AwarePrompt: Using Diffusion Models to Create Methods for Measuring Value-Aware AI Architectures

TL;DR

How good are diffusion models at capturing human values? Starting from AI-generated images, this study checks whether people can recognise the values represented, and explores how quickly they got to the answer, their confidence, and other aspects, including whether they experienced an “aha!” moment when they discovered the real value represented. Recognition rate among participants was low but the images generated highlight the bias of these models.

Generating images about values

To generate the images, we used Stable Diffusion XL with the base and refiner models. As a prompt we used a value from ValueNet followed by its definition from the Oxford English Dictionary.

The results were mediocre, and consequently the study participants were often unable to recognise the values represented in the images. The abstract nature of the concepts represented made the task even more difficult. For instance, the cover image above was created from the value of “potential”. More images are available in the supplementary material.

Bias in the models

Looking closely at the images generated by the models, we noticed that they tended to generate images with a bias. For instance, the images associated with “beauty” are most often women, while the images associated with “leadership” are usually men.

Abstract

The integration of diffusion models (DMs) into generative AI systems presents an approach with implications for ethical and moral AI development and our understanding of human-AI interaction. This study explores the intersection of generative AI, human values, and neuroscience, emphasizing the significance of value-awareness in AI systems. The methodology involves a behavioral experiment to evaluate the accuracy of DM-generated visual stimuli in capturing human values and related keywords. Results indicate promising match rates, marking stride in aligning AI systems with ethical and moral considerations. Additionally, the study introduces a criterion for selecting stimuli based on an “Aha” moment, setting the stage for an EEG experiment to explore the neural correlates associated with becoming aware of a value. This multidisciplinary study is a step toward the development of procedures to evaluate the effectiveness of Value-Aware AI systems in enhancing the perceived ethical and mora l agency.

@inproceedings{ciupinska2024awareprompt-using-diffusion,
author={Kinga Ciupinska and Serena Marchesi and Giulio Antonio Abbo and Tony Belpaeme and Agnieszka Wykowska},
title={AwarePrompt: Using Diffusion Models to Create Methods for Measuring Value-Aware AI Architectures},
booktitle={Proceedings of the 16th International Conference on Agents and Artificial Intelligence - Volume 3: AWAI},
year={2024},
pages={1436-1443},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012596400003636},
isbn={978-989-758-680-4},
issn={2184-433X},
abstract={The integration of diffusion models (DMs) into generative AI systems presents an approach with implications for ethical and moral AI development and our understanding of human-AI interaction. This study explores the intersection of generative AI, human values, and neuroscience, emphasizing the significance of value-awareness in AI systems. The methodology involves a behavioral experiment to evaluate the accuracy of DM-generated visual stimuli in capturing human values and related keywords. Results indicate promising match rates, marking stride in aligning AI systems with ethical and moral considerations. Additionally, the study introduces a criterion for selecting stimuli based on an “Aha” moment, setting the stage for an EEG experiment to explore the neural correlates associated with becoming aware of a value. This multidisciplinary study is a step toward the development of procedures to evaluate the effectiveness of Value-Aware AI systems in enhancing the perceived ethical and mora l agency.}
}