animals by predicting their behavior and location. Its video exhibits how a well-intentioned motivation can become susceptible to a flawed outcome when programmed with narrow intentions (Figure 5).

 

THE VALUE OF SPECULATION

By revealing the limitations and biases of AI, Committee of Infrastructure demonstrates how AI systems are subject to the same fallibility that is present in human-to-human interaction. Not only are the code, personality, and data necessary to create a functioning AI, but the ethical framework that guides how these systems interact is fundamental too.

AI cannot be blindly trusted; it should
be subject to the same form of scrutiny
as a bill, law, or ballot measure. The
project illustrates why developers and
researchers should create a record of
motivations, origin, and data sources
visible to everyone [ 8]. In detailing
all the variables, a discussion among
a broader set of stakeholders should
challenge preexisting assumptions
and provide evidence to thoroughly
negotiate how these systems influence
daily life.

Ultimately, Committee of Infrastructure proposes that speculation is essential for working through the uncertainty, complexity, and ambiguity of AI problems.

Designers have the opportunity to make sense of what is not yet truly understood in both capability and application. Further, AI issues are incredibly difficult and esoteric.

Committee of Infrastructure utilizes
speculation to show how humor and
play can be used to examine serious
issues like surveillance and governance,
allowing designers, developers, citizens,
and policymakers to have a dialogue
about how we want AI to influence our
daily life. By using humanistic values,
designers can promote new forms of
interaction that facilitate inclusive and
compassionate experiences. This field
affords new opportunities that call for
radical ways of working.

Endnotes

1. Deep learning is a method that utilizes vast amounts of data to learn. It uses the model of neural networks to detect and predict.

2. Google Duplex, a virtual assistant released in 2018, can schedule an appointment so skillfully that it is impossible to discern whether a human or computer is speaking. The assistant pauses, intonates, and affirms just like a human.

3. Crowdflower and Amazon’s Mechanical Turk are platforms that solicit humans to annotate data to be used by ML. ML needs copious amounts of data to be implemented.

4. Biases such as stereotyping, attentional bias, and confirmation bias can become amplified when incorporated into machine learning. For example, the COMPAS algorithm used by the Department of Corrections in Wisconsin, New York, and Florida has led to harsher sentencing toward African Americans. See Angwin, J., Larson, J., Mattu, S., and Kirchner, L. Machine bias. ProPublica. May 23, 2016; https:// www.propublica.org/ article/machine- bias-risk-assessments-in-

 

5. Dunne, A. and Raby, F. Speculative Everything: Design, Fiction, and Social Dreaming. MIT Press, Cambridge, MA, 2013.

6. Karpathy, A. The unreasonable effectiveness of recurrent neural net works.” Andrej Karpathy Blog. May 1,

2015; karpathy.github.io/2015/05/21/
rnn-effectiveness/.
7. For example, the L. A. DOT representative
learned to speak from City of Los Angeles
Transportation Impact Study Guidelines
and Traffic Studies Policy and Procedures.

8. The idea of data provenance was introduced at the 2017 AAAI symposium.

 

Jason Shun Wong is a designer, researcher, and strategist of interactions and critical media working in emerging technology. His work focuses on the intersection of smart cities, networking, behavioral psychology, object-oriented ontology, civics, Chinese science fiction, and meme culture. → [e-mail]

https://jasonshun wong.com/ work/ committee-of-infrastructure-part-2/

DOI: 10.1145/3274568 COPYRIGHT HELD BY AUTHOR. PUBLICATION RIGH TS LICENSED TO ACM. $15.00

Figure 4. Transcript excerpts created using char-rnn. Human and AI representatives arguing for their respective organizations in the vernacular learned by training the ML algorithm on seminal texts important to the ethos of each organization ( https://vimeo.com/250998475).

Figure 5. PE TA’s image-classification algorithm over-optimizes its computer-vision analysis, leading to incorrect classification of vehicles and people as animals (https://vimeo. com/213373126).