IHS Working Paper 6
April 2019
Artificial Intelligence: Socio-Political
Challenges of Delegating Human
Decision-Making to Machines
Robert Braun
Author(s)
Robert Braun
Editor(s)
Beate Littig
Title
Artificial Intelligence: Socio-Political Challenges of Delegating Human Decision-Making
to Machines
Institut für Höhere Studien - Institute for Advanced Studies (IHS)
Josefstädter Straße 39, A-1080 Wien
T +43 1 59991-0
F +43 1 59991-555
www.ihs.ac.at
ZVR: 066207973
License
„Artificial Intelligence: Socio-Political Challenges of Delegating Human Decision-Making
to Machines“ by Robert Braun is licensed under the Creative Commons: Attribution 4.0
License (http://creativecommons.org/licenses/by/4.0/)
All contents are without guarantee. Any liability of the contributors of the IHS from the
content of this work is excluded.
All IHS Working Papers are available online:
https://irihs.ihs.ac.at/view/ihs_series/ser=5Fihswps.html
This paper is available for download without charge at: https://irihs.ihs.ac.at/id/eprint/5043/
ARTIFICIAL INTELLIGENCE: SOCIO-POLITICAL CHALLENGES OF DELEGATING HUMAN
DECISION-MAKING TO MACHINES
ROBERT BRAUN
1
Abstract: Artificial intelligence is at the heart of current debates related to ethical, social and
political issues of technological innovation. This briefing refocuses attention from the technoethical
challenges of AI to artificial decision-making (ADM) and the questions related to
delegating human decisions to ADM. It is argued that (a) from a socio-ethical point of view
the delegation is more relevant than the actual ethical problems of AI systems; (b) instead of
traditional responsible AI approaches focusing on accountability, responsibility and
transparency (ART) we should direct our attention to trustworthiness in the delegation
process; and (c) trustworthiness as a socio-communicational challenge leads to questions that
may be guided by a responsible research and innovation framework of anticipation,
reflexivity, inclusion, and responsiveness. This may lead to different questions policymakers
and other interested publics may ask as well as novel approaches, including regulatory
sandboxes and other measures to foster a more inclusive, open and democratic culture of
human-ADM relations.
Key words: AI, arithmetic decision-making, delegation, Arendt, RRI.
JEL codes: M14; M31; O31; O32; O33.
Acknowledgement: I would like to thank my colleagues Johannes Starkbaum, Tamara Brandstätter, Thomas
König, Helmut Honigmayer, the participants of our regular Techno Science and Societal Transformation
research group seminar and the anonymous reviewer(s) for their valuable comments to earlier drafts of this
working paper.
1
Institute for Advanced Studies, Vienna
Techno Science and Societal Transformation Research Group
2
1. Introduction
Artificial Intelligence (AI) is currently seeing major media and popular interest, significant
attention from regulatory and policy making bodies both on the national and on the European
level, from academia and from society at large. A term coined by at Dartmouth mathematics
professor John McCarthy, artificial intelligence as a research discipline was initiated at the
Summer Research Project of 1956 (McCarthy, Minsky, Rochester, & Shannon, 2006). The
field of AI research was launched not by agreement on methodology, choice of problems or
general theory, but by the shared vision that computers can be made to perform specific tasks
that may be termed as ‘intelligent’. This vision was stated boldly in the proposal for the 1956
conference: “The study is to proceed on the basis of the conjecture that every aspect of
learning or any other feature of intelligence can in principle be so precisely described that a
machine can be made to simulate it” (McCarthy et al., 2006, p. 12). The study and research in
AI from the beginning, beyond problems of mathematics or control and operations theory,
embraced the idea of duplicating human faculties such as creativity, self-improvement, and
language use. AI also remained a branch of computer science that experiments with building
machines that will function autonomously in complex, changing environments. From its
conceptualization AI operated at the confluence of technology and society and required
transdisciplinary approaches to tackle the multifold and multilayered challenges, both
computational and social, that it posed.
From its origins in the 1950s, to early optimistic predictions of its founders and to recent
ethical and economic challenges as well as the growing awareness of the potentially
transformational nature of development, AI has seen its share of ups and downs in public
interest. There has been steady growth in the past 50-60 years in basic AI research, the
availability of massive amounts of data, and vast advances in computing power have brought
us to a unique phase in AI history. It is now up to society in general to shape the future of the
3
developments and the potentially major societal impacts of artificial intelligence. While
artificial intelligence became the generally accepted name of the field to signal its separation
from other computational fields, it has been challenged from the outset. AI is a fuzzy term
that encompasses a wide range of controversial ideas invoking connotations of human-like
autonomy and intentionality as well as an ‘artificial-natural’ dichotomy. AI involves ideas
about the possibility of machines thinking and acting humanely and/or acting and thinking
rationally as well as a number of other key issues, besides those of computer science, ranging
from the philosophy of the mind to probabilistic mathematics, from decision theory studied in
economics to complex issues in neuroscience, psychology and linguistics (Perez, Deligianni,
Ravi, & Yang, 2018).
As this working paper concentrates on the societal and human aspects of the wide field of AI
it will focus on ‘arithmetic decision-making’ (ADM) as opposed to artificial intelligence (AI)
in general. Arithmetic decision-making is part of the fuzzy field of artificial intelligence
research: it deals with the challenges of data-driven machine learning-based algorithms
tackling complex problems (Gillespie, 2014; Lepri, Oliver, Letouzé, Pentland, & Vinck,
2017; Willson, 2016). This limitation better describes the main social problems at hand.
