=Paper=
{{Paper
|id=Vol-2491/abstract113
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-2491/abstract113.pdf
|volume=Vol-2491
|dblpUrl=https://dblp.org/rec/conf/bnaic/Schraffenberger19
}}
==None==
https://ceur-ws.org/Vol-2491/abstract113.pdf
Investigating people’s attitudes towards AI with a
smart photo booth
1 1,2 2 3
Hanna Schraffenberger , Yana van de Sande , Gabi Schaap , Tibor Bosse
1
Artificial Intelligence, Radboud University, Nijmegen, The Netherlands
h.schraffenberger@ru.nl
j.vandesande@donders.ru.nl
2
Communication Science, Radboud University, Nijmegen, The Netherlands
g.schaap@maw.ru.nl
t.bosse@ru.nl
Extended abstract. With the increasing impact of AI in people’s everyday lives,
multidisciplinary research on the public perception and understanding of AI is more
important than ever. The development and implementations of smart technologies
and AI in our society raises many urgent, societally and culturally relevant
questions: How do people think and feel about AI? Can they recognise, understand
and evaluate the processes involved in AI-driven decision-making? What mental
models do they use when interacting with AI systems — are these models similar
to those of humans or more like models of machines? Yet, research that attempts to
answer these questions is scarce.
In this paper, we emphasise the need to address such questions concerning
AI attitude, the public understanding of AI and Human-AI Interaction from a
multidisciplinary perspective. We combine expertise from both social science and
computer science, and take steps in the direction of such a multidisciplinary
approach towards AI problems. We present a first pilot study, illustrating how AI
systems can be used to research people’s AI attitude. The proposed setup takes the
form of an intelligent photo booth called “misidentify.me” capable of detecting
humans and identifying a range of objects and animals.
During this pilot, visitors of a film festival were challenged to fool the AI
and take a selfie on which the intelligent photo booth would not identify them as a
human being. This idea is rooted in the observation that machines are getting
smarter and the consequent question whether we humans are still smart enough to
fool them. More specifically, we were interested whether and how people’s interac-
tion with AI systems and their ability to outsmart them affect their AI attitude.
The photo booth was realized with web technologies, which makes it pos-
sible to also offer the experience online and to extend the study with an online ex-
periment in the future. The core functionalities are realized with a combination of
p5.js (see https://p5js.org/) and ml5.js (see https://ml5js.org/). In particular ml5’s
version of PoseNet was used to estimate a “humanness” score, whereas its Mo-
bileNet model for image classification was used to determine a “somethingness”
Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License At-
tribution 4.0 International (CC BY 4.0). This paper is a compressed version of the 2019 xCoAx paper:
Schraffenberger, H., Van de Sande, J., Schaap, G., & Bosse, T. (2019) Can you fool the AI? Investigat-
ing people’s attitudes towards AI with a smart photo booth. xCoAx 2019 proceedings,
http://2019.xcoax.org/pdf/xCoAx2019-Schraffenberger.pdf.
score. The two scores were compared to determine the ultimate label assigned to
each selfie, tagging it either as a “human” or labelling it with the name of the iden-
tified object/animal.
In order to investigating people’s attitude towards AI, we used a quasi-
experimental 1 factorial (identified as human yes/no) pretest-posttest design, in
which two dependent variables – thoughts and feelings about AI (Cronbach α = .67
- .95) were measured both before and after interacting with the smart photo booth.
Participants were allocated in two groups depending on whether or not the AI iden-
tified them as humans. After cleaning up the data, we ended up with a small sample
size (N=25). Because of this small size, no inferential analyses were conducted.
However, the descriptive statistics show some interesting things. In table 1 the gen-
eral feelings and thoughts on AI are displayed. As one can see, the overall attitude
is more on the negative side, with a score well below the midpoint of the 0-100
scale. In addition, the pretest and posttest data suggest that once participants had
endured interaction with the photo booth, their feelings and thoughts became even
more negative.
Table. 1. Descriptive statistics (N=25)
Feelings Thoughts
M(SD) Min-Max M(SD) Min-Max
36.73 1- 40.79 16.56-
Pretest
(15.63) 63.80 (10.57) 57.44
32.70 .87- 35.97 .89-
Posttest
(16.39) 64.0 (14.90) 56.89
Comparison of means between the groups show participants have a more
positive attitude towards AI when they succeeded in fooling the AI (Feelings:
M=34.68, SD=17.85, Thoughts: M = 37.88, SD = 15.5) than the participants that
were successfully identified as human by the AI (Feelings: M=30.56, SD=15,12,
Thoughts: M=33.90, SD=14.57).
The many conversations and informal observations revealed that our in-
stallation creates AI awareness and fosters dialogue and reflection. At least some
people seemed rather surprised and impressed by the capabilities of the AI system
and/or the outcome – even though similar AI technology is already a part of many
people’s everyday lives. This raises the question whether people are aware of the
AI systems that are already part of their lives and highlights the need for research
into what we call “AI literacy” – the question whether people can recognise, under-
stand and evaluate the involvement of AI systems when using technology. Based
on exploratory observations, we conclude that multidisciplinary research into AI
attitude, Human-AI-interaction, AI literacy and the social impact of AI in people’s
immediate vicinity is a promising and much needed research direction.