This paper explores differences between the commercial image recognition of Google, Amazon and Microsoft in categorizing images on religion. Based on our findings we discuss the empirical and ethical implications of the potential use of these services for the study of religion.
Paper long abstract:
Since the so called visual turn in the construction of meaning in media-saturated societies, the study of religion is also faced with new methodological challenges. Development of AI technologies and readily available automated image recognition services would appear to provide one solution for the study of large amounts of image data currently available in digital form.
Visual recognition services, however, like other machine learning systems are not just tools, but also systems of representation that take part in constructing perceptions of religion as they organize and categorize the data available to them. This paper studies the empirical and ethical implications of employing commercial visual recognition systems to the study of religion.
Empirically, we conduct a comparison of categorizations of religion related images by three commercial image recognition services: Google Vision AI, Microsoft Azure Computer Vision and AWS Rekognition. Our study focuses on how the services represent religion by labeling the images and on the differences and similarities between the three services. Based on our findings it is safe to say that in the context of religion the image recognition systems of Google, Amazon, and Microsoft are in many ways problematic in ethical terms and (re)produce cognitive and cultural, representational biases.