Arithmetically controlled decision-making is a procedure in which decisions are partly or
completely delegated – via other persons or corporate entities – to automatically executed
decision-making models to perform an action. This action, then, will have deep social
implications from hiring, firing and substituting human labor to addressing social
contingencies in mechanized movement on the road and, potentially, in the air, just to name a
few applications being researched as we speak. Arithmetic decision making is used in setting
up schemes to aid the unemployed, to design robotic vacuum cleaners, to create face
recognition systems and many other technologies that are currently being experimented with.
4
AI is perceived to make ‘intelligent’ decisions modeled on the human mind. This poses a
number of social, legal and philosophical challenges; especially as such decisions involve
actions that contain ethical assumptions and moral consequences. The most well-known
example of the ethical questions posed to AI is the application of the trolley problem to
autonomous vehicles driven by AI. The trolley problem is the ethical model of moral
decision-making when a runaway tram may kill one person or five on a railway depending on
the choice of the onlooker to potentially alter its way (Bonnefon, Shariff, & Rahwan, 2016;
Foot, 1978; Goodall, 2014; Thomson, 1985)). In its basic version the model describes a
situation in which a tram gets loose and reaches an intersection where its route may be
altered. If it continues it kills five people tied to the rail, whereas if a switch is applied the
tram changes track and on that track kills one person tied to the track. There are other
versions with options to stop the tram by pushing someone on the track, or deciding between
tracks based on different characteristics, not the number of people to be killed by the choice.
The dilemma(s) may be resolved by choosing different outcomes based on the ethical theory
one subscribes to. The trolley problem points to the moral dilemma of AI: there is no one
right answer, as there is no one ethical theory or moral outcome that may be applied to any
specific case or circumstance. AI or arithmetic decision making in general cannot be
‘ethical’: certain decisions may be made according to specific ethical theories that work
(morally) better (or are more acceptable to a community) in certain cases than do other
potential choices. These technologies are also parts of socio-technical systems, such as
autonomous mobility-as-a-service provision or autonomous stock trading methods, health
diagnostics arrangements or corporate sales and marketing endeavors that offer complexities,
both moral and social, that there is no linear or path determined way to tackle.
5
2. Framing arithmetic decision-making
In the plethora of politico-social challenges related to AI our attention here will focus on the
delegation of human decisions to data-driven, arithmetically controlled systems. The
delegation is part of a socio-technical framework in which ADM is utilized. This includes a
delegation process, a decision-making model, an algorithm that translates the model into a
computational code, the sets of data and its modes of collection the code uses as input, the
‘learning’, ‘processing’ and ‘analysis’ procedure the code utilizes and the entire social,
economic and political environment that surrounds the delegation process. Also, the
consideration and decision to apply ADM for a certain purpose, its development,
procurement and deployment are parts of the framework to be analyzed.
I will look at ADM from a techno-social perspective, and discuss the social context of ADM
technology as well as the techno-political consequences of delegating human decisions to
machines. Questions of politics have long been part of science and technology studies (STS),
and STS scholars have carefully studied how technology became intertwined with politics
(Brown, 2015). Here I will focus on the socio-political challenges of delegation and discuss
how research and innovation in ADM should approach these challenges as well as what are
the options of regulating ADM and the process of delegation. Delegation is this context is not
primarily a question of technology, but a techno-political problem that relates to the sociotechnical
‘imaginary’ as well as the ‘techno-political framework’ of ADM. By ‘imaginary’ I
mean a commonsense understanding of the delegation process: a framework that contains
elements of a shared vision created, sustained and reproduced through rhetoric and power.
Like other worldviews, imaginaries “become so deeply embedded in commonsense
understandings that they are taken for granted and beyond question” (Harvey 2017, 24). This
taken-for-grantedness animates technology innovation processes as well as policies,
regulation and decisions on multiple levels, from individuals to governments.
6
The ‘techno-political framework’ refers to the complex interrelatedness of the principal (be it
an individual or an organization) and the agent (the ADM operation of a technology) in the
delegation process, as well as the socio-political context the delegation takes place and the
multifold impact of this delegation and the actions that follow on society. Imaginary and
framework are interrelated as the imaginary creates a commonsense understanding of takenfor-grantedness
that using ADM is ‘better’ (whether ‘better’ is defined from eg. a neoliberal
efficiency perspective (Means, 2015) or from a securitization angle (Yampolskiy, 2018)) than
human decision-making is in specific contexts. This provides the rationale of delegation for
the principal and suppresses concerns impacting society. This approach is exemplified in the
EU communication on AI for Europe (date: 25.04.2018) stating that “[b]eyond making our
lives easier, AI is helping us to solve some of the world's biggest challenges: from treating
chronic diseases or reducing fatality rates in traffic accidents to fighting climate change or
anticipating cybersecurity threats” (EC, 2018b, p. 1). Most communication on AI
acknowledge that “[s]ome AI applications may raise new ethical and legal questions, related
to liability or fairness of decision-making” (EC, 2018a), however, they also claim that
addressing these ‘problems’ head on will make them manageable and our lives can be made
easier by ADM.
