., 2018). In the output phase, the user can order the items or modify the recommendation list provided (Bostandjiev et al., 2012; , give feedback to the RSs that provide explanations for their recommendations (Lamche et al., 2014), or navigate through connected lists of recommendations (Helberger et al., 2018). The degree of user control depends on how many phases allow user control and the amount of control that the settings of a RS allow. Such conditions depend in turn on the RS provider and thus the user control of a RS has to be assessed case by case.

3 The Theoretical Basis of the Moral Defense of Diversity in RSs

3.1 The Ethics of the Future Self

Self-regarding morality regulates the relationship in which the individual stands with themselves, and the ethics of the future self is a subset of self-regarding morality concerning the relationship between one’s present and future selves. I chose to adopt the perspective of the ethics of the future self for defending diversity in RSs for three reasons. First, RSs are egopoietic technologies or technologies of self-construction (Floridi, 2011), namely technologies that change our self-understanding and construction of our own identities as individuals (Milano et al., 2020; Reviglio, 2019). Accordingly, such technologies need to be regulated by self-regarding morality. Second, our future selves cannot protect themselves from the consequences of using such technologies in the present, as I will show. Third, the negative effects of lack of diversity in RSs on our lives accumulate with time; thus, they mainly affect us in the future, i.e., our future selves.

The relevant entities or objects of analysis of the ethics of the future self are a person’s diachronic selves, namely the past, present, and future selves. I conceive a person’s diachronic selves as practical agents, namely subjects that autonomously take actions for which they can provide reasons, who are situated at a temporal stage of a person and coexist with that person. The motivations providing an agent with reasons for action are personal commitments, plans, values, etc. I call them normative principles of action. One’s set of normative principles of action determines one’s preferences in decision-making.

A diachronic self is not metaphysically distinct from the person they belong to, they can be rather considered a morally relevant attribute of that person. Similarly, a person’s present self is not metaphysically distinct from a person’s future self. I distinguish one’s present and future selves not at the metaphysical level but at the practical level, namely in the sphere of action, because the temporal stage at which a person’s self is located determines several morally salient features of the self and the latter’s relationship with the other diachronic selves of a person.Footnote 2

In the case of RSs, the present self is the user of a RS in the present, and the future self is the same user in a future that is distant from the present time in terms of years. The exact temporal distance between the two selves varies with context and thus cannot be specified once for all cases.

Two moral features of the present-self–future-self relationship are relevant for the application of the ethics of the future self to RSs.Footnote 3 The first is the asymmetry of decisional power: the present self affects the future self with their decisions but the future self cannot affect the present self with theirs. Therefore, one’s earlier selves have more decisional power than one’s later selves and are in a more advantageous position in time than the later selves. For instance, one’s present self can decide to spend their money for a comfortable life and leave the future self with no savings. A clear consequence of the asymmetry of decisional power is that the future self’s normative principles of action (and thus preferences) and even existence are affected by the present self’s decisions. At the same time, the future self’s normative principles of action and existence are influenced by unpredictable future events (e.g., a job opportunity, a change in one’s fundamental values, an accident) and thus not (completely) determined in the present. The second feature of the present-self–future-self relationship is precisely the indeterminacy of the future self’s preferences and existence. When the present self makes a decision with effects on the future self, the former does not know whether the latter will agree with the decision and whether there will be a future self at all.

3.2 The Right to an Open Present

The present and future selves are both practical agents and thus they have the same moral worth and are equally entitled to pursue their normative principles of action. The future self’s reduced decisional power is not morally justified; it is only due to the direction of time and causality. I contend that an agent’s moral worth is not affected by the temporal location of a self in a person’s life: the fact that the future self has yet to come does not reduce their moral worth. As (i) the present and future selves are equally entitled to realize their own normative principles, (ii) their moral worth is not influenced by time, and (iii) the present self’s greater decisional power is not justified, I contend that the future self should have the same freedom to pursue their normative principles as that of the present self. This freedom is granted if the necessary conditions for the pursuit of any set of normative principles are preserved, no matter how the future self’s normative principles will be different from the present self’s.Footnote 4 I cannot provide an exhaustive list of such necessary conditions in this contribution because this would lead my contribution astray. Yet I can list some fundamental ones on which most of the approaches on agency would agree, such as health, adequate education, income, and basic rights.Footnote 5

