1022702

research-article2021
NMS0010.1177/14614448211022702new media & societyHermann

Review Article

new media & society

Artificial intelligence and

2022, Vol. 24(5) 1258­–1277
© The Author(s) 2021
Article reuse guidelines:
mass personalization of sagepub.com/journals-permissions
DOI: 10.1177/14614448211022702
https://doi.org/10.1177/14614448211022702
communication content— journals.sagepub.com/home/nms

An ethical and literacy

perspective

Erik Hermann
IHP—Leibniz-Institut für innovative Mikroelektronik, Germany

Abstract
Artificial intelligence (AI) is (re)shaping communication and contributes to (commercial
and informational) need satisfaction by means of mass personalization. However, the
substantial personalization and targeting opportunities do not come without ethical
challenges. Following an AI-for-social-good perspective, the authors systematically
scrutinize the ethical challenges of deploying AI for mass personalization of
communication content from a multi-stakeholder perspective. The conceptual analysis
reveals interdependencies and tensions between ethical principles, which advocate the
need of a basic understanding of AI inputs, functioning, agency, and outcomes. By this
form of AI literacy, individuals could be empowered to interact with and treat mass-
personalized content in a way that promotes individual and social good while preventing
harm.

Keywords
Artificial intelligence, ethics, literacy, mass, personalization

Introduction
Artificial intelligence (AI) is not just a technology but constitutes an encompassing
power (re-)shaping daily practices, individual and professional interactions, and environ-
ments (Taddeo and Floridi, 2018). Its transformative impact also pertains to how people

Corresponding author:
Erik Hermann, Wireless Systems, IHP—Leibniz-Institut für innovative Mikroelektronik, Im Technologiepark
25, 15236 Frankfurt (Oder), Germany.
Email: [email protected]
Hermann 1259

communicate, which content they encounter, and how content is generated and dissemi-
nated (Guzman and Lewis, 2020; Hancock et al., 2020; Lewis et al., 2019). Among other
things, AI has been and is a powerful force in the personalization of communication
content (Sundar, 2020). The sophistication and computational power of AI applications
in combination with the availability of big data (e.g. individuals’ digital traces) facilitates
the unprecedented personalization of communication content and messages at an indi-
vidual level and likewise on a massive scale to a large audience (Winter et al., 2021).
That is, AI enables the mass personalization of communication content (i.e. commercial,
editorial/journalistic, and user-generated messages and information). Targeting by tai-
lored persuasive appeals (e.g. Matz et al., 2017), entertainment and commercial (vendor)
platforms based on recommender systems (e.g. Milano et al., 2020), news feeds of social
network sites (e.g. Bakshy et al., 2015), and automated news production and dissemina-
tion including newsbots (e.g. Lewis et al., 2019)—to name but a few—have become part
of our daily lives (Kitchin, 2017; Willson, 2017). In spite of the general benefits of per-
sonalization such as increased personal relevance and satisfaction of individuals’ wants
and needs (Sundar and Marathe, 2010), AI-driven mass personalization does not come
without ethical concerns, which relate to privacy (e.g. Matz et al., 2019a), agency (e.g.
Soffer, 2019; Sundar, 2020), biases and transparency (e.g. Hancock et al., 2020), and
filter bubbles and echo chambers (e.g. Levy, 2021)—among other things.
Generally, the mounting pervasiveness of AI systems and application has sparked the
debate of ethical principles and values guiding AI development and use (e.g. Cowls
et al., 2021; Floridi et al., 2018, 2020; Hagendorff, 2020; Jobin et al., 2019; Mittelstadt,
2019; Morley et al., 2020). To date, the AI ethics landscape is rather fragmented and
entails recurring principles (Jobin et al., 2019) that are of high-order, deontological
nature (Hagendorff, 2020). Accounting for these principles in practice while taking into
account the different stakeholder interests might demand tradeoffs. In light of AI’s impact
on the individual, economic, and societal level, the AI ethics literature increasingly
focuses ethical frameworks of AI for (social) good (Cowls et al., 2021; Floridi et al.,2018,
2020; Taddeo and Floridi, 2018). That approach addresses and attempts to solve the ten-
sion between leveraging the benefits and preventing (or at least mitigating) potential
harms of AI—to achieve a “dual advantage” for society (Floridi et al., 2018: 694).
To the best of our knowledge, our conceptual analysis is the first study to systemati-
cally scrutinize the ethical principles related to AI to AI-driven mass personalization
from a multi-stakeholder perspective. Thereby, we provide two important contributions
to the AI ethics and communication literature. First, we reveal several interdependent
ethical challenges in respect to AI-driven mass personalization. Second, we suggest AI
literacy as a mean to leverage these ethical challenges to empower individuals, which
eventually benefits society at large.
The remainder of our study is structured as follows. After delineating our methodo-
logical approach and illustrating the role and use of AI in mass personalization, we pre-
sent an overview of the AI ethics literature. Afterwards, we consolidate both perspectives
by applying selected ethical principles to AI-powered mass personalization. We con-
clude our investigation with proposing AI literacy as a potential individual remedy to
address the interdependent ethical challenges.
1260 new media & society 24(5)

Methodology
We conducted a comprehensive literature search of published papers to identify relevant
scholarly work. First, we performed a keyword search of electronic databases (Web of
Science, EBSCO, and Google Scholar) using the following keywords: “ethic*,” “guide-
lines,” “principles,” “framework,” (for AI ethics) and “communication,” “mass person-
alization,” “personalization,” “customization,” (for AI in communication) each in
combination with “artificial intelligence,” “AI,” “artificial,” “machine learning,” “algo-
rithm*,” “bots.” Second, we examined references of review and seminal articles in both
fields (e.g. Guzman and Lewis, 2020; Hancock et al., 2020; Sundar, 2020 and Floridi
et al., 2018; Jobin et al., 2019; Mittelstadt, 2019 respectively) and applied an ancestry
tree search by screening all papers citing these articles. Third, we performed manual
search of journal outlets that turned out to be major sources for journal articles dealing
with AI in communication (e.g. Computers in Human Behavior, Information,
Communication & Society, Journal of Computer-Mediated Communication, New Media
& Society) and AI ethics (i.e. Ethics and Information Technology, Minds and Machines,
Nature Machine Intelligence, Science and Engineering Ethics).

