community portraits
An interactive artwork for Bedford Community Arts exploring the aesthetics of AI and emergent biases
Fergus Laidlaw, Arnab Chakravarty
.png)
The training of image generation models is optimised for 'beauty'
But what is beauty? These systems have become aligned with predominantly western beauty standards, through the curation of datasets and carefully prescribed reinforcement learning with human feedback (RLHF).
By drawing a comparison between what the image generation models can create and the set of community photographs from BCA, we can identify gaps in the the models 'knowledge'. More importantly we encourage visitors to try and remedy these gaps, giving them first hand experience with AI's bias and promoting critical thinking around these systems.
Highlighting the subtle biases and limitations of image generation models
The training of image generation models is optimised for 'beauty'
But what is beauty? These systems have become aligned with predominantly western beauty standards, through the curation of datasets and carefully prescribed reinforcement learning with human feedback (RLHF).
Highlighting the subtle biases and limitations of image generation models
By drawing a comparison between what the image generation models can create and the set of community photographs from BCA, we can identify gaps in the the models 'knowledge'. More importantly we encourage visitors to try and remedy these gaps, giving them first hand experience with AI's bias and promoting critical thinking around these systems.
We're using the Community photographs of Bedford Community Arts as a sample of real world photography, that captures diverse cultural iconography and imagery. By generating images that aim to replicate these photographs, we can identify gaps in AI's ability to generate certain features. Visitors are able to intervene and encouraged to try to generate images closer to the authentic photographs. Through this process the visitors experience first hand the difficulty in achieving representative imagery. At the same time, we gain a better understanding of what visitors feel are missing from the generative images. We aim to gain insight into both the inherent biases of the model as well as insight into what holds emotional and cultural value to the public in this context.
How it works
Design Process
We were intrigued by the disparity between AI-generated images and authentic human photography. Because of its training on commercial and stock image photography, AI-generated images often exhibit a heroic, well-constructed aesthetic, which fails to capture the nuances and diversity of human life. Our aim is to use the BCA community portraits as a repository of images that tell an authentic human story, and compare them against generative imagery to explore the shortcomings of generative imagery in its current form.
By focusing on these themes, we hope to uncover whether AI can be nudged into creating representative, imperfect images that resonate with authenticity. This exploration will highlight the contrasts between the polished perfection of AI art and the authentic irregularities found in real-world photography. Our goal is to challenge and expand the boundaries of AI's artistic capabilities, questioning the potential and limitations of machine learning in capturing the essence of everyday human life.
In the design of the objects themselves, we wanted to give life to the often abstract concept of AI, by housing it in the familiar nostalgic casing of an overhead projector. Although the personification of AI can be detrimental at times, leaning in here enabled us to foster engagement with the audience and critical observation about the way the system 'thinks'.