Today the emphasis of ‘responsible AI’, a term used to tackle the socio-ethical challenges of
ADM, is on the algorithmic decision-making itself and not the relationship between humans
and ADM or the conditions when and if delegation should or should not take place. The
confluence of taken-for-grantedness of delegation and the assumption that technologies are
(always) socially fit for purpose create a fascination with the idea of AI induced technological
progress. The constant search for a perceived betterment, whether of an individual or of
society, is underscored by the belief in ADM: ‘solutionism’ provides the drive to both
innovate and delegate. Innovation is framed as the unsolved not-yet-innovated (Dewandre,
7
2018) and the illusion that there is no barrier to what can be achieved by delegating our
human decisions to ADM. This casts the present in terms of deficit: what is lacking, what is
not proper and the future where ‘solutions’ (perfect or at least better decisions) lie. It also
creates a structural dissatisfaction coupled with a deeply anchored impatience. This
impatience, then, is the driver of delegation, but also the reason to put aside societal
considerations in exchange of ADM solutionism.
While marginally related, this impatience is exemplified for instance in the current
enthusiasm in transforming our education system to offer more STEM (science, technology,
engineering and mathematics) as opposed to liberal arts and social science (Benson, 2017).
The potential reordering of our systems of education may seriously inhibit our abilities to
tackle the social challenges, ethical concerns and humane aspects of innovation in ADM and
beyond. Exclusive STEM education (without social or humanities awareness) will create
‘ADM subjects’ -- rational principals indoctrinated by enthusiasm for technological fixes and
betterment via engineering solutionism2
. Ethical or social alarms as well as delegation
concerns will be removed; ADM subjects will learn to ‘think’ like machines and machines
will learn to think like STEM humans. One should be reminded that dystopias are not created
by making machines that take over humans, but by creating humans that are not capable of
controlling machines as all skills required for this control (ethics, social and political
awareness) are unlearned.
Taken-for-grantedness permeates public discourse on an ‘expert’ and on a commonsense
level as communities of power sharing similar epistemological assumptions make
solutionism-visions collectively held and institutionally stabilized (Jasanoff & Kim, 2015).
2
This is manifested in the findings of our research looking at H2020 funding principles in specific technological
innovation program lines. Cf.: Akca Prill, Melek, Lindner, Ralf, Allinger, Matthias, Bernstein, Michael J.,
Bratan, Tanja, Braun, Robert, Gianni, Robert, Goos, Kerstin, Ikonen, Veikko, Schrammel, Maria,
Nieminen, Mika, Seebacher, Lisa Marie, Tumbrägel, Tessa, Tyynelä, Janika and Wunderle, Ulrike (2018)
New HoRRIzon Deliverable 4.1. Diagnosis: RRI in Societal Challenges. [Research Report] 256 p.
http://irihs.ihs.ac.at/4918/
8
We are witnessing techno-optimists and the industrial-innovation complex formulating AI
enthusiastic epistemic communities of power (Antoniades, 2003) acting in concert to create
the imaginary of AI inevitability in all possible domains of life. This working paper entertains
the idea that the moment as AI emerges from the research labs to be applied in commonly
used sociotechnical artifacts (vehicles, health diagnostic instruments and commercial
systems) generates the urgency to scrutinize the overall taken-for-grantedness of delegation.
This paper refocuses attention from asking the question ‘How can ADM (made to) be better?’
to the complementary question of ‘How to make our decisions better on when, why and how
to delegate human decisions to ADM?’
The remainder of the paper will first briefly discuss how delegation is dealt with in political
science, business and marketing literature. Next I will turn to current socio-political
challenges to ADM, and critically discuss the responsibility, accountability, transparency
framework used by ‘responsible AI’ scholars to address these challenges. Reframing
responsibility as trustworthiness, I will show how a trustworthy ADM approach may be more
appropriate to address the political problems related to ADM. It is argued that responsible AI
focuses on epistemic ideals and procedures of ADM, while trustworthy ADM delegation
concentrates on the socio-political context of delegation, the plurality and diversity of
potential outcomes as well as the risks involved in the process of delegation. The paper will
end in discussing the relevance of this to changing regulatory practices as well as draw some
wider political conclusions related to our technofutures.
It is clear that the questions of ADM betterment and improving our judgment on delegation
are interrelated, however not straightforwardly following from each other. The first question,
important in its own right, focuses on the socio-technical appropriateness of the arithmetic
decision making system as such, the second question concerns our relationship to ADM as
well as the consequences of this relationship (embodied in the delegation) on others. While
9
the first question addresses the ‘epistemic’ and ‘moral’ qualities of ADM (and our judgments
about these epistemic and moral qualities), the second focuses on the political aspects of
ADM: the embeddedness of ADM in societies and our relationship to each other. The core
(political) challenge is this: What our polis will look like and how we will deal with the
socio-ethical problems when we delegate many (or most) of our socially relevant human
decisions (from mobility to health and beyond) to machines with ADM.