Mass personalization and AI

Personalization refers to the “degree to which receivers perceive a message reflects their
distinctiveness as individuals differentiated by their interests, history, relationship net-
work, and so on” (O’Sullivan and Carr, 2018: 1166). As a form of system-initiated per-
sonalization, it differs from (user-initiated) customization, where individuals deliberately
tailor content by choosing options and/or creating new content (Sundar and Marathe,
2010) and become sources of communicative interactions (i.e. self-as-source; Kang and
Sundar, 2016). Mass personalization unifies characteristics of mass communication, that
is, technologically mediated content is delivered to large audiences, and interpersonal
communication, that is, personalized content reflecting recipients’ uniqueness, distinc-
tiveness, or identity (e.g. Kalyanaraman and Sundar, 2006), which is comparable to
interpersonal messages by traditional definition (O’Sullivan and Carr, 2018). Although
digital media have substantially simplified content personalization, it existed long before
the advent of digital media (Sundar and Marathe, 2010). In a seminal essay on personali-
zation of mass media, Beniger (1987) noted,

The capacity of such mass media for simulating interpersonal communication is limited only
by their output technologies, computing power, artificial intelligence; their capacity for
personalization is limited only by the size and quality of data sets on the households and
individuals to which they are linked. (p. 354)

The computational power of AI and the availability of massive amounts of (digital and
social media) data (e.g. Cappella, 2017; Harari et al., 2016; Matz and Netzer, 2017;
Stachl et al., 2020a; Stahl et al., 2021; Winter et al., 2021)—of metrified and tracked
individuals (Gerlitz and Helmond, 2013; König et al., 2020)—have relativized some of
the limits stressed by Beniger (1987). Thereby, AI allows to personalize content at
Hermann 1261

unprecedented speed, scale, intensity, and responsiveness. Correspondingly, Kotras

(2020) defined mass personalization as “algorithmic processes in which the precise
adjustment of prediction to unique individuals involves the computation of massive data-
sets, compiling the behaviors of very large populations” (p. 2). Thus, mass personaliza-
tion and algorithms are inextricably intertwined. Put simply, algorithms produce outputs
from inputs by means of certain (deterministic) rules or procedures (Hill, 2016). Machine-
learning algorithms are methods that utilize data to identify (novel) patterns and underly-
ing rules (Kaplan and Haenlein, 2019). They have the capacity to define or modify
decision-making rules autonomously (Mittelstadt et al., 2016). Although machine learn-
ing is a focal part of AI, AI is broader due to its ability to perceive data (e.g. language
processing) and other human-like capabilities such as moving objects (e.g. robots) or
conversation capacities (e.g. chatbots; Kaplan and Haenlein, 2019). The concept of AI is
polysemous and not well-defined (Guzman and Lewis, 2020; Stahl et al., 2021). In its
simplest sense, AI refers to technologies performing tasks that are associated with some
level of human intelligence (Guzman and Lewis, 2020). In more technical terms, AI can
be defined as “a system’s ability to correctly interpret external data, to learn from such
data, and to use those learnings to achieve specific goals and tasks through flexible adap-
tation” (Kaplan and Haenlein, 2019: 17).
Algorithms are the integral components of recommender systems (e.g. Lury and Day,
2019; Milano et al.,2020, 2021; Mittelstadt et al., 2016), that are the foundation of and/
or intensively used by news intermediaries and aggregators, social media, entertainment,
and commercial (vendor) platforms (Bozdag, 2013; Helberger, 2019; Lury and Day,
2019). Recommender systems refer to (algorithmic) functions utilizing information
about individual preferences (e.g. products or news items) as inputs to predict how indi-
viduals would rate certain items under evaluation (e.g. new items available) and how
they would rank a set of items individually or as a bundle (Milano et al., 2020). Inputs
from individuals can include any form of reactions (e.g. comments, likes, ratings,
reviews) to news, products, or other social, political, cultural, or entertainment stimuli—
all of which being indicative of social norms evaluating sociocultural entities. That
makes respective recommendations a form of platform-mediated interpersonal commu-
nication (Cappella, 2017). Recommender systems can take the form of collaborative
filtering, content-based filtering, or hybrid methods (Bozdag, 2013; Lury and Day, 2019;
Milano et al., 2020). Collaborative filtering algorithms base their recommendations on
target individuals’ past behavior, choices, and preferences, and on preferences of other
individuals being structurally similar to them (Cappella, 2017; Lury and Day, 2019).
These recommendations of future choices based on similar tastes and patterns of past
choices can be considered as a surrogate for social influence (Cappella, 2017) or auto-
mated word-of-mouth (Bozdag, 2013). However, content-based filtering algorithms
make use of discrete characteristics and properties of items to generate recommendations
of items with similar characteristics and properties that individuals preferred in the past
(Bozdag, 2013; Cappella, 2017; Lury and Day, 2019). Besides, algorithmic content fil-
tering and ranking can personalize, prioritize, and curate content (e.g. search engines,
news feeds of social network sites) for individuals (e.g. Bakshy et al., 2015; Bozdag,
2013; Lazer, 2015; Möller et al., 2018; Scharkow et al., 2020; Schwartz and Mahnke,
2021).
1262 new media & society 24(5)

Personalization might not only be based on individuals’ preferences, interests, demo-

graphics, and past behavior, item features and characteristics, or similar tastes of others,
but also on psychological factors—the method of psychological targeting (Hirsh et al.,
2012; Matz and Netzer, 2017; Matz et al., 2017; Matz et al., 2019b; Stachl et al., 2020a;
Winter et al., 2021; Zarouali et al., 2020). AI-powered psychological targeting offers
considerable opportunities to tailor (persuasive) appeals to individuals’ psychological
traits (variability across consumer such as personality traits or values) and states (varia-
bility within consumers over time such as mood or emotions) that are computationally
predicted from their digital footprints (Matz and Netzer, 2017; Matz et al., 2017).
Combining large-scale (digital) data with the computational power of AI allows “an
unprecedented understanding of consumers’ unique needs as they relate to the situation-
specific expressions of more stable motivations and preferences” (Matz and Netzer,
2017: 9). Therefore, AI could be leveraged for both content creation of psychologically
tailored appeals and situation-specific and context-aware dissemination of such appeals.
Apart from commercial appeals, the dual role of AI in personalized content produc-
tion and dissemination also pertains to news and journalistic content (e.g. Bodó, 2019;
Bodó et al., 2019; Ford and Hutchinson, 2019; Guzman, 2019; Helberger, 2019; Lewis
et al., 2019; Milosavljević and Vobič, 2019; Thurman et al., 2019a, 2019b). While the
first generation of news personalization incorporated receiver-initiated customization
based on explicitly expressed preferences, the second generation features implicit per-
sonalization techniques building on individuals’ digital profiles and indirect preference
signals (Bodó, 2019; Kunert and Thurman, 2019; Thurman and Schifferes, 2012). In
addition, newsbots have developed from rebroadcasters of news content to disseminators
of news incorporating chatbot conversation capacities, thereby becoming a third party
(person) mediating the sender-receiver relationship (Ford and Hutchinson, 2019; Lokot
and Diakopoulos, 2015) or conversational agents (Guzman and Lewis, 2020).
Generally, AI does not only facilitate, mediate, and channel communication (e.g.
Hancock et al., 2020), but also functions as a communicator and participant in commu-
nicative exchanges itself—a role that has been historically attributed to humans from a
communication-theoretical perspective (Gunkel, 2012; Guzman and Lewis, 2020;
Schwartz and Mahnke, 2021). Before we shed light on the ethical questions of AI in mass
personalization, we provide an overview of the AI ethics literature.

