Individual and organizational actions have social consequences that call for the implementation of recommendations of good conduct at multiple levels of granularity. This work focuses on the accountability value, on how to express it, and on the challenges of handling accountability in a computational way, in a system of interacting agents.
In recent years we have seen a growing attention of the scienti c community towards the theme of Arti cial Intelligence (AI) and Ethics. In 2016, the major conferences on AI (IJCAI, ECAI, AAAI) proposed in their programs workshops that focused on the societal implications of building AI systems. Questions were posed and issues discussed about the adoption of AI techniques and methodologies in search engines, self-driving cars, electronic markets, smart homes, military technology, big data analysis, care robots. Similar discussions reached a somewhat more mature stage of development within the scienti c community that works on Robotics, where the focus mainly rests on the ethical design and application of robots and robotic systems and on the legal implications of using hardware or software devices. Signi cant in this respect is the recent BS 8611:2016 standard, published by BSI in April 2016.
This paper presents a work in progress that introduces and studies Computational Accountability. Computational Accountability is an example of a di erent way in which AI and Ethics may intertwine, that concerns the traceability, evaluation, and communication of values and good conduct. This is currently an open challenge, that we believe can be faced with the support of intelligent systems, and that has plenty of potential applications in such diverse elds as nance and business transactions, fair business practices, (human-resource) management, consumer protection, economic systems, corruption, taxation, sales and marketing, health care, public administration, research, smart cities, and decision support.
What is accountability? Generally speaking, accountability is the acknowledgment and assumption of responsibility for decisions and actions that an individual, or an organization, has towards another party. The concept implicitly includes the expectation of account-giving: individuals are expected to account for their actions and decisions when put under examination. In human societies, such an examination is usually carried out by a \forum" of auditors. Accountability is an ethical value, and it is crucial in many (either institutionalized or relational) contexts in which humans interact. Human beings are naturally capable of understanding and tackling accountability. However, humans are being more and more often assisted in their work, and in their lives, by sophisticated sociotechnical systems or even by socio-cognitive technical systems (see, e.g., [11, 8]) { i.e. systems where the interacting individuals may be humans as well as arti cial agents. In this setting it is extremely important to realize via software the abilities to trace, evaluate, and communicate accountability, to support the interacting parties, and to help solve disputes as a forum of auditors would do. However, tracing, communicating, and evaluating accountability is a complex task, as the next example shows.
Ted and Bill are two painters, called by Jim for estimating the cost of painting a room. Ted makes a better o er and Jim decides to assign the work to him. The walls were originally white, and Jim would like to have them painted of the same color. Bill, however, is not a good loser. Since he has a spare tin of black paint in the night he paints the walls black. When Ted nds out what happened he realizes he will not be able to satisfy the commitment he made with Jim because, in order to do a nice job, he will have to use twice as much paint as expected at a much higher cost.
This simple example shows many challenges brought about when trying to tackle accountability in a computational way. Let us suppose the agreement between Ted and Bill was formalized in some way. Ted is unable to ful ll the contract that binds him to Jim. Should the simple fact that he took a commitment, which is now impossible for him to ful ll, make a computational system conclude he is accountable for the failure? Clearly, conditions changed since when he, Bill, and Ted inspected the room. One may argue that contextual conditions that constitute the prerequisites, for Ted, for the execution of the work should have been formalized. In the real world, however, contextual conditions that hardly change over time are presumed implicitly stipulated even when they are not formalized. How could Ted foresee a possible change in the color of the room to paint? Is, then, Jim accountable? Reasoning about causes, the door may have been left unlocked. Then, it could be argued that it is Jim, and not Ted, the one to blame for not having taken care of the room. On the other hand, we know that Bill is the one who changed the colors of the room but how can Bill be considered in the process of deciding who is accountable? He is not \in the system": he was not assigned the job and has no relationship with Jim or Ted. An alternative ending of the story is that Ted, feeling responsible because of his commitment, will paint the room at the agreed price because he values the satisfaction of the contract more than earning money. 3
De nitions of accountability vary in approach and scope and di erent
communities do not share a same understanding. This is mostly due to the fact that the
notion strictly depends on socio-cultural aspects that characterize di erent
communities and di erent application domains. In this section, we particularly refer
to [
{ Accountability is a composition of processes. Accountability can be divided
up into three processes. First, one must identify accountable parties, for
what they are accountable, and to whom they are accountable. Second, if a
condition veri es for which a party was previously identi ed as accountable,
a forum of some kind convenes to gather the necessary information and
passes a judgment as to the accountability of said party. Third, one must
assign positive or negative sanctions to the accountable party.
{ Accountability does not hinder autonomy. An accountable agent has complete
freedom to do as they choose, but only will later potentially be taken to
account for their actions [
Modern enterprises [
Two are the main approaches to the development of business models (and enterprise software): the business process approach and the artifact-centric approach. A business process describes how a set of interrelated activities can lead to a precise and measurable result (e.g., a service) in response to an external event (e.g., a new order). Business processes are used for developing software systems that concretely support the work of a rm. In this light, business processes become work ows that connect and coordinate di erent people, o ces, organizations, and software in a compound ow of execution. This process-centric view enables things like the analysis of an enterprise functioning, and the identi cation of criticalities, like bottlenecks, but since interactions among the actors are only indirectly represented by means of input/output dependencies between activities, it hinders reasoning about accountability (because accountability mainly concerns the interactions that emerge and evolve among the parties).