3. Delegation
There is a wide array of research available in different areas about impacts, challenges and
rationale of delegation in political science, marketing and management studies (Aggarwal &
Mazumdar, 2008; Bell & Bodie, 2012; Sengul, Gimeno, & Dial, 2012 ). Some of the findings
may be applicable to our inquiry in delegating human decisions to ADM. In political science
legislative delegation to the bureaucracy is criticized as diminishing electoral accountability
and exacerbating legislative shirking. In electoral theories of delegation the general consensus
is that optimal political representation consists of a mixture of the delegate (no discretion)
and trustee (full discretion) models of representation. One of the most important findings is
that whatever model of delegation (ally, credible commitment, strategic, etc.) is applied, such
models do not take into consideration the decisional as well as the social context of
delegation. Decisional contexts involve the fact that principals delegate with an ongoing
authority and not design their delegation strategies on a blank slate, while agents also have
some already existing level of agency or de facto authority (Bendor, Glazer, & Hammond,
2001). In management literature the principal-agent problem (a version of delegation of
authority in a corporate arrangement) is seen as resulting in moral hazard and conflicts of
interest (related to the costs of controlling the agent and the potential benefits thereof, and/or
the cost accrued by the agent to appropriately represent the interests of the principal) (Jensen
10
& Meckling, 1976). The principal-agent relationship generates social uncertainties tackled by
regulating the relationship through individual actions, rule based processes or policy and
regulatory attention. In traditional principal-agency setups agents are motivated to comply
with the interests of principals through incentive mechanisms. In recent research on the
behavioral aspects of the principal-agent problem principal side overconfidence as behavioral
bias emerged as expected utility destructive (de la Rosa, 2011). In marketing literature
delegation of consumer choice is seen as reducing felt responsibility for negative choice
outcomes, and is mainly used when surrogates are perceived as knowledgeable and
trustworthy. (Aggarwal & Mazumdar, 2008)
At this point it suffices to say that the problem of delegation of decisions under risk and
uncertainty by differing actors is not new, therefore there is no need to reinvent the
delegation-theory wheel. We may also see that human decision-making are complex setups of
personal, societal and cultural determinations individualized by context, personal weighing of
expected utility and a number of biases and heuristics. There is no one solution or applicable
theory that says it all, therefore the analysis of delegation needs to focus on context and
available delegation options as well. The core of delegation, the principal-agent problem, is
not addressed via assessing the epistemic qualities of the agent, but by her intent to represent
the interests of the principal suitably. This marks a shift from epistemology to a political
understanding of delegation. Principal side overconfidence and the consequences thereof
remind us to pay special attention to brashness aka taken-for-grantedness.
4. Political and ethical challenges of delegation in ADM
As advances in decisions to delegate human decisions to ADM take place at a growing pace a
set of questions arise related to social, economic, political, technological, legal and
philosophical issues (Dignum, 2017). Beyond ethical considerations, societies have not yet
worked out ways to deal with developing new technologies that utilize machine learning,
11
multi-dimensional connectivity, multi-layered data collection to fit societal concerns and
expectations.
(A) Accountability, responsibility and transparency
In order for ADM to pass as ‘intelligent’ responsibility is argued to have to be an inbuilt
feature. ‘Responsible AI’ theorists suggest that AI systems should be augmented with the
principles of accountability, responsibility and transparency (ART) as driving principles of a
human centered and intelligent system (Dignum, 2017). Accountability refers to the need to
explain and justify decisions and actions to users and other stakeholders with whom the
system interacts and is ensured if decisions are derivable from, and explained by, the ADM
used. Responsibility refers to the capability of ADM systems to answer for their decisions, to
identify errors or unexpected results. This also means the need to link ADM to the fair use of
data and to the actions of stakeholders involved in the system’s decision. Transparency refers
to the need to describe, inspect and reproduce the mechanisms through which AI systems
make decisions and learn to adapt to their environment, and to the governance of the data
used or created. Thus, according to responsible AI, methods are required to inspect
algorithms and their results and to manage data, their provenance and their dynamics
(Bostrom & Yudkowsky, 2014).
ART is problematic and may not be instructive in understanding the politics of ADM.
Algorithms in ADM tend to be ‘black boxes’: devices that can be viewed in terms of inputs
and outputs, but principals often have no knowledge of their internal workings. As algorithms
that enable ADM to make decisions and navigate the complexities of their environments
become more specialized and complex, even creators may no longer be able to understand
them (Stilgoe, 2018). ADM is tasked to engage with complexities that cannot be captured by
a set of simple and formal rules. Deep learning mechanisms operate probabilistic setups of
nonlinear transformations on input to reach an acceptable level of accuracy of output. ADM
12
systems using such probabilities unsupervised create social uncertainties that, by design,
make algorithmic decision outcomes inscrutable (Bornstein, 2016).
This is troublesome from a political ADM perspective. Politics, from a technoscience
perspective, may be defined “as purposeful activities that aim for collectively binding
decisions in a context of power and conflict” (Brown, 2015, p. 19). All elements of this
definition sketch (human agency, purposefulness, decisions, power and conflict) require
certain level of transparency related to how decisions are arrived at, what is the context of
such decisions and what room the agent has to maneuver in arriving at collectively binding
power setups to resolve conflict(s). Once decisions become inscrutable they may offer a
technology fix to specific challenges, however the conflicts involved in finding solutions are
resolved without the possibility of human agency (purposefulness) involved. This may be
attractive to those who see technology as essentially ‘natural’ and path determined, however,
from an STS perspective it annuls the well documented ‘mutual constitution’ of society and
technology (Jasanoff, 2004). Apolitical approaches deprive technosocial situations or events
of relations of power, justice, morality, and group identity as well as deny its value-laden,
therefore contestable, nature. ADM thus is intertwined with a technological essentialism that
abstracts technology from society.
A ‘right to explanation’ (Kaminski, 2018) (information about individual decisions made by
algorithms) is called to help increase the accountability and transparency of the ADM
process. However ‘right to explanation’ may be problematic as machine learning and ADM
mirrors the architecture of (our current knowledge of) (Mercier & Sperber, 2017) human
brains by building complex and multilayered representations of information. These
representations are sometimes delusive (misreading certain information that it has mistakenly
categorized) and may cause ‘mental health’ problems similar to hallucination or other social
disorders like schizophrenia and Asperger (Hills, 2018).