AI ethics
The discourse on moral and ethical implications of AI dates back from 1960 (Samuel,
1960; Wiener, 1960). The increasing pervasiveness and encompassing impact of AI
applications and systems have intensified calls for and discussions of accompanying
ethical guidelines. That is, “the ethical debate has gone mainstream” (Morley et al.,
2020: 2141). Ethical principles related to AI focus on ethical issues in respect to particu-
lar features of the technology or the consequences of its use (Stahl et al., 2021). That is,
in the tradition of computer and (information) technology ethics (e.g. Brey, 2000, 2012;
Moor, 1985, 2005; Royakkers et al., 2018; Wright, 2011), which incorporate recurring
principles and themes such as autonomy, justice, beneficence, non-maleficence, dignity,
and privacy (Brey, 2012; Royakkers et al., 2018; Wright, 2011).
Hermann 1263

These principles also characterize ethical frameworks related to AI. In a comprehen-

sive review of 84 documents of principles and guidelines for ethical AI issued by private,
public, and research institutions, Jobin et al. (2019) found convergence around the prin-
ciples transparency (referenced in 73 out of 84 documents), justice and fairness (68),
non-maleficence (60), responsibility (60), privacy (47), beneficence (41), and freedom
and autonomy (34). However, no single ethical principle was referenced in all 84 docu-
ments. The prevalence of transparency could be attributed to the reasoning that it “is not
an ethical principle in itself but a proethical condition for enabling or impairing other
ethical practices or principles” (Turilli and Floridi, 2009: 105). While frequently refer-
enced principles such as justice and fairness, non-maleficence, and privacy reflect a cau-
tious view on potential risks of AI, the more frequent occurrence of non-maleficence as
compared to beneficence implies the moral obligation to avoid any negative impact of AI
and a certain negativity bias (Jobin et al., 2019). The role of trust as an AI governance
principle is not without opposition and ambiguity, particularly, whether trust is a princi-
ple in itself or rather an outcome of other foundational principles (e.g. Floridi, 2019;
Ryan, 2020; Thiebes et al., 2020). The solidarity principle is only featured in 6 out of 84
documents, although it refers to redistributing the benefits of AI to not jeopardize social
cohesion (Jobin et al., 2019). In light of persistent and mounting global inequalities,
prosperity, and burdens created by AI should be shared, that is, solidarity should be con-
sidered as a focal ethical principle of AI (Luengo-Oroz, 2019).
The solidarity and beneficence principles already hint at the need to (equitably) lever-
age the benefits of AI on a societal level. Ethical frameworks for AI for (social) good
(Cowls et al., 2021; Floridi et al.,2018, 2020; Taddeo and Floridi, 2018) are in line with
this stance and advocate the following five focal ethical principles: beneficence, non-
maleficence, autonomy, justice, and explicability (Floridi et al., 2018). Figure 1 provides
a systematization of the ethical principles identified by Jobin et al. (2019) and Floridi
et al. (2018) and how they align.
While the tenet of beneficence refers to the promotion of well-being as well as social,
environmental, and common good (Jobin et al., 2019; Thiebes et al., 2020), the non-
maleficence principles caution against the potentially negative aspects of AI. Central to
non-maleficence are safety, security, privacy, and generally, the prevention of risks and
any harm—both accidentally or unintentionally (overuse) and deliberately (misuse)
caused (Floridi et al., 2018; Jobin et al., 2019). Across ethical frameworks, privacy often
constitutes a principle on its own. However, given the emphasis on avoiding infringe-
ments of privacy, breaches of data protection, and misuse of data, that is, prevention of
harms and risks in respect to personal data and privacy, privacy can be subsumed under
the non-maleficence principle. It is noteworthy that beneficence and non-maleficence are
not opposite ends of a continuum but coexist, although they seem logically equivalent
(Floridi et al., 2018). The principle autonomy entails self-determination and the power to
and whether to decide in an uncoerced way, that is, seeking a balance between human
and AI agency and decision-making power (Floridi et al., 2018; Morley et al., 2020). The
justice principle stresses fairness, avoiding unwanted or unfair biases, and discrimination
(Jobin et al., 2019; Morley et al., 2020; Thiebes et al., 2020), as well as sharing benefits
and prosperity and fostering solidarity (Floridi et al., 2018). Thus, the solidarity principle
merges in the justice principle.
1264 new media & society 24(5)

Figure 1. AI ethics map.

Principles in bold are taken from Floridi et al. (2018), while principles in normal font and italics are taken
from Jobin et al. (2019). Principles in italics were not subsumed under beneficence, non-maleficence, justice,
or explicability, but listed as independent principles by Jobin et al. (2019).

Finally, explicability comprises intelligibility (i.e. how AI works—the epistemologi-

cal sense) and accountability (i.e. who is responsible for the way AI works—the ethical
sense; Floridi et al., 2018). In the literature, different nomenclature and concepts, that is,
intelligibility, comprehensibility, interpretability, explainability, and transparency, are
used interchangeably and inconsistently (Barredo Arrieta et al., 2020), and are partly
misconceived (Rudin, 2019). In a comprehensive review, Barredo Arrieta et al. (2020)
identified intelligibility, that is, human understanding of a model’s function without any
need for explaining its internal structure or underlying data processing algorithm, as the
most appropriate conceptualization. The narrow relation between intelligibility and
accountability (e.g. Lepri et al., 2018; Martin, 2019; Morley et al., 2020) justifies their
subsumption under explicability. That is, judgments about accountability necessitate a
certain understanding of the underlying processes of AI systems and applications (i.e.
intelligibility; Lepri et al., 2018). Understanding the functionalities (i.e. intelligibility)
and responsibilities (i.e. accountability), in turn, informs evaluations of the other princi-
ples by comprehending if and how AI benefits or harms individuals and society (benefi-
cence and non-maleficence), by drawing conclusions about whether to delegate decisions
to AI systems (autonomy), and by knowing whom to hold accountable in case of failures
or biases (justice; Floridi et al., 2018; Thiebes et al., 2020).
AI ethics frameworks have in common that they focus high-level ethical principles
with little reference to philosophical ethical theories (Stahl et al., 2021). However, a
predominantly principled approach is called into question for at least two reasons (e.g.
Hagendorff, 2020; Mittelstadt, 2019; Theodorou and Dignum, 2020). First, deontologi-
cal and normative imperatives and principles (Hagendorff, 2020) lack translation into
Hermann 1265

Figure 2. Multi-stakeholder model of ethical principles related to AI-powered mass

personalization.

practice through mid-level norms and low-level requirements taking into consideration
the legal, technical, and social circumstances (Mittelstadt, 2019). Applied AI ethics could
close the gap between principles (what) and the practice of how to develop ethical AI
(Morley et al., 2020). Second, AI is not developed and deployed in isolation, but within
the socio-technical system (i.e. people, organizations, their interactions, and processes
organizing these interactions) it is operating and unfolding. Therefore, concrete ethical
and socio-legal governance and policies are in demand (Cath, 2018; Theodorou and
Dignum, 2020).
In the following, we examine ethical principles and controversies of deploying AI for
the mass personalization of communication content.

The ethics of AI in mass personalization of communication

content
We investigate the ethical implications and concerns of AI-driven mass personalization
of communication content from a multi-stakeholder perspective comprising content
senders, content receivers, and society at large (see Figure 2). Following communica-
tion-theoretical conceptualizations, we refer to senders and receivers. Admittedly, this
distinction is far from unequivocal, and transitions can be fluid. For instance, the tradi-
tional receivers (i.e. individuals) can now create and share content on their own (i.e.
user-generated content). We simplistically define senders as the entities operating and/or
economically profiting from AI systems that create or disseminate content, while receiv-
ers are the targets the content is directed to. By adopting a multiperspectivity approach,
we want to provide a holistic picture of the ethical considerations beyond the individual
content receivers. Particularly, phenomena such as filter bubbles, echo chambers, and
respective (ideological) polarization that can arise from algorithmic content filtering can
have adverse effects for democracy and society at large (e.g. Bozdag and van den Hoeven,
2015; Helberger, 2019). Our multilevel analysis further accounts for the AI-for-social-
good perspective stressed by prior AI ethics literature (Cowls et al., 2021; Floridi
1266 new media & society 24(5)

et al.,2018, 2020; Taddeo and Floridi, 2018). Correspondingly, our analysis is based on
the ethical principles suggested by this stream of research, that is, beneficence, non-
maleficence, autonomy, justice, and explicability (Floridi et al., 2018; Morley et al.,
2020).