The artifact-centric approach, e.g. [
We believe that in order to realize systems of accountability it is necessary to
combine the two cited levels [
Multiagent systems o er abstractions that provide a promising basis of development. Two fundamental characteristics of agents are autonomy and situatedness. Agents are autonomous in the sense that they have a sense-plan-act deliberative cycle, which gives them control of their internal state and behavior. Agents are situated because they can sense, perceive, and manipulate the environment in which operate.
In particular, it is possible to reify the social environment of the agents in a way that supports accountability. A social commitment C(x; y; s; u) models the directed relation between two agents, a debtor x and a creditor y [14]. The debtor commits to its creditor to bring about the consequent condition u when the antecedent condition s holds. Commitments evolve along a standard lifecycle as a consequence of the actions agents perform. For instance, a conditional commitment, whose antecedent condition results being true after some action is executed, becomes detached.
Commitments have a normative value because the debtor of a detached commitment is expected to bring about, sooner or later, the consequent condition of that commitment otherwise it will be liable for a violation. The fact that debtors should satisfy their commitments creates social expectations on the agents' behaviors. Nevertheless, agents remain norm-autonomous [9] in two ways: an agent becomes debtor of a commitment by its own decision, an agent decides whether satisfying the obligation entailed by a commitment of which it is debtor.
An agent creates commitments towards other agents while it is trying to
achieve its goals (or precisely with the aim of achieving its goals) [16]. The
creation of a commitment starts an interaction of the debtor with its creditor
that coordinates, to some extent, the activities of the two, thus supporting the
achievement of goals that an agent alone could not achieve. An agent creates a
conditional social commitment towards some other agent, based on its own
beliefs and goals [16]. The creditor agent will detach the conditional commitment if
and when it deems it useful for its own purposes, thus activating the obligation
of the debtor agent. So, conditional commitments play a fundamental role in
the realization of interactivity, intended by the fact that a message relates to
previous messages and to the way previous messages related to those preceding
them [12]. In other words \there is a causal path from the establishment of a
commitment to prior communications by the debtor of that commitment" [15,
Sect. 4.4]. This aspect is fundamental for reasoning about accountability and
responsibility. In general, the di erence between responsibility and accountability
can be expressed temporally: responsibility's domain lies intuitively in the future
while accountability looks to the past. Responsibility requires a forward-looking
approach in the form of a capability study. An entity cannot be responsible if
that entity lacks the capacity to in uence results. Accountability, on the other
hand, involves a posterior analysis to understand how a given result came to be.
With accountability, an entity might be deemed accountable even if that entity
was not preemptively identi ed as potentially accountable in a given context.
The actual analysis of accountability can be accomplished by looking at
commitment relationships, which collectively form chains of action from one context
to another thanks to exchanged creditor/debitor and antecedent/consequent.
On this foundation, commitments can be used to realize a relational
representation of interaction, where agents, by their own action, directly create
normative binds (represented by social commitments) with one another, and use such
binds to coordinate their activities, e.g. through responsibility assignment, as
well as to identify liabilities. Future work will focus on the temporal distinction
and connections between the two concepts in order to create a comprehensive
system of accountability exible enough to move from system to system yet
speci c enough to understand contextual particularities and unique relationships.
A starting point will be the 2COMM framework, that is described in [
Acceptance. In ATAL, pages 99{112, 1998. 10. Simon Kramer and Andrey Rybalchenko. A Multi-Modal Framework for Achieving Accountability in Multi-Agent Systems. In Proc. of Workshop on Logics in Security, pages 148{174, 2010. 11. Pablo Noriega, Julian Padget, and Mark d'Inverno. The challenge of arti cial sociocognitive systems. In Pre-proceedings of Coordination, Organizations, Institutions, and Norms in Agent Systems COIN@AAMAS, 2014. 12. Sheizaf Rafaeli. Sage Annual Review of Communication Research: Advancing Communication Science: Merging Mass and Interpersonal Processes, chapter (Chapter 4) Interactivity: From new media to communication, pages 110{134. Sage, 1988. 13. Judith Simon. The Online Manifesto: Being human in a hyperconnected era, chapter Distributed Epistemic Responsibility in a Hyperconnected Era. Springer Open, 2015. 14. Munindar P. Singh. An ontology for commitments in multiagent systems. Artif.
Intell. Law, 7(1):97{113, 1999. 15. Munindar P. Singh. Commitments in multiagent systems some controversies, some prospects. In The Goals of Cognition. Essays in Honor of Cristiano Castelfranchi, chapter 31, pages 601{626. College Publications, London, 2011. 16. Pankaj R. Telang, Munindar P. Singh, and Neil Yorke-Smith. Relating goal and commitment semantics. In ProMAS, volume 7217 of Lecture Notes in Computer Science, pages 22{37. Springer, 2011.