Part 2 of the Alphabet Soup of Responsible AI Adjectives: Definitions

Explainable, interpretable, transparent, responsible, trustworthy, principled, ethical, antiracist, good. These are many different though often overlapping adjectives that have been used to describe techniques to improve AI in an effort to ensure that it is beneficial to society. But what do these terms actually mean? Let’s dive into the definitions and attempt to disentangle these adjectives as they relate to AI.

See Part 1 for a landscape of the uses of these AI adjectives.

Explainable AI: Often abbreviated XAI, an explainable AI can be understood as, “one that produces details or reasons to make its functioning clear or easy to understand” (Barredo Arrieta et al., 2020). Explanation types can include explanation by example, explanation by simplification, and explanation with counterfactuals.

Interpretable AI: Interpretability is often understood as the degree to which a model can be explained in terms understandable to a human (Barredo Arrieta et al., 2020). Some researchers use “interpretable” and “explainable” interchangeably (Miller, 2017) while others describe a “chasm” between interpretability and explainability (Rudin, 2019). Interpretability is considered difficult to define mathematically (Molnar, 2022) and domain-specific (Rudin, 2019).

Transparent AI: Barredo Arrieta (2020) says that, “a model is considered to be transparent if by itself it is understandable.” This definition is extended to “algorithmic transparency” such that a model is fully explorable by mathematical methods.

Responsible AI: Responsible AI considers the larger framework within which AI is developed and used. This includes considerations of fairness, accountability, privacy, explainability, and societal impact (Dignum, 2019). Responsible AI is often linked with ethical AI, but responsible AI also considers legal, economical and cultural concerns along with ethical priorities (Zhu, 2021).

Trustworthy AI: The NIST AI Risk Management Framework (2022) defines trustworthy AI as, “valid and reliable, safe, fair and bias is managed, secure and resilient, accountable and transparent, explainable and interpretable, and privacy-enhanced.” Transparency, explainability, and interpretability overlap with other definitions in this list.

Principled AI: Middelstadt (2019) advocated for a principled approach to implement AI. This type of approach relies on defining AI principles and establishing clear fiduciary duties for data subjects and users. Methods can then be developed to translate principles into practice.

Ethical AI: Ethics is the study of morals and values, and ethical AI considers how these morals and values apply to the development, deployment, and decision making of AI technologies. Ethical AI is considered more limited in scope than responsible AI (Dignum, 2019).

Antiracist AI: Antiracism is the practice of actively identifying and opposing racism through changing policies and behaviors. Antiracist AI involves the centering of racial equity in work related to AI development and deployment (Waikar, 2021).

AI for Good: AI for Good should generate a benefit for humanity and the common good (Barredo Arrieta et al., 2020). AI for Social Good (AI4SG) is a specific variety of AI for Good that advocates for alignment with the United Nations Sustainable Development Goals (SDGs). This includes formalizing AI development and deployment to, as described in Cowls et. al (2021), “(i) prevent, mitigate and/or resolve problems adversely affecting human life and/or the wellbeing of the natural world, and/or (ii) enable socially preferable or environmentally sustainable developments, while (iii) not introducing new forms of harm and/or amplifying existing disparities and inequities.”

Commentary

Although many of the definitions of these terms overlap, no one term describes the entire space. Stakeholders including researchers should evaluate which term(s) most closely apply to their work, and consider which areas may benefit from additional exploration and investment.

This blog post is part of a Deep Dive into Responsible Data Science and AI series.

References

Barredo Arrieta A, Díaz-Rodríguez N, Del Ser J, et al. Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI. Information Fusion. 2020;58:82–115. doi:10.1016/j.inffus.2019.12.012

Cowls J, Tsamados A, Taddeo M, Floridi L. A definition, benchmark and database of AI for social good initiatives. Nat Mach Intell. 2021;3(2):111–115. doi:10.1038/s42256–021–00296–0

Dignum V. Responsible Artificial Intelligence: How to Develop and Use AI in a Responsible Way. Springer; 2019.

Miller T, Howe P, Sonenberg L. Explainable AI: Beware of Inmates Running the Asylum. IJCAI 2017 Work- shop on Explainable Artificial Intelligence (XAI). Published online 2017:7.

Mittelstadt B. AI Ethics — Too Principled to Fail? SSRN Journal. Published online 2019. doi:10.2139/ssrn.3391293

Molnar C, König G, Herbinger J, et al. General Pitfalls of Model-Agnostic Interpretation Methods for Machine Learning Models. In: Holzinger A, Goebel R, Fong R, Moon T, Müller KR, Samek W, eds. XxAI — Beyond Explainable AI: International Workshop, Held in Conjunction with ICML 2020, July 18, 2020, Vienna, Austria, Revised and Extended Papers. Lecture Notes in Computer Science. Springer International Publishing; 2022:39–68. doi:10.1007/978–3–031–04083–2_4

NIST AI RMF Playbook. NIST. Published online July 8, 2022. https://www.nist.gov/itl/ai-risk-management-framework/nist-ai-rmf-playbook

Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 2019;1(5):206–215. doi:10.1038/s42256–019–0048-x

Waikar, S. Designing Anti-Racist Technologies for a Just Future. Stanford HAI. 2021. https://hai.stanford.edu/news/designing-anti-racist-technologies-just-future

Zhu L, Xu X, Lu Q, Governatori G, Whittle J. AI and Ethics — Operationalising Responsible AI. arXiv:210508867 [cs]. Published online May 18, 2021. http://arxiv.org/abs/2105.08867

Disclaimer: Support for this blog series was provided by RTI International. The opinions expressed by the author are their own and do not represent the position or belief of RTI International. Material in this blog post series may be used for educational purposes. All other uses including reprinting, modifying, and publishing must obtain written consent.