I call the claim for the preservation of the necessary conditions of action the right to an open present. Feinberg, 1992 defended the children’s right to an open future, which is an intergenerational right aiming to protect, in the present, the autonomy that children will have when they will be adults from their parents’ violation. The violation consists in cutting off fundamental opportunities (i.e., education) that will restrict the autonomy of the children, once they will be autonomous adults. I apply Feinberg’s right to an open future to one’s future self.Footnote 6 I define the future self’s claim for the preservation of the necessary conditions of action as the right to an open present and not future for two reasons. First, the necessary conditions that such a right aims to preserve are present at the moment in which a self takes action. Second, while Feinberg attributes the right to an open future to the metaphysical person in its infancy (the child), I apply the right to an open present to the future self, who is a practical—not metaphysical—identity. that is situated at a temporal stage of a person. The right to an open present preserves the self’s agency, i.e., the capacity to be an agent, as it enables a self to pursue their set of normative principles.

3.3 Principle M and the Right to be Treated as an Individual

A person qua agent is the result of their previous choices with which they determined their life and the principle of their future choices (Anscombe, 1957; Davidson, 1963; Frankfurt, 1971; Mele, 2003; Kant, 2006, 2015; Dworkin, 2017; Eidelson, 2013). Respecting autonomy implies treating a person as the author of their choices and not as an object determined by causal laws and thus in principle predictable, once the causal laws are known (Basu, 2019). Various dimensions make a person different from another one and thus unique (e.g., personality traits, body shape, genetic conditions), but autonomy differentiates a person from the other ones in a way that is morally relevant and demands respect (Dworkin, 1988 p. 110; Eidelson, 2013, p. 211). What makes a person a unique individual in a morally significant sense is the capacity to determine or shape one’s life on the basis of, for instance, personal commitments, plans, values, namely what I defined as one’s normative principles of action.

The principle of treating people as individuals, which I call from now onwards principle M, respects people’s autonomy by taking into account their morally relevant uniqueness and avoiding considering them just as specimens of a group (e.g., female, Muslim) (Viganò et al., 2022). I define principle M as the moral requirement prescribing that in the beliefs we form about a person, we should give reasonable weight to their capacity to autonomously determine their lives and thus be an exception to the predictions about their behavior based on statistical evidence concerning other people’s behavior.

Statistical evidence is a collection of information that helps to prove whether something is true through generalizations, namely inferences drawn from data (Schauer, 2003, p. 172). Principle M implies that, to form a belief about somebody, we should take into account the available information about that specific individual, namely individual evidence. This means that when gathering information about somebody is costly, impossible, time-consuming, or inappropriate due to, for instance, an emergency, principle M can be disregarded or overridden by other moral considerations. Consider, for example, choosing the therapy in an emergency room for a case in which using statistical evidence about the patient’s gender can save their life. Principle M is thus a pro tanto moral principle (Viganò et al., 2022). My characterization of principle M synthetizes Lippert-Rasmussen’s (2011),Footnote 7 Eidelson’s (2013),Footnote 8 and Moss’ (2018)Footnote 9 interpretation of the moral norm of treating people as individuals.

From the perspective of the person that is the object of a belief, principle M is the right to be treated as an individual meant as an autonomous being. Similarly to the right to an open present, this right protects agency.

It may seem that principle M and the right to be treated as an individual are epistemic principles, as they concern statements. However, we relate to other people not only through actions but also through our beliefs about them (Basu, 2019, pp. 922, 928). Therefore, when a statement is about people, it may entail moral encroachment (Moss, 2018, p. 223). The predictions about users’ preference of a RS are subject to the right to be treated as an individual as they are statements about the users.Footnote 10