Beneficence
First and foremost, AI-powered mass personalization of communication content can be
harnessed to match individuals’ preferences (e.g. Matz and Netzer, 2017), to increase
personal relevance and to satisfy—or at least approach—individuals’ wants and needs
(e.g. Sundar and Marathe, 2010), and to improve the attractiveness and usability of prod-
ucts, services, messages, and content, which, in turn, increases acceptance, usage, satis-
faction, and loyalty (e.g. Stachl et al., 2020b). Moreover, mass-personalized content can
serve as substitute or shortcut for extensive information search and gathering through
information (pre-)filtering and selection (Cappella, 2017) leading to better information
and efficiencies (e.g. time savings) on the content receiver side (e.g. Helberger, 2019).
These advantages on the content receiver level also benefit the content senders in terms
of adoption rates, satisfaction, loyalty and retention, profit, and resource efficiencies.
In general, AI-powered mass personalization is not configured and employed for the
sake of it but relates to specific benefits and a clear purpose, which implies justifica-
tion—one requirement for beneficence (Morley et al., 2020). Nevertheless, judgments of
benefits, goodness, and hence, the beneficence principle can be ambiguous (D’Acquisto,
2020). That is, beneficence on the content sender and receiver level does not necessarily
imply beneficence on the societal level. As mentioned earlier, personalization entails the
(pre-)selection of content and recommendations individuals are exposed to, which can
eventually result in content receivers’ selective exposure to content. Selective exposure
and limited content diversity could lead to polarization, that is, strengthening of indi-
viduals’ original attitude or position (Stroud, 2010), echo chambers, or filter bubbles,
that is, individuals are only encountering content from like-minded individuals or con-
tent selected by algorithms according to individuals’ previous behavior and interactions
with the system, respectively (Bakshy et al., 2015; Bozdag and van den Hoeven, 2015).
While some studies found evidence that mass personalization induces selective expo-
sure and polarization (e.g. Dylko et al., 2017; Levy, 2021), other studies challenge the
assumption and concerns that personalization algorithms necessarily or solely cause fil-
ter bubbles, echo chambers, or polarization (Bakshy et al., 2015; Berman and Katona,
2020; Flaxman et al., 2016; Fletcher and Nielsen, 2018; Haim et al., 2018; Messing and
Westwood, 2014; Möller et al., 2018; Nechushtai and Lewis, 2019; Scharkow et al.,
2020; Zuiderveen Borgesius et al., 2016). Despite these equivocal findings, constraints
on cross-cutting and counter-attitudinal content that undermines balanced content diver-
sity would limit the beneficence of AI-driven mass personalization on a societal level. On
the content sender and receiver levels, exploitation of existing individual information for
tailoring content by AI systems often constitutes the optimal (standard) strategy to maxi-
mize individual utility and satisfaction. However, exploitation strategies could also run
the risk of choosing prediction accuracy over satisfaction (e.g. Kotkov et al., 2016) and
of underrepresenting new or alternative content (i.e. limiting content diversity) as
Hermann 1267

compared to more explorative strategies (e.g. Milano et al., 2021). Therefore, some
scholars emphasize the importance and value of serendipity, that is, recommendations of
items that are relevant, novel, and unexpected (and thus, relatively unpopular and signifi-
cantly different from individuals’ profiles; Kotkov et al., 2016), and diversity-sensitive
designs to promote content diversity and satisfaction (e.g. Fletcher and Nielsen, 2018;
Helberger et al., 2018; Kotkov et al., 2016; Levy, 2021; Reviglio, 2019). Taken together,
the unconditional beneficence of AI-powered mass personalization can be called into
question, which draws the analogy to the non-maleficence principle.

Non-maleficence
In the case of limited content diversity resulting from AI-driven mass personalization,
judgments of beneficence and non-maleficence from a societal perspective are related.
That is, mass personalization does not necessarily foster social good (beneficence princi-
ple not met) but can compromise it (non-maleficence principle not met). Contrarily, the
respective ethical judgments do not coincide or are even inversely related when compar-
ing the content sender and receiver levels. That means, mass personalization could be
beneficent (e.g. for content senders) and maleficent (e.g. for content receivers) at the
same time (Milano et al., 2021).
Notably, personal privacy, accuracy, data protection, and quality are central to the
non-maleficence principle (e.g. Floridi et al., 2018; Morley et al., 2020). Privacy risks
can arise (1) when data are gathered without informed consent of individuals, (2) after
storage when they are leaked or de-anonymized (i.e. data breaches), (3) when AI systems
draw inferences from both individual data (directly), or interaction data with others (indi-
rectly; Milano et al., 2020). The large-scale (digital) data feeding AI systems and appli-
cations can aggravate privacy and data protection issues (Baruh and Popescu, 2017), as
discussed for algorithms (e.g. Mittelstadt et al., 2016), recommender systems (e.g.
Milano et al., 2020), psychological targeting (e.g. Matz et al., 2019a), and personaliza-
tion in general (e.g. Cloarec, 2020).
In the case of AI-powered mass personalization, there is a tension and potential trade-
off between the scale and scope of data inputs for mass personalization and privacy con-
cerns, that is, the heightened amount of data inputs to achieve predictive validity and
accuracy of personalization efforts could interfere with data protection and privacy. Since
AI system’s inferences and predictions are as accurate and reliable as the underlying data,
quality and integrity of data are decisive. Biases, inaccuracies, and errors inherent in data
could bias results and lead to false conclusions (e.g. Barredo Arrieta et al., 2020; Hancock
et al., 2020; Morley et al., 2020). Furthermore, algorithmic predictions are directed to
individuals, although inferences are drawn from populations. That becomes problematic
when algorithmic inferences are based on (potentially spurious) correlations found in
large datasets, because causality is often not established prior to algorithmic decisions
(Mittelstadt et al., 2016). Inferior or biased predictions and recommendations can be par-
ticularly adverse for individuals if they depend too much on algorithm-generated recom-
mendations (i.e. algorithm overreliance) that could then harm their well-being (Banker
and Khetani, 2019). Whether or not individuals are exposed to and influenced by mass-
personalized content is narrowly related to the autonomy principle.
1268 new media & society 24(5)

