Skip to main content
Springer Nature Link
Log in

Should explainability be a fifth ethical principle in AI ethics?

  • Original Research
  • Published:
AI and Ethics Aims and scope Submit manuscript

Abstract

It has been recently claimed that explainability should be added as a fifth principle to AI ethics, supplementing the four principles that are usually accepted in Bioethics: Autonomy, Beneficence, Nonmaleficence and Justice. We propose here that with regard to AI, on the one hand explainability is indeed a new dimension of ethical concern that should be paid attention to, while on the other hand, explainability in itself should not necessarily be considered an ethical “principle”. We think of explainability rather (i) as an epistemic requirement for taking into account ethical principles, but not as an ethical principle in itself; (ii) as an ethical demand that can be derived from ethical principles. We do agree that explainability is a key demand in AI Ethics, with practical importance for stakeholders to take into account; but we argue that it should not be considered as a fifth ethical principle, to maintain a philosophical consistency in the organization of AI ethical principles.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Log in via an institution

We’re sorry, something doesn't seem to be working properly.

Please try refreshing the page. If that doesn't work, please contact support so we can address the problem.

Similar content being viewed by others

Explore related subjects

Discover the latest articles, news and stories from top researchers in related subjects.

Notes

  1. We will adopt the term “explainability”; the terminology of the concept is contested, as we will see below. We will use other terms only when commenting on other authors’ ideas, in which case we preserve the terms used by them.

  2. Using the term “intelligibility”, the House of the Lords report from 2018, for instance, already suggests “five hierarching principles”, that one could associate in some way to the four principles of bioethics and to explainability [18].

  3. Floridi et al. [10], use the term “explicability”, whereas the more common term in the domain is “explainability”. On “explicability” as a fifth principle, see also [11].

  4. We thank a reviewer who suggested this possible research path to us.

  5. This is similar to accountability of a company in some aspects: it should not declare and explain everything it does, but in a particular case (for instance being judged for a crime), it should be capable of showing relevant information and justifying its decisions.

  6. Besides, one should keep in mind that several people are involved in the process of the development of an algorithm. As Coeckelbergh [6] says, AI is a problem of “many hands”, in the sense that “many people are involved in technological action”, which makes responsibility attribution difficult in this case—we could think that to know to whom address explainability and how to do is also made harder by this “many hands” aspect of AI.

  7. An explicit quotation is made at page 11 of the document [15].

  8. The document does not absolutely ignore beneficence, since it proposes that the implementation of a Trustworthy AI “entails seeking to maximise the benefits of AI systems while at the same time preventing and minimising their risks” [15]. One should recall that at the Belmont Report [1] one could find the three principles of “Respect for Persons”, “Beneficence” and “Justice”. Beauchamp and Childress [2] defend that we should consider two different principles, Beneficence and Nonmaleficence, having a total of four principles.

  9. One could think that if a system is transparent, it does not, for that very same reason, need to be explained—if this is the case, simply saying that both transparency and explainability are desirable is not enough.

  10. One should not think that for a principle to be adopted the nature of its concept should be clear. This is not our argument for explainability not being an ethical principle: as we will see, the problem lies rather in the ethical demand for explainability, that we see as originating (when it is the case) from the other principles.

  11. We will address below the fact that, in some cases, explainability can be valued per se—but we will show that even though it is not for that reason a principle.

  12. Coeckelbergh does not say that we can attribute responsibility to an AI as a moral agent, but rather that “only humans can be responsible agents” [5]. However, he claims that the aristotelian “philosophical analyses of ignorance”—in particular, to know the technology you are using—“can guide discussions about knowledge problems with AI” [5]. In this sense, one can say that having an explainable algorithm can contribute to giving the agent an epistemic condition.

  13. Herzog [14] claims that, for Floridi et al. [10], “the principle of explicability is introduced as enabling the other principles of ‘beneficence’, ‘nonmaleficence’, ‘autonomy’ and ‘justice’, rather than being a primary principle”. We cannot agree with this reading of the passage: if it is not a principle, how could the authors write in the continuation of the article about the “addition” of a “principle” concerning explicability? Herzog does not explain how one should differentiate between a principle and a “primary principle”.