13
(B) Trustworthiness
Trustworthiness is a key factor of interpersonal relations that enable delegation (Homburg &
Stock, 2005; Sargeant & Lee, 2004). Trustworthiness may be defined as a function of assured
positive anticipations in situations involving risk. Trustworthiness is generally discussed in
terms of social or contractual accountability, as well as expertise involving responsibility and
transparency (Keating & Thrandardottir, 2016). Some theorists of trustworthiness, however,
claim that trustworthiness is not necessarily based on ART principles, but on the perceived
honesty, competence and reliability of the other person (O’Neill, 2018). Honesty would
address the claims and commitments made, competence entails the perceived ability to
perform the relevant task at hand and reliability is provided to enable those communicated
with to reach an intelligent judgement. Trustworthiness in this rendering is situational and
communicational – it is about the institutional and social context and not the actual
knowledge involved in the perceived delegated action. It is also political as it is relational: it
is about assured positive anticipations of the principal about the agent in situations involving
risk.
The idea of relationality is important to address the political nature of delegation. Traditional
delegation frames theorize agents (both as principals and as surrogates) as ‘rational’ and
conceptualize their relationship in terms of dominance and interest induced conflict.
Delegation is based on some form of cost-benefit analysis and the implications of such
analysis to the management of the relation to reach optimal outcomes. Optimal outcomes,
even if social contingencies are accounted for, are framed based on starting hypotheses of
short or long term interests of principals and agents, or the collective ideals of the social
contexts they are embedded in. In ideal delegation setups rational agents weigh costs and
benefits, analyse and adjust to contexts to arrive at utility maximizing outcomes. The
relational agent, a concept originating in Hannah Arendt’s conceptualization of the human
condition as essentially rooted in plurality, offers an alternative route.
14
As opposed to the rational, the relational agent is aware of the fact that “action [is] the only
activity that goes on directly between men without the intermediary of things or matter,
corresponds to the human condition of plurality, to the fact that men, not Man, live on the
earth and inhabit the world” (Arendt, 1958, p. 9). Plurality or relationality is a concept that
challenges the modern omniscience-omnipotence utopia of ‘optimal outcomes’ and enables
embracing diversity of the human condition both in terms of the diversity of individuals and
the diversity of social contexts. It is also important that relational agents are aware of social
interdependence and of the fact that there is no ultimate guarantee for a given optimal
outcome as “relational selves are conscious of experiencing the ‘calamities of action’ arising
from plurality” (Dewandre, 2018, p. 511). The calamities of action are essential to the human
condition as doing away with plurality leads to the abolition of the public realm altogether
(Arendt, 1958).
Politicizing ADM delegation in terms of trustworthiness means that we refocus on relational
as opposed to rational agents embedded in the human condition of plurality of outcomes and
the diversity of social situations. A responsible AI framed delegation seeks ‘optimal
outcomes’ and focuses on epistemic ideals and procedures of ADM to achieve such
outcomes. Trustworthy ADM delegation, in turn, focuses on the socio-political context of
delegation, the plurality and diversity of potential outcomes as well as the risks involved in
the process of delegation. It also embraces the diversity of potential principals as well as
understands the contingencies involved in human decision making stemming from the
diversity and uniqueness of identities, a basic human condition. This then allows for a more
nuanced and heuristic approach to delegation, one that first looks at context, potential
impacts, then rationale of delegation. Trustworthy ADM delegation also keeps the option of
non-delegation open, even if delegation seems to be utility maximising on a cost-benefit
basis, but utility diminishing socially or politically.
15
One of the socio-political contexts of machine learning in ADM to be addressed is
democratizing the process of (social) learning. Machine learning advances are to be made
public and shared, not to be kept proprietary to one company or technology provider. ADM
applications are not only a set of engineering ‘tasks’ but also impact socialities that, during
the delegation process, require relational awareness. This entails that addressing ADM
challenges means being attentive not only to risks and challenges of the new technologies,
but also to public concerns as to how and why specific innovations happen in ADM systems;
why and how human decisions are transferred to machines and whether keeping human
decision-making in certain contexts instead of delegating decisions to ADM would actually
lead to other, alternative risk eliminating social impacts.
A good example of a trustworthy AI approach is the European Ethics Guidelines on
Trustworthy AI created by the High Level Expert Group on Artificial Intelligence. It
embraces many of the ideas including ethical purpose, stakeholder involvement and constant
reflection expressed in this working paper (EC, 2018c). The guidelines focus on humancentric
AI stating that should be developed, deployed and used with an ethical purpose,
grounded in, and reflective of, fundamental rights, societal values and the ethical principles of
beneficence (do good), non-maleficence (do no harm), autonomy of humans, justice, and
explicability. It also focuses on paying appropriate attention to situations involving
vulnerable groups and to situations with asymmetries of power or information. These are
essential to the delegation of decision making from humans to ADM. It calls attention to
incorporate data governance, design for all principles, human oversight when required, nondiscrimination
and respect for privacy from the earliest design phase. The document also
suggests providing, in a clear and proactive manner, information to stakeholders about the
system’s capabilities and limitations, allowing them to set realistic expectations of the
16
capabilities and potentials of the ADM operated AI. This is crucially important to assess
when and for what end to delegate.