Autonomy
Autonomy relates to a meta-autonomy or decide-to-delegate model, that is, “humans
should always retain the power to decide which decisions to take” on their own or when
to cede decision-making control (Floridi et al., 2018: 698). Human agency (i.e. autono-
mous decisions) and human oversight are focal requirements of autonomy (Morley et al.,
2020).
AI in mass personalization act as (secondary) gatekeeper creating, selecting, filtering,
and disseminating the content individuals eventually encounter (Just and Latzer, 2017;
Singer, 2014; Soffer, 2019). Therefore, content receivers’ autonomy is concerned to the
effect that decisions (i.e. agency) are delegated to AI systems at the information collec-
tion stage of the decision-making process, particularly, (pre-)filtering of information and
options individuals are exposed to. Because personalization, psychological targeting,
and recommender systems can serve as adaptive, structural, or informational nudges
(Floridi, 2016; Milano et al., 2020; Sunstein, 2016), individuals’ decision-making pro-
cesses are influenced. That is, these kind of interventions shape individuals’ choice sets
or information related to choices and eventually preferences and decisions. That can be
beneficial due to resource efficiency (e.g. time, cognitive resources) and personally rel-
evant content, but also detrimental in case of overreliance on mass-personalized recom-
mendations (e.g. Banker and Khetani, 2019) or due to manipulated or deceptive content
(e.g. Hancock et al., 2020; Milano et al., 2020). Human agency could be fostered by
preferring reactive personalization (i.e. obtaining permission before providing personal-
ized content) and overt data gathering over proactive personalization (i.e. automatically
providing personalized content) and covert data gathering (Sundar, 2020; Sundar and
Marathe, 2010).
On the content sender level, AI systems are granted autonomy both at the content
creation and dissemination stage (Hancock et al., 2020). Therefore, governance mecha-
nism should be implemented to facilitate human agency and oversight and to keep
humans in the loop (e.g. Thiebes et al., 2020), particularly, in ethically or morally salient
contexts. Generally, the question of autonomy on the content sender and receiver levels
should not be framed as a dichotomy between human and AI agency, since humans are
either treated as passive victims of AI predictions or they are entirely held accountable
for any potential negative effect of (neutral) AI that mediates human inputs (Schwartz
and Mahnke, 2021).

Justice
As human judgments can be error-prone, biased, and discriminating, so can AI predic-
tions and inferences (Kleinberg et al., 2020; Rich and Gureckis, 2019). Personalization
could “segment a population so that only some segments are worthy of receiving some
opportunities or information, re-enforcing existing social (dis)advantages” (Mittelstadt
et al., 2016: 9). Such profiling leads to “industrialized social discrimination” that creates
“winners” and “losers” being worth or not to receive content (Turow and Couldry, 2018:
417). Accordingly, AI-powered mass personalization could discriminate content receiv-
ers on the basis of psychological, economic (e.g. income), and demographic factors,
Hermann 1269

reinforce gender, age, and racial disparities, prejudices, and stereotypes (e.g. Bol et al.,
2020; Datta et al., 2015; Kleinberg et al., 2020), and/or target (psychologically, ideologi-
cally, or economically) vulnerable groups (e.g. Matz and Netzer, 2017; Matz et al., 2017).
As mentioned earlier, biased outcomes, unfair and unequal treatments, and targeting can
be attributed to biases in and skewness of underlying data (Barredo Arrieta et al., 2020;
Hancock et al., 2020; Morley et al., 2020). Sources of bias include but are not limited to
over- and underrepresentation of demographic groups or sensitive features, considera-
tion of misleading proxy features (Barredo Arrieta et al., 2020), or data sparsity in respect
to certain individuals, features, and items (Batmaz et al., 2019; Rich and Gureckis, 2019).
In light of these multiple sources of biases, diligence and monitoring along the entire
data lifecycle and in respect to AI development (e.g. model specification) and deploy-
ment are advisable if not indispensable.
Discrimination is not limited to content receivers but can also affect content senders.
In the commercial domain, mass personalization (e.g. recommender systems) can be
discriminatory by decreasing sales diversity (i.e. a lack of serendipity) and by increasing
market share concentration for popular items (Lee and Hosaganar, 2019). Besides, dis-
crimination can arise from senders’ presence versus absence on multisided (e-commerce)
platforms deploying AI-driven mass personalization and respective unequal market
access (Milano et al., 2021). Apart from commercial content, informational and attitudi-
nal content and its senders can be subject to unbalanced representation caused by the
issue of selective exposure, echo chambers, and filter bubbles delineated earlier.
Taken together, content senders and receivers can suffer from biases, discrimination,
and amplification of existing inequalities due to AI-driven mass personalization, which
can, in turn, diminish social good and well-being, which establishes the connection to the
beneficence and non-maleficence principles.

Explicability
Due to the black-box nature of AI systems (i.e. black-box AI), their opacity and lack of
accountability (Ananny and Crawford, 2018; Milano et al., 2020; Mittelstadt et al., 2016;
Thiebes et al., 2020; Willson, 2017), explicability (i.e. intelligibility and accountability)
features a prominently and controversially debated ethical principle, particularly, when
high-stake decisions and sensitive, personal data are involved (e.g. Barredo Arrieta et al.,
2020; Rudin, 2019). For content receivers, a basic understanding of how AI functions
(i.e. intelligibility) and personalizes content might be more effective and satisfying than
complicated and methodologically detailed explanations causing information overload,
irritation, and frustration (Barredo Arrieta et al., 2020). Moreover, content receivers have
a legitimate interest in knowing who to hold accountable (i.e. accountability) for adverse,
biased, or discriminatory outcomes of personalized recommendations and targeting
activities. That becomes particularly important if AI systems and algorithms are consid-
ered and conceptualized as value-laden rather than neutral (e.g. Martin, 2019).
However, addressing the black-box and opacity issue of AI-driven mass personaliza-
tion can be challenging for content senders. First, intelligibility can interfere with privacy
concerns and proprietary boundaries aiming at facilitating exclusivity or competitive
advantages (e.g. Ananny and Crawford, 2018; Willson, 2017). Second, intelligibility can
1270 new media & society 24(5)

be undermined by cognitive (i.e. excessively or insufficiently detailed information, lack

of understanding), technical (i.e. methodological and technical complexity), and tempo-
ral constraints (i.e. rapid advancements and development cycles; Ananny and Crawford,
2018; Rudin, 2019). Third, transparency and disclosure of non-human identity of AI
systems such as (news)bots can compromise their performance and efficiency, which
raises the question of how to weigh the benefits and costs of disclosing the non-human
AI nature (Ishowo-Oloko et al., 2019). Finally, overall (commercial) content diversity
can decline when content receivers are explained why they received certain recommen-
dations (Milano et al., 2020). That is, the item popularity among (similar) other users as
explanation could further amplify desirability and popularity of items—a self-reinforc-
ing, popularity-enhancing process could emerge.