4 The Moral Defense of Diversity in RSs

RSs filter the content that the user receives on the basis of what is allegedly most interesting to them. As indicated by behavioral economics and psychology (Thaler & Sunstein, 2009), choice architecture—the way a set of options is presented—influences a decision-maker. For this reason, many scholars contended that RSs manipulate the user’s choices (Lanzing, 2019; Milano et al., 2020; Mittelstadt et al., 2016; Sahebi & Formosa, 2022; Specker Sullivan & Reiner, 2019; Tsamados et al., 2022). As RSs cannot avoid presenting choice options, from a moral perspective, they cannot be required to avoid influencing users’ choices; rather, they can be required to do so in a way that avoids potentially harmful choices, respect some moral principles, and/or nudges users toward beneficial choices. At the same time, RSs should provide relevant suggestions to the user—i.e., suggestions the user likes—otherwise they do not fulfill their function of reducing information overload and making the user’s choices easier.

In this section, after sketching the moral defense of diversity in RSs of information that was put forth in the literature, I present my moral defense of diversity in RSs, which is divided into two parts. In the first part, I argue that, except knowledge-based RSs, the rest of RSs that do not include diversity in their design negatively impact on the user’s autonomy and good life. In the second part of my moral defense, I argue that RSs in which diversity is implemented promote the user’s autonomy.

4.1 The Moral Defense of Diversity in RSs of Information in the Literature

The literature on RSs has identified two detrimental effects of RSs of news and scholarly content that prioritize accuracy at the expense of diversity: filter bubbles and echo chambers. A filter bubble is an enclosed media space in which the user is insulated from information that is different from what they are likely to consume (Pariser, 2011), which basically means insulated from experiences that are different from the ones they already had. An echo chamber is an enclosed media space in which only some information circulates and “has the potential to both magnify the messages delivered within it and insulate them from rebuttal” (Jamieson & Cappella, 2008, cit., p. 76). However, so far empirical evidence of filter bubbles and echo chambers is weak, mixed or even absent in some studies (Bruns, 2019; Figà & Arfini, 2022; Ross Arguedas et al., 2022).

In the RSs literature, the defense of diversity in news RSs is founded on two arguments. The first is from the societal perspective and is the contention that diversity is an instrumental value to media pluralism (Bernstein et al., 2020; Elahi et al., 2022; Heitz et al., 2022; Helberger et al., 2018; High Level Expert Group on Media Freedom and Pluralism, 2013). The second is from the individual perspective and is the contention that diversity expands people’s access to information and thus favors their autonomy (Helberger et al., 2018; Milano et al., 2020). Scholars warned that news RSs that do not include diversity reduce the information exposure of the users (see, e.g., Schoenbach, 2007; Zhang & Hurley, 2008; Bozdag & van den Hoven, 2015; Burri, 2016). Helberger et al. (2018) contended that limited information exposure undermines the basis of a democratic society made of informed citizens making well-considered decisions on the basis of information received from diverse sources and viewpoints. Milano et al. (2020) identified six ethical challenges of RSs. In the sixth ethical challenge, which is negative social effects, they propose diversity in RS as a remedy to filter bubbles brought about by news RSs and social media filters. This proposal can be interpreted as a moral defense of diversity in terms of social utility.

In the case of news RSs offered by public service media organizations, diversity is a core value included in the public interest mission of such organizations (Ada Lovelace Institute, 2022). One of the meanings given to diversity in RSs by the European Broadcasting Union is giving voice to a plurality of viewpoints and it is justified with the same argument employed in defending diversity in news recommendations provided by private organizations: “to help build a more inclusive, less fragmented society” (European Broadcasting Union, 2022, cit. p. 5).

In the case of scientific RSs, scholars contended that such RSs risk insulating researchers from exposure to different viewpoints—especially viewpoints proposed by unknown scientists—which has in turn detrimental effects on scientific progress (Monteiro-Krebs et al., 2021; Polonioli, 2021). Among the recommendations for limiting the negative effects of scientific RSs on scientific advances, Polonioli (2021) suggested the application of diversification re-ranking approaches to the outputs of scientific RSs and Monteiro-Krebs et al. (2021) the introduction of serendipity in the design of such RSs.