  14. The Belmont Report deals with “respect for persons”, which was later associated with an “respect for autonomy” principle [1].

  15. Regarding the fact that explainability should be presented to stakeholders with different backgrounds that interact with the system at different moments, one could think for instance of an approach to explainability that tries to present a “minimal” explainability, comprehending demands that appear for all users; another solution would be to differentiate explanations according to the stakeholders (Herzog [14], for instance, claims that “we should not only be interested in the developing party as the responsible entity, but also in the commissioning, deploying and, ultimately, the utilizing parties.”, thinking in a conceptualization of “explicability” that encompasses several stakeholders). However, this question will not be further developed here.

  16. We should note that, despite adopting a different theoretical approach, Mirbabaie and colleagues [23] propose a framework in which “explainability/explicability”, by means of the concept of “transparency”, appears under the Autonomy principle (the four principles of Bioethics being assumed).

  17. "Obligations of nonmaleficence include not only obligations not to inflict harms, but also obligations not to impose risks of harm." [2]

  18. Here, we assume accuracy as a good proxy of performance.

  19. One could say that in general it is better to understand a technology than not—of course, this is not absolute, as better understanding it typically causes lack of some other aspect. These are the trade-off situations we will analyse.

  20. Ethical principles are not absolute, in the sense that in the case of a conflict between principles one can be postponed for the sake of another. There is always a demand for a principle—what is sometimes designed as prima facie principles: the principle should be considered in itself, except in the case of a more urgent demand.

  21. The European Commission’s High-Level Expert Group on Artificial Intelligence [15] declares that the degree of explicability that is needed is dependent on the severity of consequences of it, to which they comment: “for example, little ethical concern may flow from inaccurate shopping recommendations generated by an AI system, in contrast to AI systems that evaluate whether an individual convicted of a criminal offence should be released on parole”. One could think, however, that, even in this case, to evaluate whether a risk of infringement of a principle occurs is not so simple.

  22. It is not clear whether every end in ethics is necessarily intrinsically valued—cf. Korsgaard [19], which claims for two different distinctions, between means and ends and between intrinsic values and extrinsic values. These problems should be addressed in future work regarding AI ethics.

  23. As we said, there is an open question regarding if one should talk about something that is “valued in itself”, “intrinsically good” or an “end” for the present question, and we leave it for future works.

References

  1. The Belmont Report: Ethical Guidelines for the Protection of Human Subjects. Washington: DHEW Publications (OS) 78-0012 (1978)

  2. Beauchamp, T.L., Childress, J.F.: Principles of Biomedical Ethics, 8th edn. Oxford University Press (2019)

    Google Scholar 

  3. Beauchamp, T. L., & DeGrazia, D.: Principles and principlism. In: G. Khushf (ed), Handbook of Bioethics, Springer, Dordrecht, pp. 55–74 (2004)

  4. Buchanan, B.G., Smith, R.G.: Fundamentals of expert systems. Ann. Rev. Comput. Sci. 3, 23–58 (1988)

    Article  Google Scholar 

  5. Coeckelbergh, M.: Artificial intelligence, responsibility attribution, and a relational justification of explainability. Sci. Eng. Ethics 26(4), 2051–2068 (2020)

    Article  Google Scholar 

  6. Coeckelbergh, M.: AI Ethics. The MIT Press, Cambridge (2020)

    Book  Google Scholar 

  7. DARPA: Explainable Artificial Intelligence (XAI). Broad Agency Announcement DARPA-BAA-16-53 (2016)

  8. Fjeld, J., et al.: Principled Artificial Intelligence: Mapping Consensus in Ethical and Rights-based Approaches to Principles for AI. Berkman Klein Center for Internet & Society (2020)

    Google Scholar 

  9. Floridi, L.: Infraethics – on the conditions of possibility of morality. Philos. Technol. 30(4), 391–394 (2017). https://doi.org/10.1007/s13347-017-0291-1

    Article  Google Scholar 

  10. Floridi, L., et al.: AI4People—an ethical framework for a good AI society: opportunities, risks, principles, and recommendations. Mind. Mach. 28(4), 689–707 (2018)