AI literacy: a remedy for multiple interdependent ethical

challenges?
Our analysis reveals several interdependencies between ethical principles. First and fore-
most, explicability in the form of intelligibility can be considered as an enabling principle
for the other ethical principles. That is, an entire lack thereof (i.e. black-box AI) impedes
individuals’ judgments about beneficence, non-maleficence, justice, and autonomy.
Furthermore, justice and autonomy can determine judgments about beneficence and non-
maleficence, and the latter are related as well—even an inverse relationship between them
is possible. In sum, a basic understanding of how AI in the mass personalization context
works can be a prerequisite for individuals to assess biases (i.e. justice), their decision-
making power and agency (i.e. autonomy), privacy concerns and underlying data (i.e.
non-maleficence), and benefits (i.e. beneficence). When it comes to subtle and often
unconscious techniques of AI-powered mass personalization, content receivers’ reflection
of everyday practices and interactions with respective content (e.g. Schwartz and Mahnke,
2021) might not suffice. Instead, a form of techno-capital unifying digital media and
information literacy (e.g. Choi et al., 2020) might be in demand. We advocate a form of
AI literacy to empower consumers in and beyond the mass personalization context. As
media literacy is multi-dimensional—typically, cognitive, emotional, aesthetic, and moral
dimensions (Potter, 2010)—so AI literacy also requires personal development along vari-
ous dimensions. We conceptualize AI literacy as individuals’ basic understanding of (a)
how and which data are gathered; (b) the way data are combined or compared to draw
inferences, create, and disseminate content; c) the own capacity to decide, act, and object;
(d) AI’s susceptibility to biases and selectivity; and (e) AI’s potential impact in the aggre-
gate. Increasing awareness and empower individuals to develop AI literacy constitutes a
non-trivial and ambitious objective, but it could be an important step to leverage AI for
social good. While AI literacy mainly relates to the content receiver level, remedies on the
sender and societal level should be contemplated, too.
The existing deontological AI ethics approach might be inappropriate to account for
complex interdependencies of ethical principles. Instead, a utilitarian approach weigh-
ing benefits and costs across all stakeholders could better account for multiple values,
objectives, and utilities at the sender, receiver, and societal levels. Therefore, senders
Hermann 1271

could equip AI systems with multi-objective utility concepts taking into account ethical
principles (e.g. Vamplew et al., 2018). The respective challenging conceptualization
and implementation might require embedded ethics approaches proactively integrating
ethicists (e.g. McLennan et al., 2020). That could be supplemented by appropriate gov-
ernance and auditing mechanisms (e.g. Floridi et al., 2018). On the societal level, bind-
ing ethical and socio-legal policies (e.g. Stahl et al., 2021; Theodorou and Dignum,
2020) could provide regulatory guidance—the European Union’s General Data
Protection Regulation is a first step to address the non-maleficence (i.e. privacy) princi-
ple (Selbst and Powles, 2017).

Conclusion
This article synthesizes the research streams of AI ethics and AI-driven mass personali-
zation. Our conceptual analysis shows that ethical principles related to the AI-for-social-
good perspective interdepend and collide, partly, in dependence of the stakeholders
concerned. In particular, beneficence and non-maleficence are not necessarily met, since
mass personalization can limit content diversity and/or interfere with privacy. Besides,
explicability (i.e. intelligibility and accountability) turns out to be a precursor to the other
principles. Therefore, we propose that AI literacy (i.e. a basic understanding of AI inputs,
functioning, agency, and outcomes) could empower individuals to judge beneficence,
non-maleficence, autonomy, and justice related to AI-driven mass personalization them-
selves. AI literacy could further help individuals to retain autonomy and agency, since
they are better able to identify and assess choice architectures generated through
AI-driven mass personalization.
At their core, ethical principles should not be conceived as impediments of actions or
(technological) progress; conversely, they should rather enhance the scope of action,
autonomy, freedom, and self-responsibility (Hagendorff, 2020). We follow this path and
suggest to leverage the ethical principles and respective challenges related to AI-powered
mass personalization to conceptualize AI literacy at the receiver level and recommend
ethically aligned mechanisms and policies at the sender and societal levels. AI literate
and empowered individuals and effective governance mechanisms could promote indi-
vidual and social good while limiting harm, so that a dual advantage for society could
evolve. We hope that some of our thoughts motivate public, private, and/or research
institutions to find a balanced set of bottom-up (i.e. AI literacy, embedded ethics
approaches) and top-down (i.e. regulations) measures to exploit AI’s potential while
minimizing its negative externalities.

Declaration of conflicting interests

The author declared no potential conflicts of interest with respect to the research, authorship, and/
or publication of this article.

Funding
The author received no financial support for the research, authorship, and/or publication of this
article.
1272 new media & society 24(5)

ORCID iD
Erik Hermann https://orcid.org/0000-0003-0895-3562

References
Ananny M and Crawford K (2018) Seeing without knowing: limitations of the transparency ideal
and its application to algorithmic accountability. new media & society 20(3): 973–989.
Bakshy E, Messing S and Adamic LA (2015) Exposure to ideologically diverse news and opinion
on Facebook. Science 348(6239): 1130–1132.
Banker S and Khetani S (2019) Algorithm overdependence: how the use of algorithmic recom-
mendation systems can increase risks to consumer well-being. Journal of Public Policy &
Marketing 38(4): 500–515.
Barredo Arrieta A, Díaz-Rodríguez N, Del Ser J, et al. (2020) Explainable artificial intelli-
gence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI.
Information Fusion 58: 82–115.
Baruh L and Popescu M (2017) Big data analytics and the limits of privacy self-management. new
media & society 19(4): 579–596.
Batmaz Z, Yurekli A, Bilge A, et al. (2019) A review on deep learning for recommender systems:
challenges and remedies. Artificial Intelligence Review 52(1): 1–37.
Beniger JR (1987) Personalization of mass media and the growth of pseudo-community.
Communication Research 14(3): 352–371.
Berman R and Katona Z (2020) Curation algorithms and filter bubbles in social networks.
Marketing Science 39(2): 296–316.
Bodó B (2019) Selling news to audiences—A qualitative inquiry into the emerging logics of
algorithmic news personalization in European quality news media. Digital Journalism 7(8):
1054–1075.
Bodó B, Helberger N, Eskens S, et al. (2019) Interested in diversity: the role of user attitudes,
algorithmic feedback loops, and policy in news personalization. Digital Journalism 7(2):
206–229.
Bol N, Strycharz J, Helberger N, et al. (2020) Vulnerability in a tracked society: combining track-
ing and survey data to understand who gets targeted with what content. new media & society
22(11): 1996–2017.
Bozdag E (2013) Bias in algorithmic filtering and personalization. Ethics and Information
Technology 15(3): 209–227.
Bozdag E and van den Hoeven J (2015) Breaking the filter bubble: democracy and design. Ethics
and Information Technology 17(4): 249–265.
Brey PAE (2000) Method in computer ethics: towards a multi-level interdisciplinary approach.
Ethics and Information Technology 2(2): 125–129.
Brey PAE (2012) Anticipating ethical issues in emerging IT. Ethics and Information Technology
14(4): 267–284.
Cappella JN (2017) Vectors into the future of mass and interpersonal communication research:
Big data, social media, and computational social science. Human Communication Research
43(4): 545–558.
Cath C (2018) Governing artificial intelligence: ethical, legal and technical opportunities and chal-
lenges. Philosophical Transactions of the Royal Society A 376(2133): 20180080.
Choi JR, Straubhaar J, Skouras M, et al. (2020) Techno-capital: theorizing media and information
literacy through information technology capabilities. new media & society. Epub ahead of
print 27 May 2020. DOI: 10.1177/1461444820925800.
Hermann 1273

Cloarec J (2020) The personalization–privacy paradox in the attention economy. Technological