Reviglio (2019) provided a moral defense of serendipity as a design and ethical principle in RSs but he did not conceive serendipity as a property of recommendations. He rather meant serendipity as a subjective experience of the user depending on environmental and individual capability to accidentally encounter relevant and unexpected information. Such a defense is mainly based on the capacity of serendipity to foster media pluralism and strengthen the individual rights to receive information (protected by the Article 10 of the European Convention on Human Rights) and not to be manipulated by choice architecture.

The main differences between my moral defense of diversity in RSs and the ones proposed in the literature are, first, that I provide a moral justification that is valid regardless of the type of content (news, music, food, exercises, etc.) RSs suggest; second, I conceive diversity as a property of the recommendations and not as an experience of the user; third, my justification is grounded exclusively on the user’s agency.

4.2 The Violation of the Right to an Open Present in Content-Based and Utility-Based Systems Without Diversity

To predict what the user will most probably like, the various types of RSs employ different techniques, as seen in Section 2. Contend-based RSs use the user’s past choices and ratings, collaborative-filtering RSs use the choices and ratings of individuals labeled as similar in tastes to the user, demographic systems use the choices and ratings of demographic groups to which the user belongs, utility-based RSs compute the user’s utility function, and knowledge-based RSs use explicit knowledge from the user.

Apart from knowledge-based RSs, if we deprive the rest of the RSs of diversity, we obtain RSs that make the user’s digital environment homogeneous and immutable. Without diversity, content-based and utility-based systems do not expose the user to content that is new with respect to or different from the user’s past choices. Without diversity, collaborative-filtering and demographic RSs do not expose the user to content that is new with respect to or different from similar people’s choices. Knowledge-based RSs do not create a homogeneous and immutable digital environment if deprived of diversity. However, they are employed only in specific sectors in which the user actively provides information on their preferences or are combined in a hybrid system – and in this case, have the same issue of the rest of the RSs.

Content- and utility-based RSs without diversity do not expose the user to content that is new or different from the user’s past choices. Therefore, they create a digital environment in which the future self’s normative principles tend to be the present self’s ones. Such RSs insulate the user from alternative life experiences and developments of their potential. The cumulative effect of this insulation is that the preferences, plans, and aims (i.e., the normative principles of action) of the user’s future self tend to remain the same as the user’s present ones; I call this phenomenon identity inertia (Loi et al., 2020).

The right to an open present requires preserving the future self’s freedom to pursue their own normative principles. Content- and utility-based RSs without diversity violate the right to an open present as they hinder the possibility that the user finds out new or changes some interests, preferences, values, basically some normative principles of action. By making the user’s choice sets homogeneous and past-biased, such content- and utility-based RSs make the user’s future self predictable (as the latter will be similar to the present self) and narrow the future self’s open present (as the latter is a reproduction of the past). If deprived of diversity, content- and utility-based RSs violate an aspect of the user autonomy in a form that may be interpreted as a past-biased manipulation of the user’s choices: they make the user choose items similar to those chosen in the past since the user is provided only with such a kind of items.

4.3 The Violation of the Right to be Treated as an Individual in Collaborative-Filtering and Demographic RSs Without Diversity

The homogenization of a user’s digital environment is not only caused by content- and utility-based RSs in which diversity is removed. If the latter is eliminated from collaborative-filtering and demographic RSs, such RSs do not expose the user to content that is new or different from the choice of people with whom the user is associated.

Collaborative-filtering and demographic RSs without diversity isolate the users from different lifestyles, namely exposure to different experiences, stimuli, and opportunities. The homogenization of people’s lifestyles in the long run makes the lives of users that are assigned to the same groups progressively resemble each other and conform to a standard that is the most common lifestyle of the users assigned to the same group.

The right to be treated as an individual prescribes to acknowledge that the individual can be an exception to generalizations about their behavior deriving from statistical evidence based on other people’s data. Collaborative-filtering and demographic RSs without diversity predict a user’s preferences only on the basis of their membership to a certain group. They suggest to the user an item because, for instance, people of the same age positively rated it. Yet each individual distinguishes themselves from other people, as each individual determines their life on the basis of their own choices, which are in turn based on their values, commitments, projects, etc. This means that collaborative-filtering and demographic RSs without diversity violate the right to be treated as an individual.