However, the document falls short of mentioning that non-delegation is also an option for
specific contexts and by specific stakeholders. The document calls attention to make
trustworthy AI part of the organization’s culture, and provide information to stakeholders on
how trustworthy AI is implemented into the design and use of AI systems. The document also
emphasizes to ensure participation and inclusion of stakeholders in the design and
development of the AI system. The document provides an excellent overview of how, for
what purpose and by which means trustworthiness may become part of AI/ADM. It comes
short however, of drawing the attention to potential barriers and pitfalls of the delegation
process as well as the contexts where delegation is beneficial and where it may or should be
avoided.
(C) Responsible & trustworthy ADM
When applying basic principles of trustworthiness to ADM, honesty as claims and
commitments of ADM may mean that ADM development avoids an overall technology-fix
approach that looks at social challenges as requiring a better (more efficient, faster, safer)
technology. These concepts are socially loaded and may lead to unintended social
consequences. In case of ADM public deliberation on the specific technologies, their benefits
and pitfalls, wellbeing impacts as well as their alternatives should be evaluated. Disruptive
technologies, ADM included, claim to offer remedies to past social pathologies of
technological development, such as inequality, social exclusion or ethical dilemmas. This
may, once again, lead to solutionism.
The concept of competence as the perceived ability to perform relevant task means that
delegation of human decisions to ADM requires new regimes of governance that do not take
technology as fixed and seen to provide solutions to deficits of human behaviour,
17
infrastructure or the law. Technological systems should be shaped actively, including the
(sociotechnical) imaginaries that animate them and (re)connecting ADM with their social and
natural environment in multiple ways and forms. This would also mean that keeping nondelegation
options open is also part of the assessment of competence. Involving different
stakeholders early on in the innovation process would create a more open discussion about
what domains and areas should ADM technology be applied to. This is extremely important
as the delegation of human decision making to ADM becomes unavoidable once a
technology has been developed and emerged as market ready. When considering this
question, responsible innovators as well as other stakeholders – policy makers, funders and
regulators in ADM should first reflect on the research and innovation process itself and
openly discuss what roles ADM run artefacts should or should not play in society.
A more open, anticipatory and democratic approach to ADM would ask what different
publics would want from ADM – where do they think delegating decisions may be useful.
This would require that these publics have a better understanding what ADM can and cannot
do, while openly conversing about the potential intended and unintended social impacts from
job loss to creating lethal autonomous robots for terrorist operations. Such programs would
also empower and engage the public in debates about technology futures, require substantial
learning on their side to have an educated conversation. Public engagement would offer
inclusion of different points of views as well as different aspects and arguments for the social
desirability of specific ADM systems and their governance, including collaborative practices
of co-design and co-regulation. Reliability as enabling those communicated with to reach an
intelligent judgement would empower responsible innovation in ADM to regard humans as
resource and not as problem.
Claims of ADM that they provide a solution to human errors in driving is living proof. While,
for example, safety is an utmost concern of mobility, in the course of autonomous vehicle
18
ADM innovation, the question ‘safe enough for what?’ should also be posed. Automobility
for instance traditionally ‘accepts’ roadside death and injury as part of the system implicitly
applying a consequentialist ethical position in which the benefits of the car (both in mobility
and in social terms) are assumed to outweigh the malign impacts of accidents, while other
mobility technologies, e.g. elevators, became widespread and societally transformational only
after the technology was considered to be absolutely safe and accidents were eliminated
(Bernard, 2014).
Institutions and individuals need to build and develop appropriate reflexive capacity to
diverge from a technology fix approach and focus on social learning, complex assessments of
impacts and responsiveness to challenges thereof, both in the sense that people learn and
assess impacts socially and that societies learn, reflect and respond constantly, offering ways
to better understand and democratize the social experiment of ADM transition.
19
5. Regulatory sandboxes
The framework and practice of responsible research and innovation (RRI), a concept aiming
at the betterment of the social embeddedness of research and innovation (R&I), may be
applied to ADM. Responsible innovation addresses the growing distrust in scientific
knowledge by challenging the established neoliberal model of R&I in which research and
innovation have been understood as a linear, efficiency and progress driven development,
directly leading from basic research to application, from experts to consumers. It points
towards a research and innovation ecosystem that recognizes non-linearity and addresses
inclusion of different publics with regards to research goals and questions, processes,
outcomes and impacts. RRI, according to one definition, is “taking care of the future through
collective stewardship of science and innovation in the present,” which may be achieved
through applying a framework of anticipation, reflexivity, inclusion, and responsiveness
(Stilgoe, Owen, & Macnaghten, 2013).
Applying an anticipatory and reflective ADM assessment framework would focus on
diversity of potential outcomes and the implications thereof, as well as on the potential of
cooperation as opposed to delegation. Currently ADM is perceived to substitute human
decision making in many areas; an alternative social vision could look at options of
complementing the strength of human intelligence with ADM and other forms of AI. This
would involve complex strategies from education to participatory decision-making, focusing,
as discussed earlier, on options like not introducing more STEM into education but on how to
utilize STEM for arts, humanities and vice versa. The confluence of different forms of arts
and STEM3
as well as the experiments of the Socio-Technical Integration Research
embedding social scientists and humanities scholars in laboratories (Fischer, 2007) provided
3
One example is the experiment, in a different context, in the Joint Research Center: cf.
https://ec.europa.eu/jrc/en/event/exhibition/resonances-science-arts-politics.
20
experimental avenues to bring a more value and human centered design to ADM and AI in
general.