Forecasting & Social Change 161: 120299.
Cowls J, Tsamados A, Taddeo M, et al. (2021) A definition, benchmark and database of AI for
social good initiatives. Nature Machine Intelligence 3(2): 111–115.
D’Acquisto G (2020) On conflicts between ethical and logical principles in artificial intelligence.
AI & Society 35(4): 895–900.
Datta A, Tschantz MC and Datta A (2015) Automated experiments on ad privacy settings: a tale
of opacity, choice, and discrimination. Proceedings on Privacy Enhancing Technologies
1: 92–112.
Dylko I, Dolgov I, Hoffman W, et al. (2017) The dark side of technology: an experimental inves-
tigation of the influence of customizability technology on online political selective exposure.
Computers in Human Behavior 73: 181–190.
Flaxman S, Goel S and Rao JM (2016) Filter bubbles, echo chambers, and online news consump-
tion. Public Opinion Quarterly 80(S1): 298–320.
Fletcher R and Nielsen RK (2018) Automated serendipity. Digital Journalism 6: 8976–8989.
Floridi L (2016) Tolerant paternalism: pro-ethical design as a resolution of the dilemma of tolera-
tion. Science and Engineering Ethics 22(6): 1669–1688.
Floridi L (2019) Establishing the rules for building trustworthy AI. Nature Machine Intelligence
1(6): 261–262.
Floridi L, Cowls J, Beltrametti M, et al. (2018) AI4People—An ethical framework for a good AI
society: opportunities, risks, principles, and recommendations. Minds and Machines 28(4):
689–707.
Floridi L, Cowls J, King TC, et al. (2020) How to design AI for social good: seven essential fac-
tors. Science and Engineering Ethics 26(3): 1771–1796.
Ford H and Hutchinson J (2019) Newsbots that mediate journalist and audience relationships.
Digital Journalism 7(8): 1013–1031.
Gerlitz C and Helmond A (2013) The like economy: social buttons and the data-intensive web.
new media & society 15(8): 1348–1365.
Gunkel DJ (2012) Communication and artificial intelligence: opportunities and challenges for the
21st century. Communication+ 11(1): 1.
Guzman AL (2019) Prioritizing the audience’s view of automation in journalism. Digital
Journalism 7(8): 1185–1190.
Guzman AL and Lewis SC (2020) Artificial intelligence and communication: a human–machine
communication research agenda. new media & society 22(1): 70–86.
Hagendorff T (2020) The ethics of AI ethics: an evaluation of guidelines. Minds and Machines
30(1): 99–120.
Haim M, Graefe A and Brosius HB (2018) Burst of the filter bubble? Effects of personalization on
the diversity of Google News. Digital Journalism 6(3): 330–343.
Hancock JT, Naaman M and Levy K (2020) AI-mediated communication: definition, research
agenda, and ethical considerations. Journal of Computer-Mediated Communication 25(1):
89–100.
Harari GM, Lane ND, Wang R, et al. (2016) Using smartphones to collect behavioral data in psy-
chological science: opportunities, practical considerations, and challenges. Perspectives on
Psychological Science 11(6): 838–854.
Helberger N (2019) On the democratic role of news recommenders. Digital Journalism 7(8):
993–1012.
Helberger N, Karppinen K and D’Acunto L (2018) Exposure diversity as a design principle for
recommender systems. Information, Communication & Society 21: 2191–2207.
Hill RK (2016) What an algorithm is. Philosophy & Technology 29(1): 35–59.
1274 new media & society 24(5)

Hirsh JB, Kang SK and Bodenhausen GV (2012) Personalized persuasion: tailoring persuasive
appeals to recipients’ personality traits. Psychological Science 23(6): 578–581.
Ishowo-Oloko F, Bonnefon JF, Soroye Z, et al. (2019) Behavioural evidence for a transparency–effi-
ciency tradeoff in human–machine cooperation. Nature Machine Intelligence 1(11): 517–521.
Jobin A, Ienca M and Vayena E (2019) The global landscape of AI ethics guidelines. Nature
Machine Intelligence 1(9): 389–399.
Just N and Latzer M (2017) Governance by algorithms: reality construction by algorithmic selec-
tion on the Internet. Media, Culture & Society 39(2): 238–258.
Kalyanaraman S and Sundar SS (2006) The psychological appeal of personalized content in web
portals: does customization affect attitudes and behavior? Journal of Communication 56(1):
110–132.
Kang H and Sundar SS (2016) When self is the source: effects of media customization on message
processing. Media Psychology 19(4): 561–588.
Kaplan A and Haenlein M (2019) Siri, Siri, in my hand: who’s the fairest in the land? On the inter-
pretations, illustrations, and implications of artificial intelligence. Business Horizons 62(1):
15–25.
Kitchin R (2017) Thinking critically about and researching algorithms. Information, Communication
& Society 20(1): 14–29.
Kleinberg J, Ludwig J, Mullainathan S, et al. (2020) Algorithms as discrimination detectors.
Proceedings of the National Academy of Science 117(48): 30096–30100.
König R, Uphues S, Vogt V, et al. (2020) The tracked society: interdisciplinary approaches on
online tracking. new media & society 22(11): 1945–1956.
Kotkov D, Wang S and Veijalainen J (2016) A survey of serendipity in recommender systems.
Knowledge-Based Systems 111: 180–192.
Kotras B (2020) Mass personalization: predictive marketing algorithms and the reshaping of con-
sumer knowledge. Big Data & Society 7(2): 1–14.
Kunert J and Thurman N (2019) The form of content personalization at mainstream, transatlantic
news outlets: 2010–2016. Journalism Practice 13(7): 759–780.
Lazer D (2015) The rise of the social algorithm. Science 348(6239): 1090–1091.
Lee D and Hosaganar K (2019) How do recommender systems affect sales diversity? A cross-
category investigation via randomized field experiment. Information Systems Research 30(1):
239–259.
Lepri B, Oliver N, Letouzé E, et al. (2018) Fair, transparent, and accountable algorithmic decision-
making processes. The premise, the proposed solutions, and the open challenges. Philosophy
& Technology 31(4): 611–627.
Levy R (2021) Social media, news consumption, and polarization: evidence from a field experi-
ment. American Economic Review 111(3): 831–870.
Lewis SC, Guzman AL and Schmidt TR (2019) Automation, journalism, and human–machine
communication: rethinking roles and relationships of humans and machines in news. Digital
Journalism 7(4): 409–427.
Lokot T and Diakopoulos N (2016) News bots: automating news and information dissemination
on Twitter. Digital Journalism 4: 6682–6699.
Luengo-Oroz M (2019) Solidarity should be a core ethical principle of AI. Nature Machine
Intelligence 1(11): 494.
Lury C and Day S (2019) Algorithmic personalization as a mode of individuation. Theory, Culture
& Society 36(2): 17–37.
McLennan S, Fiske A, Celi LA, et al. (2020) An embedded ethics approach for AI development.
Nature Machine Intelligence 2(9): 488–490.
Hermann 1275