4.4 The Limitation of the User’s Good Life in RSs Without Diversity

A person’s good life is a life that goes well for the person living it (Crisp, 2021). The uniqueness of a person makes their good life unique and different from another person’s good life. At the same time, though, there are components of the good life that are the same for everyone, as they are necessary for the pursuit of any good life. For instance, a person’s autonomy makes it essential to the realization of any good life that the person is not forced to make choices.

I contend that the homogeneous and immutable environment of RSs deprived of diversity progressively reduces the user’s good life because it limits three factors that are essential to any good life, which I call formal conditions of the good life. These factors are life experience, self-knowledge, and authenticity. Notwithstanding the debate on the components of the good life, many theories in psychology and philosophy acknowledge these three formal conditions as contributing to any good life. For instance, in Griffin’s theory, the list of prudential values includes knowing about oneself (which belongs to the value of understanding) and choosing one’s course of life (which belongs to the components of human existence) (Griffin, 1986, pp. 66-67). Ryff’s theory of psychological well-being incorporates personal growth meant as the constant development of one’s potential, of which openness to experience is a key factor (Ryff, 1989b, 1989a). Sumner’s theory comprises self-knowledge conceived as the authentic and autonomous self-assessment of life satisfaction (Sumner, 2011, pp. 156-170). In Nussbaum’s list of fundamental capabilities, the sixth one—practical reason—is the capacity to form a conception of the good and critically reflect on the planning of one’s life (Nussbaum, 2016, pp. 76-78).

The reduction of the three formal conditions of the good life due to a lack of diversity in RSs takes place progressively and its effect on the good life can be seen after a few years: it is the good life of the user’s future self that is reduced. A homogeneous and immutable environment reduces the user’s opportunity to discover their potential, aspects of themselves, new interests, preferences, and values, which in turn limits self-knowledge and life experience. Authenticity is the capacity to live according to ideals that one reflectively accepts. The reduced knowledge of oneself hinders authenticity because if the individual has limited knowledge of what they want, like, and aspire to do, they cannot be authentic to themselves (Anderson, 1991; Mill, 1859). Even in the case the individual wants to focus on one specific experience or activity, the possibility to try out different experiences enables the individual to critically choose to focus on one experience by comparing the latter with other experiences and not just focus on it because it was the only one available. Therefore, a heterogeneous and mutable environment creates the basis for making choices that are more informed, better grounded, and thus more authentic, than the choices that are available in a homogeneous and immutable environment.Footnote 11

4.5 The Right to be an Exception to Predictions in RSs with Diversity

Content- and utility-based RSs without diversity predict the user’s behavior on the basis of their past choices; collaborative-filtering and demographic RSs deprived of diversity predict the user’s behavior on the basis of a group’s choices. Let us now ethically examine RSs endowed with diversity.

Content- and utility-based RSs which include diversity in their design provide the user with new and diverse content that expands the user’s horizon in terms of heterogeneity of choices and thus enables the user to explore interests, preferences, and contents that in RSs without diversity would never be available and thus the user could not pursue. Content- and utility-based RSs with diversity create a mutable and heterogeneous digital environment in which the users are not treated as their past selves. Therefore, on the one hand, they respect the right to an open present, which requires preserving the future self’s freedom to pursue their own normative principles. On the other, as such RSs expand the user’s heterogeneity of choices, they favor the user’s autonomy in terms of development of oneself by trying out new experiences and exploring one’s potential.

Collaborative-filtering and demographic RSs which include diversity introduce alternatives that do not derive from the choices of the group with which the user is associated. In this way, they avoid treating the individual as a specimen of a group; they rather treat the individual as an autonomous being that can choose alternatives that are different from those chosen by other users in the same group.

In sum, content-, utility-based, collaborative-filtering, and demographic RSs with diversity appropriately acknowledge the morally salient fact that a RS user is the principle of their actions, and thus their future choices cannot be just inferred from their previous ones, nor from the choices of people associated with them.

