This paper offers a solution to the Problem of Ethical Integralism (PEI). The PEI affects artificial agents when they are unable to perform an operation in different ways respecting, in each case, a given set of ethical principles. For example, a bank customer wants to be able to transfer money from her account either on-line or in person, and wishes her privacy to be respected in both situations. In this case, an automated accounting system should be able to assist the customer by performing the same operation in two different ways while respecting the ethical principle of privacy.

The PEI is particularly relevant when part of the activity of an organization is automated, and the overall ethical conduct of the organization has to respect a set of ethical principles. Typical examples of these organizations include banks with automated client management [1], hospitals with automated delivery of medicines [2] or digitalized diagnostic processes [5], trading companies [3], and information providers [4].

The paper is divided into three sections. The first section is dedicated to the analysis and definition of the PEI, and opens with a brief review of the solutions proposed to solve the Problem of Ethical Integralism when designing artificial agents [6-9]. The attempts to embed deontological reasoning capacities into artificial agents are compared with the different approach advocated in this paper. Here, it is argued that inscribing ethical principles into artificial agents ‘by design’ leads to localized behaviors that can prove to be ethically consistent with the common practice of private or public organizations. This approach renounces the flexibility that could be achieved by an artificial deontological system, but gains the possibility of proving, from the early stages of the design of the software agents, that their actions will be bound by the given set of ethical principles.

The definition of the PEI relies on two main concepts: flexibility and ethical exception. An artificial agent is flexible when it is able to perform the same operation in different ways; it makes no ethical exception when it maintains its flexible behavior within the boundaries of the given ethical principles.

The definition of the PEI in terms of flexibility and ethical exception closes the first section of the paper. Given:

• 1. an ethical principle expressing a deontic constraint; and
• 2. an operation constrained by that principle;

the Problem of Ethical Integralism is the inability to perform that operation in different ways, while ensuring that the relevant deontic constraint is respected.

The second section of the paper contains the proposed solution for the PEI. This solution is based on a threefold process of translation that transforms ethical principles into ethical requirements, and the latter into ethical protocols [10]. Ethical requirements define the conditions according to which the behaviors of a computational system may be consistent with the given ethical principles. Ethical protocols specify the single operations performed by the computational system that match the constraints posed by the ethical requirements.

The first stage of the translation requires an analytical method that can be used to decompose ethical principles expressed in natural language. In the paper, a method based on the theory of level of abstractions [11] is proposed. Copyright is used as a case study to show how an ethical principle expresses a normative constraint over a class of operations performed by a set of observables (see figure 1).

The second stage of the translation process is based on a conceptual tool called the Control Closure of an operation, CC(op). A CC(op) is the set of agents whose state variables are needed to perform the operation op [12]. The CC(op) is used to translate the deontic constraint into a precondition for the execution of the relevant class of operations. An operation of that class can be executed only if that precondition – and hence the normative constraint – is satisfied.

Figure 1 Analytical decomposition of the ethical principle of copyright

The last stage of the translation produces a set of specific operations with the relative control closure. This protocol is an analytic description of the behavior of the artificial agent that is being designed.

The solution of the PEI depends on taking into account the distributed nature of the actions of artificial agents. Agents are embodied into an environment and their operations depend not only on their own states but also on those of the environment and, in case, of other agents. Therefore, the concept of CC(op) is further developed by distinguishing its cardinality – the number of agents required to perform the operation op – from its quality – the type of state variables that are required to perform that operation.

An artificial agent that does not run into the PEI performs an operation op in at least two different ways respecting, in every case, one or more ethical principles. In the paper, it is argued that two instances of the same operation are different if they are performed by a different number of agents, and that they are guaranteed to respect the same ethical principles if their CCs have the same qualitative properties.

The third and final part of the paper introduces an explanatory example of the proposed solution for the PEI. The context is the automation of driving systems for cars and the ethical principles of safety and privacy. The example begins analyzing a monitor system called Teen Driver Support System (TDSS). This system monitors and records some of a driver’s activities. The goal of the system is to allow the parents of teen drivers to monitor the driving attitudes of their children.

In the paper it has been posited that while the TDSS can be used without breaching the driver’s right to privacy, it cannot be used to guarantee the principle of safety. Different systems are then considered, produced by Volkswagen and Ford, which allow the full automation of some or all the driving activity. The ethical analysis of these systems shows how they can be affected by the PEI. The explanatory example is completed by a description of how the proposed solution to the PEI can be used to inscribe the ethical principle of privacy and safety into those systems without incurring the PEI.

In the conclusion, some ethical implications of the proposed solution are pointed out. Might the responsibility of the driving actions be ascribed to the automated system or is it still held by the driver? Is the minimal moral commitment a guideline to endorse the ethical design of artificial agents? These and others relevant questions are important consequences of the research presented into this paper.

References

1. Ranger, S., Bank automates to boost customer service. Case study: HSBC speeds up queries with workflow automation, in Silicon.com. 2006.

2. Fitzpatrick, R., et al., Evaluation of an Automated Dispensing System in a Hospital Pharmacy Dispensary. The Pharmaceutical Journal, 2005. 274: p. 763--765.

3. Arbel, A. and A.E. Kaff, Crash: Ten Days in October. Will It Strike Again? 1989, Chicago: Longman Financial Services.

4. AP, Justice Dept. Rejects Google's Privacy Issues. The New York Times 2006(February 26): p. 23.

5. Jirotka, M., et al. Towards Understanding Requirements for eScience: the eDiaMoND case study. in UK e-Science All Hands Meeting 2004. 2004. Oxford.

6. Adam, A., Delegating and Distributing Morality: Can We Inscribe Privacy Protection in a Machine? Ethics and Information Technology, 2005. 7(4): p. 233--242.

7. Allen, C., I. Smith, and W. Wallach, Artificial morality: top-down, bottom-up, and hybrid approaches. Ethics and Information Technology, 2005. 7: p. 149-155.

8. Andries, B. and P. Laurette, Ethical design issues in cyberspace, in Proceedings of the 4th international symposium on Information and communication technologies. 2005, Trinity College Dublin: Cape Town, South Africa.

9. Wiegel, V., J.v.d. Hoven, and G.-J. Lokhorst, Privacy, Deontic Epistemic Action Logic and Software Agents. Ethics and Information Technology, 2005. 7(4): p. 251--264.

10. Turilli, M., Ethical Protocols Design. Ethics and Information Technology, 2007. 8.

11. Floridi, L. and J.W. Sanders, The Method of Abstraction, in Yearbook of the Artificial. Nature, Culture and Technology, M. Negrotti, Editor. 2004, P. Lang: Bern. p. 177--220.

12. Sanders, J. and M. Turilli, Dynamics of Control. forthcoming.