Martin K (2019) Ethical implications and accountability of algorithms. Journal of Business Ethics
160(4): 835–850.
Matz SC and Netzer O (2017) Using big data as a window into consumers’ psychology. Current
Opinion in Behavioral Sciences 18: 7–12.
Matz SC, Appel RE and Kosinski M (2019a) Privacy in the age of psychological targeting. Current
Opinion in Psychology 31: 116–121.
Matz SC, Kosinski M, Nave G, et al. (2017) Psychological targeting as an effective approach to
digital mass persuasion. Proceedings of the National Academy of Science 114(48): 12714–
12719.
Matz SC, Segalin C, Stillwell DJ, et al. (2019b) Using computational methods to predict personal
image appeal. Journal of Consumer Psychology 29(3): 370–390.
Messing S and Westwood SJ (2014) Selective exposure in the age of social media: endorsements
trump partisan source affiliation when selecting news online. Communication Research
41(8): 1042–1063.
Milano S, Taddeo M and Floridi L (2020) Recommender systems and their ethical challenges. AI
& Society 35(4): 957–967.
Milano S, Taddeo M and Floridi L (2021) Ethical aspects of multi-stakeholder recommendation
systems. The Information Society 37(1): 35–45.
Milosavljević M and Vobič I (2019) Human still in the loop. Digital Journalism 7: 81098–81116.
Mittelstadt BD (2019) Principles alone cannot guarantee ethical AI. Nature Machine Intelligence
1(11): 501–507.
Mittelstadt BD, Allo P, Taddeo M, et al. (2016) The ethics of algorithms: mapping the debate. Big
Data & Society 3(2): 1–21.
Möller J, Trilling D, Helberger N, et al. (2018) Do not blame it on the algorithm: an empiri-
cal assessment of multiple recommender systems and their impact on content diversity.
Information, Communication & Society 21(7): 959–977.
Moor JH (1985) What is computer ethics? Metaphilosophy 16(4): 266–275.
Moor JH (2005) Why we need better ethics for emerging technologies. Ethics and Information
Technology 7(3): 111–119.
Morley J, Floridi L, Kinsey L, et al. (2020) From what to how: an initial review of publicly avail-
able AI ethics tools, methods and research to translate principles into practices. Science and
Engineering Ethics 26(4): 2141–2168.
Nechushtai E and Lewis SC (2019) What kind of news gatekeepers do we want machines to be?
Filter bubbles, fragmentation, and the normative dimensions of algorithmic recommenda-
tions. Computers in Human Behavior 90: 298–307.
O’Sullivan PB and Carr CT (2018) Masspersonal communication: a model bridging the mass-
interpersonal divide. new media & society 20(3): 1161–1180.
Potter WJ (2010) The state of media literacy. Journal of Broadcasting & Electronic Media 54:
4675–4696.
Reviglio U (2019) Serendipity as an emerging design principle of the infosphere: challenges and
opportunities. Ethics and Information Technology 21(2): 151–166.
Rich AS and Gureckis TM (2019) Lessons for artificial intelligence from the study of natural stu-
pidity. Nature Machine Intelligence 1(4): 174–180.
Royakkers L, Timmer J, Kool L, et al. (2018) Societal and ethical issues of digitization. Ethics and
Information Technology 20(2): 127–142.
Rudin C (2019) Stop explaining black box machine learning models for high stakes decisions and
use interpretable models instead. Nature Machine Intelligence 1(5): 206–215.
Ryan M (2020) In AI we trust: ethics, artificial intelligence, and reliability. Science and Engineering
Ethics 26(5): 2749–2767.
1276 new media & society 24(5)

Samuel AL (1960) Some moral and technical consequences of automation—A refutation. Science
132(3429): 741–742.
Scharkow M, Mangold F, Stier S, et al. (2020) How social network sites and other online interme-
diaries increase exposure to news. Proceedings of the National Academy of Science 117(6):
2761–2763.
Schwartz SA and Mahnke MS (2021) Facebook use as a communicative relation: exploring the
relation between Facebook users and the algorithmic news feed. Information, Communication
& Society 24(7): 1041–1056.
Selbst AD and Powles J (2017) Meaningful information and the right to explanation. International
Data Privacy Law 7(4): 233–242.
Singer JB (2014) User-generated visibility: secondary gatekeeping in a shared media space. new
media & society 16(1): 55–73.
Soffer O (2019) Algorithmic personalization and the two-step flow of communication.
Communication Theory 2019: qtz008.
Stachl C, Au Q, Schoedel R, et al. (2020a) Predicting personality from patterns of behavior col-
lected with smartphones. Proceedings of the National Academy of Science 117(30): 17680–
17687.
Stachl C, Pargent F and Hilbert S (2020b) Personality research and assessment in the era of
machine learning. European Journal of Personality 34(5): 613–631.
Stahl BC, Andreou A, Brey P, et al. (2021) Artificial intelligence for human flourishing—Beyond
principles for machine learning. Journal of Business Research 124: 374–388.
Stroud NJ (2010) Polarization and partisan selective exposure. Journal of Communication 60(3):
556–576.
Sundar SS (2020) Rise of machine agency: a framework for studying the psychology of human–AI
interaction (HAII). Journal of Computer-Mediated Communication 25(1): 74–88.
Sundar SS and Marathe SS (2010) Personalization versus customization: the importance of agency,
privacy, and power usage. Human Communication Research 36(3): 298–322.
Sunstein CR (2016) The Ethics of Influence: Government in the Age of Behavioral Science.
Cambridge: Cambridge University Press.
Taddeo M and Floridi L (2018) How AI can be a force for good. Science 361(6404): 751–752.
Theodorou A and Dignum V (2020) Towards ethical and socio-legal governance in AI. Nature
Machine Intelligence 2(1): 10–12.
Thiebes S, Lins S and Sunyae A (2020) Trustworthy artificial intelligence. Electronic Markets.
Epub ahead of print 1 October 2020. DOI: 10.1007/s12525-020-00441-4.
Thurman N and Schifferes S (2012) The future of personalization at news websites. Journalism
Studies 13(5-6): 775–790.
Thurman N, Lewis SC and Kunert J (2019a) Algorithms, automation, and news. Digital Journalism
7(8): 980–992.
Thurman N, Moeller J, Helberger N, et al. (2019b) My Friends, editors, algorithms, and I. Digital
Journalism 7(4): 447–469.
Turilli M and Floridi L (2009) The ethics of information transparency. Ethics and Information
Technology 11(2): 105–112.
Turow J and Couldry N (2018) Media as data extraction: towards a new map of a transformed
communications field. Journal of Communication 68(2): 415–423.
Vamplew P, Dazeley R, Foale C, et al. (2018) Human-aligned artificial intelligence is a multiob-
jective problem. Ethics and Information Technology 20(1): 27–40.
Wiener N (1960) Some moral and technical consequences of automation. Science 131(3410):
1355–1358.
Hermann 1277

Willson M (2017) Algorithms (and the) everyday. Information, Communication & Society 20(1):
137–150.
Winter S, Maslowska E and Vos AL (2021) The effects of trait-based personalization in social
media advertising. Computers in Human Behavior 114: 106525.
Wright D (2011) A framework for the ethical impact assessment of information technology. Ethics
and Information Technology 13(3): 199–226.
Zarouali B, Dobber T, De Pauw G, et al. (2020) Using a personality-profiling algorithm to
investigate political microtargeting: assessing the persuasion effects of personality-tai-
lored ads on social media. Communication Research. Epub ahead of print 20 October
2020. DOI: 10.1177/0093650220961965.
Zuiderveen Borgesius FJ, Trilling D, Möller J, et al. (2016) Should we worry about filter bubbles?
Internet Policy Review 5(1): 1–16.

Author biography
Erik Hermann is post-doctoral researcher at the IHP—Leibniz-Institut für innovative
Mikroelektronik. He obtained his PhD from European University Viadrina. His research focuses
on the adoption and (ethical) implications of artificial intelligence in marketing, communication,
and research and development from both the individual and more holistic perspective.