From a regulatory point of view it is important to note that ADM is a technology that is being
tested and deployed, as well as work-in-progress. Therefore ADM poses challenges not only
to researchers but also to policy makers and regulators. One of the main challenges is how to
apply an iterative, anticipatory regulatory practice to the emerging field of ADM. As opposed
to advisory or adaptive regulatory practices, methods usually applied to regulation in
emerging technological fields a more open, anticipatory approach may be applied (Armstrong
& Rae, 2017). This would assure a better understanding of technology’s impact on economy
and society, as well as foresee the constantly changing regulatory needs and vision for the
future. Such regulation would involve a wide array of stakeholders in the regulatory process
including regulators, businesses, cross-industry, civil society, local authorities, cities, citizens,
early adopters and NGOs.
In order to avoid regulatory traps of both policy push, and regulatory blockage regulatory
sandboxes could be created where innovators together with citizens and other stakeholders
may experiment with ADM/AI technologies, involving and engaging knowledge of diverse
stakeholders so innovation in ADM may also be attentive to complex social impacts and
uncertainties. The focus of regulation, besides ethical AI, should also concentrate on the
delegation process and the pre-conditions thereof. In such regulatory sandboxes specifics of
delegation – when, how, why delegation should or should not happen – could be
experimented with, also unearthing intended and unintended social and ethical consequences.
Regulators may also learn and adjust regulatory regimes as work-in-progress. This is
especially important as deployment of ADM requires constant regulatory adaptation. The
challenge is that adaptation may also happen through de-facto regulation, industry initiatives
21
being standardized through industry membership organizations or by early mover practices,
not taking multiple stakeholder perspectives into account.
6. Conclusion
Ethical issues and societal challenges burgeon AI research. Within the social aspects of AI
this paper focused on discussing, based on recent literature, as to when, why and how
delegate human decisions to ADM. This marks a move from the epistemic and ethical
challenges in AI/ADM innovation towards a societal and pluralistic understanding of the
issues of ADM. ADM will become more complex and multifold, thus in most cases it will
constantly improve maneuvering the complexities of decision making and decide making
actions that are perceived to be safer, faster and more reliable. This is good. However,
societies are complex and fuzzy and thus some of the seemingly safer, faster and more
reliable decisions may not to turn out to be beneficial for society as a whole. This is the bynow-old
story of unintended consequences (Blok & Lemmens, 2015).
The emergence of AI/ADM from test labs to general use artefacts means that in more and
more instances we will delegate our human decisions to machines. Beyond unintended
consequences this may change our social world profoundly. Autonomous vehicles, robotic
diagnostic and care systems, arithmetic social platforms will dominate our worlds to make
decisions and act on our behalf. The urge for delegation is animated by the taken-for-granted
imaginary called progress (Benjamin, 2016). Responsible AI focuses on the epistemic and
ethical ideals that may be built into the machines with ADM. Politicizing ADM starts from
the relationality of the human condition and refocuses attention from the inside of the
machines to the outside where they function. Outside is the social context in which ADM
machines operate as well as the space between humans and machines. This is the space where
delegation happens. If the polis, the public realm of the political where humans nourish
22
freedom, is space in-between people (Arendt & Schell, 2006) then we are redirecting our
attention to the new technopolis of in-between of people and machines. Thus the real
challenge of ADM is not ethical or epistemic but, as has also been argued in this paper,
political. Every decision to substitute human decision-making with ADM shrinks the inbetween
of the public realm. Responsible AI would like to save the in-between and thus
recreate the public realm by making machines more humane. Focusing on trustworthiness
aims to tackle the political challenge head on. It tries to keep non-delegation options open by
forcing us to think when delegation of decisions by humans to machines does actually make
the in-between of our technopolis populated by people and machines one notch smaller.
23
References
Aggarwal, P., & Mazumdar, T. (2008). Decision delegation: A conceptualization and
empirical investigation. Psychology & Marketing, 25(1), 71-93.
doi:10.1002/mar.20201
Antoniades, A. (2003). Epistemic Communities, Epistemes and the Construction of (World)
Politics. Global Society, 17(1), 21-38. doi:10.1080/0953732032000053980
Arendt, H. (1958). The Human Condition. Chicago: Chicago University Press.
Arendt, H., & Schell, J. (2006). On Revolution: Penguin Publishing Group.
Armstrong, H., & Rae, J. (2017). A working model for anticipatory regulation. London.
Bell, R. L., & Bodie, N. D. (2012). Delegation, Authority and Responsibility: Removing the
Rhetorical Obstructions in the Way of an Old Paradigm. Journal of Leadership,
Accountability and Ethics, 9(2).
Bendor, J., Glazer, A., & Hammond, T. (2001). Theories of Delegation. Annual reviews of
Political Science, 4, 235-269.
Benjamin, W. (2016). On the Concept of History: Createspace Independent Publishing
Platform.
Benson, J. D. (2017). STEM and the Humanities. Open Access Library Journal, 4(e3476.).
Bernard, A. (2014). Lifted: A Cultural History of the Elevator. New York: NYU Press.
Blok, V., & Lemmens, P. (2015). The Emerging Concept of Responsible Innovation. Three
Reasons Why It Is Questionable and Calls for a Radical Transformation of the
Concept of Innovation. In B.-J. Koops, I. Oosterlaken, H. Romijn, T. Swierstra, & J.
van den Hoven (Eds.), Responsible Innovation 2: Concepts, Approaches, and
Applications (pp. 19-35). Cham: Springer International Publishing.