    Article  Google Scholar 

  11. Floridi, L., Cowls, J.: A unified framework of five principles for AI in society. Harv. Data Sci. Rev. 1(1), 1–15 (2019)

    Google Scholar 

  12. Gorry, G.A.: Computer-assisted clinical decision making. Methods Inf. Med. 12, 45–51 (1973)

    Article  Google Scholar 

  13. Hermann, E., Hermann, G.: Artificial intelligence in research and development for sustainability: the centrality of explicability and research data management. AI Ethics (2021). https://doi.org/10.1007/s43681-021-00114-8

    Article  Google Scholar 

  14. Herzog, C. On the risk of confusing interpretability with explicability. AI and Ethics: 1–7 (2021).

  15. High-Level Expert Group on Artificial Intelligence: “Ethics Guidelines for Trustworthy AI”. Document made public on 8 April 2019 (2019). https://digital-strategy.ec.europa.eu/en/policies/expert-group-ai

  16. Jha, S.: Value of triage by artificial intelligence. Acad. Radiol. 27(1), 153–155 (2020)

    Article  MathSciNet  Google Scholar 

  17. Kemp, P., Dahl Rendtorff, J.: The Barcelona declaration. Towards an integrated approach to basic ethical principles. Synth. Philos. 23(2), 239–251 (2008)

    Google Scholar 

  18. House of the Lords: AI in the UK: ready, willing and able? HL Paper 100 (2018).

  19. Korsgaard, C.M.: Two distinctions in goodness. Philos. Rev. 92(2), 169–195 (1983)

    Article  Google Scholar 

  20. Lacave, C., Diez, F.J.: A review of explanation methods for Bayesian networks. Knowl. Eng. Rev. 17(2), 107–127 (2002)

    Article  Google Scholar 

  21. Lipton, Z.C.: The mythos of model interpretability. ACM Queue 16(3), 1–27 (2018)

    Article  Google Scholar 

  22. Miller, T.: Explanation in artificial intelligence. Artif. Intell. 267, 1–38 (2019)

    Article  MATH  Google Scholar 

  23. Mirbabaie, M., et al.: Artificial intelligence in hospitals: providing a status quo of ethical considerations in academia to guide future research. AI Soc. (2021)

  24. Mittelstadt, B.: Principles alone cannot guarantee ethical AI. Nat. Mach. Intell. 1(11), 501–507 (2019)

    Article  Google Scholar 

  25. Nguyen, C. T. Transparency is Surveillance. Philos Phenomenol Res. (2021). https://doi.org/10.1111/phpr.12823

  26. Rana, A., Bridge, D.: Explanations that are Intrinsic to Recommendations. In: Proceedings of the 26th Conference on User Modeling, Adaptation and Personalization, 187–195 (2018)

  27. Rawls, J.: A Theory of Justice, 2 revised Belknap Press, Cambridge (1999)

    Book  Google Scholar 

  28. Ribeiro, M.T., Singh, S., Guestrin, C.: Why should I trust you? Explaining the predictions of any classifier. In: Proceedings of the ACM International Conference on Knowledge Discovery and Data Mining, 1135–1144 (2016)

  29. Richardson, H.S.: Specifying, balancing, and interpreting bioethical principles. J. Med. Philos. Forum Bioethics Philos. Med. 25(3), 285–307 (2000)

    Article  Google Scholar 

  30. Robbins, S.: A misdirected principle with a catch: explicability for AI. Mind. Mach. 29(4), 495–514 (2019)

    Article  MathSciNet  Google Scholar 

  31. Roemer, J.E.: Theories of Distributive Justice. Harvard University Press, Cambridge (1996)

    Google Scholar 

  32. Rudy-Hiller, F.: The epistemic condition for moral responsibility, The Stanford Encyclopedia of Philosophy (Fall 2018 Edition). In: Zalta, E.N. (ed.). https://plato.stanford.edu/archives/fall2018/entries/moral-responsibility-epistemic/. (2018). Accessed 22 Mar 2022