In RSs, the right to an open present forbids that the user is treated like their past self; the right to be treated as an individual forbids that the user is treated as a specimen of a class. I call the combination of the rights to an open present and to be treated as an individual in RSs as the right to be an exception to predictions. The right to be an exception to predictions is a pro tanto principle regulating how RSs should treat the users, where the regulation of the user’s treatment includes expressing appropriate statements about the user. The right to be an exception to predictions states that RSs should not provide the users with recommendations based only on their past choices or a group’s choices. Since a user is an autonomous agent, who is thus not entirely determined by their past (as the user is the principle of their future actions), nor by the group they are associated with (as the user is free in their choices),Footnote 12 the appropriate way of treating the user should include the awareness that they are at least partly non-predictable.

Of course, the right to be an exception to predictions does not apply to the RSs in which the recommendation is based on an aspect of the user on which they cannot exercise autonomy. An example of such a case is an algorithm recommending a therapy on the basis of a patient’s genetic predisposition to a certain disease (the individual cannot choose to have this genetic predisposition).Footnote 13

5 Main Objections to the Moral Defense of Diversity in RSs and Its Theoretical Basis

First, I deal with the objections targeting my moral defense of diversity in RS. A main theoretical objection in this regard is the alleged inconsistency that consists in defending diversity as a principle protecting users’ autonomy while its implementation in RSs reduces their autonomy. This is because introducing novel and different content into the user’s recommendations may be considered a form of “stealthy paternalism” (Reviglio, 2019, p. 160), as it would change the user’s choice set and thus influence their choices. Yet, as seen, choice architecture is unavoidable. Considering the impossibility of presenting a neutral choice set, it would be better to present one that protects the user’s agency, and respecting the right to be an exception to predictions enables the protection of the user’s agency. Moreover, I contended that the introduction of new and diverse options can enhance the user’s autonomy as the user is given the opportunity to know aspects of themselves and make different experiences.

In contrast with the first objection, one may say that introducing novel and diverse content does not guarantee that the user will choose that (Reviglio, 2019, p. 156). Such an objection targets the second part of my moral defense of diversity in RSs and consists in contending that RSs with diversity support the user’s autonomy and good life if and only if the user selects novel and different items. Thus, according to this objection, diversity in RSs per se does not contribute to the user’s autonomy and good life. I agree that supplying diversity does not equate with experiencing it. At the same time, I think that supplying diversity is the premise of providing the user with the formal conditions of a good life in the digital environment and guaranteeing that their present remains open and their lifestyle is not determined only by their group’s lifestyle. Even in a non-digital environment, the individual may decide not to take some novel opportunities or try diverse experiences. RSs cannot force users to choose novelty and diversity, but they are morally required to provide the users with the possibility to do so if the users want to.

Golden & Danks (2021) argued that whether a RS has a moral obligation to show novel content is partly an empirical question depending on the user’s appreciation of novelty. They analyzed the content platform TikTok and concluded that the latter is morally obliged to provide novel content because TikTok users value novelty and the company is engaged in the users’ interests. Accordingly, it may be objected that my moral defense of diversity in RSs should also include the empirical assessment of users’ preferences about diversity in content. I hold that such an empirical investigation is useful to set some low-level implementations of diversity; for instance, for making the user choose the degrees of diversity of their RSs. However, my justification for the introduction of diversity into RSs resides on a different level: its grounding is normative and consists in the respect for users’ autonomy in RSs. My thesis is that the introduction of diversity is a way to appropriately treat users in RSs, namely, to respect their agency.

My moral defense could be accused of being too demanding and encroaching on non-digital spheres of the user, since, as humans, we tend to mingle with individuals that are similar to us in certain aspects and thus be part of groups in which diversity is low and novelty is unlikely. My reply is that, even if we have this natural tendency, the non-digital world is an open space in which a new encounter or the discovery of new content is always more likely than in the enclosed space of RSs. This explains the need for diversity in RSs.

In addition, it may be objected that the risk influencing the user towards standardized and past-biased lives can be limited by opting out of the use of technologies and platforms using RSs and thus that diversity has a low ethical relevance in RSs. Yet the trend in the use of content- and utility-based RSs and collaborative-filtering and demographic RSs is expected to grow (Fayyaz et al., 2020; Grand View Research, 2021) and today there are no browsers, social media, or social platforms that do without recommendation systems. The more the user employs RSs in their life, the more their everyday environment will correspond with the digital environment, and thus the homogenization of their lives in terms of past-oriented and standardized lifestyles are a real and tangible threat.