Bonnefon, J.-F., Shariff, A., & Rahwan, I. (2016). The social dilemma of autonomous
vehicles. Science, 352(6293), 1573-1576.
Bornstein, A. M. (2016). Is Artificial Intelligence Permanently Inscrutable? Nautilus(40).
Bostrom, N., & Yudkowsky, E. (2014). The ethics of artificial intelligence. In K. Frankish &
W. M. Ramsey (Eds.), The Cambridge Handbook of Artificial Intelligence (pp. 316-
334). Cambridge: Cambridge University Press.
Brown, M. B. (2015). Politicizing science: Conceptions of politics in science and technology
studies. Social Studies of Science, 45(1), 3-30. doi:10.1177/0306312714556694
de la Rosa, L. E. (2011). Overconfidence and moral hazard. Games and Economic Behavior,
73(2), 429-451. doi:https://doi.org/10.1016/j.geb.2011.04.001
Dewandre, N. (2018). Political Agents as relational Selves: rethinking EU Politics and
Policy-Making with Hannah Arendt. Philosophy Today, 62(2), 493-519.
Dignum, V. (2017). Responsible Artificial Intelligence: Designing for Human Values. ITU
Journal: ICT Discoveries(Special Issue No. 1, ).
EC. (2018a). Artificial Intelligence. Brussels: European Commission Retrieved from
https://ec.europa.eu/digital-single-market/en/artificial-intelligence#" name=.
EC. (2018b). Artificial Intelligence for Europe. Brussels: European Commission.
EC. (2018c). ETHICS GUIDELINES FOR TRUSTWORTHY AI. Brussels: European
Commission.
Fischer, G. (2007). Designing socio-technical environments in support of meta-design and
social creativity. Paper presented at the Proceedings of the 8th iternational conference
on Computer supported collaborative learning, New Brunswick, New Jersey, USA.
Foot, P. (1978). The Problem of Abortion and the Doctrine of the Double Effect in Virtues
and Vices Oxford: Basil Blackwell.
24
Gillespie, T. (2014). The relevance of algorithms. In T. Gillespie, P. Boczkowski, & K. Foot
(Eds.), Media technologies: Essays on communication, materiality, and society (pp.
167–193). Cambridge, MA: MIT Press.
Goodall, N. J. (2014). Machine Ethics and Automated Vehicles Road Vehicle Automation
(pp. 93-102). London: Springer.
Hills, T. T. (2018). The Dark Side of Information Proliferation. Perspectives on
Psychological Science, 1745691618803647. doi:10.1177/1745691618803647
Homburg, C., & Stock, R. M. (2005). Exploring the conditions under which salesperson work
satisfaction can lead to customer satisfaction. Psychology & Marketing, 22(5), 393-
420. doi:10.1002/mar.20065
Jasanoff, S. (2004). States of knowledge: the co-production of science and the social order:
Routledge.
Jasanoff, S., & Kim, S.-H. (2015). Dreamscapes of Modernity. Chicago: Chicago University
Press.
Jensen, M. C., & Meckling, W. H. (1976). Theory of the Firm: Managerial Behavior, Agency
Costs and Ownership Structure. Journal of Financial Economics, 3(4).
Kaminski, M. (2018). The Right to Explanation, Explained.
https://doi.org/10.31228/osf.io/rgeus
Keating, V. C., & Thrandardottir, E. (2016). NGOs, trust, and the accountability agenda. The
British Journal of Politics and International Relations, 19(1), 134-151.
doi:10.1177/1369148116682655
Lepri, B., Oliver, N., Letouzé, E., Pentland, A., & Vinck, P. (2017). Fair, Transparent, and
Accountable Algorithmic Decision-making Processes: The Premise, the Proposed
Solutions, and the Open Challenges. Philosophy & Technology, 31.
McCarthy, J., Minsky, M. L., Rochester, N., & Shannon, C. E. (2006). A Proposal for the
Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955. AI
Magazine, 27(4).
Means, A. (2015). On accelerationism- decolonizing technoscience through critical
pedagogy. Journal for Activism in Science & Technology Education, 6(1), 21.
Mercier, H., & Sperber, D. (2017). The Enigma of Reason. Cambridge, MA: Harvard
University Press.
O’Neill, O. (2018). Linking Trust to Trustworthiness. International Journal of Philosophical
Studies, 26(2), 293-300. doi:10.1080/09672559.2018.1454637
Perez, J. A., Deligianni, F., Ravi, D., & Yang, G.-Z. (2018). Artificial Intelligence and
Robotics. Retrieved from London:
Sargeant, A., & Lee, S. (2004). Trust and relationship commitment in the United Kingdom
voluntary sector: Determinants of donor behavior. Psychology & Marketing, 21(8),
613-635. doi:10.1002/mar.20021
Sengul, M., Gimeno, J., & Dial, J. (2012 ). Strategic Delegation: A Review, Theoretical
Integration, and Research Agenda. Journal of Management, 38(1), 375-414.
Stilgoe, J. (2018). Machine learning, social learning and the governance of self-driving cars.
Social Studies of Science, 48(1), 25-56.
Stilgoe, J., Owen, R., & Macnaghten, P. (2013). Developing a framework for responsible
innovation. Research Policy, 42(9), 1568-1580.
Thomson, J., J. (1985). The Trolley Problem. Yale Law Journal, 1395–1415.
Willson, M. (2016). Algorithms (and the) everyday. Information, Communication & Society.
Yampolskiy, R. V. (2018). Artificial Intelligence Safety and Security. London: CRC Press.