🔬 Research summary by Ramya Srinivasan, AI Researcher at Fujitsu Research of America Inc.
[Original paper by Ramya Srinivasan and Devi Parikh]
Overview: The paper outlines some ways in which generative artworks could aid in narrowing the communication gaps between different stakeholders in the AI pipeline. In particular, the authors argue that generative artworks could help surface different ethical perspectives, highlight mismatches in the AI pipeline, and aid in the visualization of counterfactual scenarios, and non-western ethical perspectives.
A picture is worth a thousand words!
Indeed, visuals are extremely effective in conveying complex concepts in an accessible manner—they transcend language barriers, simulate engagement, trigger critical thinking, and leave lasting imprints in the minds of the observer. Backed by this understanding, the authors posit that generative artworks (i.e., artworks created by AI systems) could come handy in educating AI scientists with regards to potential pitfalls in the design, development, and deployment of the AI systems. To substantiate their argument, the authors lay out four potential pathways in which generative artworks could be leveraged in educating AI scientists about AI ethics, namely,—1) by visualizations of different ethical viewpoints, 2) by visualizations of mismatches in the AI pipeline, 3) by visualizations of counterfactual scenarios, and 4) by visualizations of non-western ethical perspectives.
Here, a brief description of each of the four aforementioned potential pathways (through which generative artworks could aid in enhancing AI ethics) is provided.
Visualizations of different ethical perspectives: Different ethical theories emphasize different principles in decision making, and can thus shed light on varying viewpoints relevant in a given context. For example, in utilitarian ethics, the emphasis is on maximizing the well-being of all stakeholders, which is not necessarily the case in deontological ethics, where the emphasis is on following the laws and regulations. Thus, even within the context of a single problem setting, there can be diverse viewpoints about what is right, fair, just, or appropriate. In order to enhance AI ethics, it thus becomes important to educate AI researchers and developers about these diverse viewpoints and thereby aid in reflexive design. Generative artworks could serve as powerful visualization tools to surface such diverse perspectives. For example, through generative artworks, it may be possible to visualize the compounded adverse effects of an AI decision in an individual’s life, a consequentialism ethical perspective.
Visualizations of mismatches in the AI pipeline: Computational systems involve quantitatively modeling abstract concepts or constructs which may or may not be observable. Furthermore, there may be unobservable factors that affect the constructs themselves. Consider, for example, a construct such as “skill” or “ability”, which is relevant across many applications such as hiring and admissions. These constructs can be influenced by both innate potential specific to the individual and other factors such as socio-economic status. Thus, a mismatch can be introduced even before measuring a construct. Generative artworks could aid in visualizing such mismatches. For example, it may be possible to highlight differences in measurement of similar constructs, thereby aiding AI researchers and developers in understanding system behavior. Consider an AI based hiring use case. Suppose one of the features in making the decision concerns measuring social skills of the candidate. In this regard, one might expect the constructs “self-esteem” and “confidence” to be related. Visualizations of AI system’s behavior under different scenarios could reveal whether it treats these constructs similarly – whether it exhibits “convergent validity” , which refers to the degree to which two measures of constructs that theoretically should be related, are in fact related.
Visualizations of counterfactuals: Generative artworks could also aid in visualizing counterfactual situations which in turn can be beneficial in reflexive design via empathy fostering. Counterfactual thinking can help in engendering empathy by enabling one to visualize situations through another person’s world. Thus, certain situations that may be irrelevant in one person’s context, but relevant in another person’s context, can be understood via such counterfactual visualizations. Generative artworks could be used as tools to visualize the consequences of AI decisions so AI researchers and developers (for instance), who may not necessarily be affected by the decision, can empathize with the impacted population, and thereby redesign their system for the better.
Visualizations of non-western perspectives: Generative artworks can serve as visualizations of social, cultural, and economic differences that exist across geographies. For example, through generative artworks it may be possible to highlight different viewpoints regarding fairness based on the local context such as social practices, religious beliefs, economic status, etc. By training generative models on data across cultures and looking at the latent visualizations, it might also be possible to view how different everyday practices (e.g. dress, food, etc.) and objects (e.g. furniture, houses, etc.) can vary across cultures thereby shedding light on local contexts which can be valuable in AI system design.
Between the lines
The ideas postulated in the paper offer promise in that they can open up new ways of reflexive design and facilitate introspection. Generative artworks could be especially beneficial in highlighting counterfactual scenarios— given that such visualizations may not exist in the real world, and thereby could shed light on new and latent perspectives. That said, for surfacing non-western perspectives and viewpoints based on various ethical theories, existing artworks could also be used. Also, as the authors acknowledge, generative artworks could themselves be biased, so it is necessary to employ these tools mindfully. Ecological costs/environmental impacts of generative artworks are however not discussed in the paper. Given that generating artworks requires significant computational resources, there exists a tradeoff between the ecological cost and educational benefit, which calls for further analysis.