From an empirical perspective, novel and diverse content can be unpleasant or make users lose time and interest in a RS. The discovery of content that is not liked is the opposite of serendipity and is called zemblanity (Boyd, 1998). Hence, it may be objected that morally defending diversity in RSs is not feasible as diversity is incompatible with or at least hinders the function of RSs, which is providing relevant content. My reply is threefold. First, computer scientist aims to develop RSs that do not only provide diverse and novel content but diverse, novel and relevant (serendipitous) content. Second, while it is expectable that the user prefers serendipity over zemblanity, the latter is anyway a source of information about oneself and one’s own preferences: it helps to know what we do not like. Third, as shown by Taleb (2012) even a diverse experience that is not liked is worthy as it increases the property of being antifragile: as a result of negative encounters, humans improve their capacity to thrive and adapt themselves to the environmental changes.

Now I deal with the objections that address the theoretical basis of my moral defense: the rights to an open present, to be treated as an individual, and to be an exception to predictions. Regarding the right to an open present, first, someone may object that it is not a right from the perspective of some theories of rights, as, for example, in the will theory (Hart, 1955). I call it a right as it is the application of Feinberg’s right to an open future to the user’s future self and is a normative claim that balances a power asymmetry; however, I do not subscribe to a specific theory of rights. I defend the normative force of the right to an open present, but I am open to the type of such a normative force; this means that it can be just defined as a moral principle.

A further objection raises from the definition of rights as only predictable of existing beings (Beckerman & Pasek, 2001; George & Richard, 1981; Macklin, 1981). The user’s future self can be considered as not existing yet and thus not entitled to a right that has to be enforced in the present. I claim that the future self’s not-yet-existence is not morally relevant by means of a thought experiment regarding the rights of a person’s selves throughout this person’s life. The case I present is an intrapersonal version of the case that Feinberg, 1984 presented to argue for the validity of rights of not-yet-existent persons. If I inject a fatal virus into a person, I violate their right to life (or in my framework, their present self’s right to life). If I inject a variant of that fatal virus that will activate only in 10 years into a person, I violate their right to life in the future (or in my framework, the right to life of this person’s future self). I contend that the two actions are both violations of the right to life, no matter when the virus activation takes place.

It may be objected that the thought experiment of the virus injection is valid only if I am sure that that person will be alive in 10 years, which is a piece of information I do not and cannot know. In the absence of that piece of information, it seems that a right of a future hypothetical person cannot be defended and thus RSs without diversity do not violate the future self’s right to an open present. My reply is that even in our everyday interactions with other people, we “risk” respecting moral principles involving somebody that could die prematurely; consider, for instance, a promise requiring some years to be kept. In such cases, we usually assume that other people have a normal lifespan. In the digital world, our synchronic interactions with other people on, for instance, social platforms, are based on the same assumption. From that, I conclude that both in our real-world and digital relations, we interact with each other as if we and other people continue existing in the future. Accordingly, we should have a similar attitude towards our future selves in the absence of clear evidence that a user’s future self will cease to exist (e.g., the user is affected by a terminal disease). This means that, as the user is assumed to still exists in the near future, RSs without diversity violate the right to an open present of the user’s future selves. The extent to which a RS is bound to the right to an open present corresponds to the user’s necessary conditions for the pursuit of their normative principles of action that the RS affects. For instance, a RS for meditation and anxiety prevention without diversity affects the user’s health, which is a necessary condition for the pursuit of their principles of action.

A final objection to the right to an open present concerns its alleged incompatibility with the essence of making a choice (which necessarily implies eliminating the unchosen options) and, when the right is applied to RSs, with RSs’ function of filtering out some options. This objection is based on the misunderstanding of considering the open present as a space in which no option is closed to the individual. The right to an open present does not require that we always leave as many options as possible open; rather it requires choosing options that protect the necessary conditions for the pursuit of any normative principle. When applied to RSs, this right forbids to provide the user only with recommendations perpetuating their past choices, which favor only their past normative principles of action.