  33. Sarkar, S., et al.: Accuracy and interpretability trade-offs in machine learning applied to safer gambling. In: CEUR Workshop Proceedings, vol. 1773, 1–9 (2016)

  34. Selvaraju, R.R., et al.: Grad-CAM: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE International Conference on Computer Vision, 618–626 (2017)

  35. Sen, A.: Commodities and Capabilities. North-Holland, Amsterdam (1985)

    Google Scholar 

  36. Tsamados, A., et al.: The ethics of algorithms: key problems and solutions. AI Soc. 37, 215–230 (2022)

    Article  Google Scholar 

  37. Turilli, M., Floridi, L.: The ethics of information transparency. Ethics Inf. Technol. 11(2), 105–112 (2009). https://doi.org/10.1007/s10676-009-9187-9

    Article  Google Scholar 

  38. Ursin, F., Timmermann, C., Steger, F.: Explicability of artificial intelligence in radiology: Is a fifth bioethical principle conceptually necessary? Bioethics 36(2), 143–153 (2022)

    Article  Google Scholar 

  39. Whittlestone, J., Nyrup, R., Alexandrova, A., Cave, S.: The role and limits of principles in AI ethics: towards a focus on tensions. In: Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, 195–200 (2019)

Download references

Acknowledgements

We thank André Levi Zanon, Cláudio Amorim, Lucas Petroni, Marcelo Santis, Marcos Menon José, Paulo Pirozelli and Tiago Lubiana for reading earlier versions of this article. We also thank an anonymous reviewer for some valuable suggestions. We thank Fapesp, CAPES and CNPq agencies, as well as the Fundação José Luiz Egydio Setúbal, for their support through research grants (the second author is partially supported by CNPq grant 312180/2018-7 and FAPESP grant 2019/07665-4).

Author information

Authors and Affiliations

  1. Núcleo de Bioética, Fundação José Luiz Egydio Setúbal, Av. Angélica, 2.071, Consolação, São Paulo, SP, 01228-200, Brazil

    João Figueiredo Nobre Brito Cortese, Marcos Paulo Lucca-Silveira & Adriano Figueiredo Bechara

  2. Center for Artificial Intelligence, C4AI at Universidade de São Paulo, Av. Prof. Lúcio Martins Rodrigues, 370, Butantã, São Paulo, SP, 05508-020, Brazil

    Fabio Gagliardi Cozman

  3. São Paulo School of Economics, FGV, Rua Itapeva, 474, São Paulo, 01332-000, Brazil

    Marcos Paulo Lucca-Silveira

  4. Departamento de Fisiologia, Instituto de Biociências, Universidade de São Paulo, R. do Matão, 321, Butantã, São Paulo, SP, 05508-090, Brazil

    João Figueiredo Nobre Brito Cortese

  5. Ibmec-SP, Alameda Santos, 2356, Jardim Paulista, São Paulo, SP, 01418-901, Brazil

    João Figueiredo Nobre Brito Cortese

  6. Department of Mechatronics and Mechanical Systems, Escola Politécnica, Universidade de São Paulo, Av. Professor Mello Moraes, 2231, Butantã, São Paulo, SP, 05508-030, Brazil

    Fabio Gagliardi Cozman

Authors
  1. João Figueiredo Nobre Brito Cortese
    View author publications

    You can also search for this author in PubMed Google Scholar

  2. Fabio Gagliardi Cozman
    View author publications

    You can also search for this author in PubMed Google Scholar

  3. Marcos Paulo Lucca-Silveira
    View author publications

    You can also search for this author in PubMed Google Scholar

  4. Adriano Figueiredo Bechara
    View author publications

    You can also search for this author in PubMed Google Scholar

Corresponding author

Correspondence to João Figueiredo Nobre Brito Cortese.

Ethics declarations

Conflict of interest

On behalf of all authors, the corresponding author states that there is no conflict of interest.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Cortese, J.F.N.B., Cozman, F.G., Lucca-Silveira, M.P. et al. Should explainability be a fifth ethical principle in AI ethics?. AI Ethics 3, 123–134 (2023). https://doi.org/10.1007/s43681-022-00152-w

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1007/s43681-022-00152-w

Keywords