Regarding the right to be treated as an individual, three main objections may arise. The first is that even when we use individual evidence about an individual, we use some generalizations based on statistical evidence (Schauer, 2003, pp. 101, 103, 172). My reply is that the right to be treated as an individual regulates the use of statistical evidence based on inductive generalizations about other people’s preferences, choices, and traits as a basis for the recommendations of a user (Viganò et al., 2022). The right does not regulate inductive generalization based on the individual’s choices. The right to an open present gives constraints to the use of information about the individual’s past choices and preferences.

The second objection to the right to be treated as an individual is the contention that in democratic societies, this right opposes the ideal of a society of equals, who, for example, have the same right to vote (Lippert-Rasmussen, 2011). My reply is that the right to be treated as an individual is not applicable in contexts in which the uniform application of the same rule to all people has been established, such as the principle stating that we are all equal before the law in courtrooms. RSs though are not part of such contexts. They have been designed to provide tailored and personalized recommendations; thus, it would make no sense to provide everyone with the same recommendations to respect a principle of equity in which the same rule is applied to everyone.

The third objection against the right to be treated as an individual consists in arguing that principle M does not provide the user with a claim that has the normative force of a right. Similarly to my reply to the objection that the right to an open present is not as such, I defend the normative force of treating people as individuals, not a specific type of this normative force, thus I am open to considering the right to be treated as an individual as just a moral principle.

Regarding the right to be an exception to predictions, it may be objected that in everyday life and the non-digital world, one tends to predict other people’s behavior on the basis of what the latter did in the past or of certain behaviors and traits of a group these people (are thought to) belong to. I reply that we need to distinguish the contexts in which we use the right to be an exception to predictions. This right is founded on principle M and thus subject to the same constraints of principle M. The right to be an exception to predictions can be disregarded in cases in which gathering individual evidence is time-consuming and/or costly, and when relying on statistical evidence fulfills a more stringent moral principle (e.g., saving a life by choosing the treatment on the basis of a person’s gender). These constraints indicate the contexts in which RSs are exempted from respecting this right, which include the cold start problem, i.e., when a user starts using a RS and thus the latter has limited data about them.

6 Conclusion

In this contribution, I provided a moral defense of diversity in RSs based on the user’s agency. I showed that without diversity most of RSs, namely content-based, utility-based, collaborative-filtering, and demographic RSs render the user’s digital environment homogeneous and immutable. I contended that such an environment progressively limits three factors that are essential to any good life, which are life experience, self-knowledge, and authenticity. I also argued that the exclusion of diversity from the design of RSs morally wrongs the RSs user by violating either the right to an open present or the right to be treated as an individual. The former right protects the freedom of one’s future self to pursue their set of normative principles of action (i.e., plans, values, and preferences that motivate them to take an action). Content- and utility-based RSs deprived of diversity violate the right to an open present because they provide the user with content based on their past choices (thus past normative principles) and therefore hinder the possibility that the user finds out new or changes some normative principles of action. The right to be treated as an individual is based on the moral principle requiring treating people as individuals, namely as autonomous agents that are principles of their actions. Collaborative-filtering and demographic RSs deprived of diversity predict a user’s preferences only on the basis of their membership to a certain group, thus they violate this right.

By combining the rights to an open present and to be treated as an individual in RSs, I derived the right to be an exception to predictions. The latter states that the user, qua agent, is not entirely determined by their past, nor by the group they are associated with and thus should be treated as a partly non-predictable being. I showed that diversity in RSs enables respecting this right by suggesting items that are different from the ones they or the group they are associated with previously chose.

My moral defense of diversity in RSs clarifies and protects the interests of the user, who, in absence of an ethical defense of diversity, is in a vulnerable position among the stakeholders of RSs. Also, it applies to RSs of all types of items, which, if deprived of diversity, create a homogeneous an immutable environment. Finally, my moral defense shows that diversity in RSs is not only an empirical matter of subjective tastes, but it has also a normative side that needs to be investigated from the perspective of ethics and thus makes all RSs subject to ethical scrutiny, not only those recommending news and scientific articles.