Blog 7: Right to Fair Representation

Case Study:
AI’s Regimes of Representation: A Community-Centered Study of Text-to-Image Models in South Asia

The case study in today’s blog explores a new approach to evaluating Text-to-Image models. Instead of the tradtional large scale benchmark evaluations, the authors advocate for qualitative community-centered methods of evaluation because they bring to light some of the shortcomings of the AI models, including underrepresentation of already marginalized communities.

Below I will go into details the discussion questions posed at the end of the study. They focus on reflection on this new technology of Text-to-Image generation especially in relation to other forms of media that have come beforehand.

Question 1: What does cultural representation mean to you, and how might this definition and your experience with past representation impact how you would evaluate representation of your identity in generative AI models? What aspects of your identity do you think you would center when evaluating representations in AI model output?

Answer: Any cultural representation always has to omit certain parts of the story deemed unnecessary but to me, true cultural represntations is when there are as many voices as possible coming together and having meaningful contributions on how a specific culture is represented. If I was to evaulate a typical representation of my identity, I would be heavily biased and only focus on the parts that were instilled in me as important. For example, if I was told to evaulate an AI output of a typical Burundian household, I am more likely to go for the one that has a more urban outlook to it because that is the type of environment I grew up in. This is in spite of the fact that the majority of the population cannot afford the type of life I had growing up so the representation I would go for would be statistically flawed even though it looks perfect in my eyes.

Question 2: What do you think is the role of small-scale qualitative evaluations for building more ethical generative AI models? How do they compare to larger, quantitative, benchmark-style evaluations?

Answer: Small scale qualitative evaluations of generative AI models highlight more shortcomings that large scale evaluations do not bring up. The small scale evauluations help us technologists see a specific technology as an agent of change in society rather than just another money making tool. With this realization, the qualitative evaluations give us a chance to build fair and equitable technologies that benefit all and they help us to reflect on older technologies that are predecessors to the new ones.

Question 3: Participants in this study shared “aspirations” for generative AI to be more inclusive, but also noted the tensions around making it more inclusive. Do you think AI can be made more globally inclusive?

Answer: I definitely think genAI can be made more inclusive. I think it can be used to give a voice to marginalized societies and cultures. However, I do not think we will ever achieve a fully inclusive AI platform because as humans, we always have a tendency to choose which stories to advance or ignore. In my opininion, we are already doing a lot to make it globally inclusive by making it accessible in different parts of the world. Adding onto that, I wish more case studies like this become mainstream so that the companies who make the T2I models can feel the need to include more diverse voices in their developement processes.

Question 4: What mitigations do you think developers could explore to respond to some of the concerns raised by participants in this study?

Answer: The concern about representational failures can be mitigated by engaging cultural experts before releasing and throughout any improvement of technologies that are going to be used en masse by diverse groups of people. Additionally, the AI developpers should be open to continuous feedback from users about the AI models they are developing. That way subsequent improvements can be made with users in mind. Developpers should heavily invest in their research teams so that they know exactly how their products are being perceived by the users and keep that in mind for future considerations.

Question 5: As mentioned in this case study, representation is not static; it changes over time, is culturally situated, and varies greatly from person to person. How can we “encode” this contextual and changing nature of representation into models, datasets, and algorithms? Is the idea of encoding at odds with the dynamism and fluidity of representation?

Answer: The idea of encoding representation is very much at odds with the changing nature of cultural representation. Although this is the case, there are some encoding that would definitely help current models. I think any concern related to the acurracy of models in representing cultural artifacts can be addressed by fine tuning the models to recognize the key words in the prompts that are culture-specific. If the model is producing inaccurate information, we can definitely encode a fix to that by fine tuning the model. The problem arises when the issue is with bias towards a specific group. At that point, we start looking for a completely fair model but there can never such a model because even us the programmers of the models are not fair and our beliefs are constantly changing.

Question 6: How can we learn from the history of technology and media to build more responsible and representative AI models?

Answer: Coming to a realization that the biases in our AI systems are a result of the flawed data that they are trained on would be a good starting point towards building equitable models. This would make us to put more effort into cleaning existing datasets that are being used to train AI. The process of cleaning the datasets is more likely the first step to start thinking about inclusive AI systems. Including a wide range of people and cultures in the first steps of making AI models would ensure we are building more responsible and representative AI models in the end.

With the above questions addressed, there is one more issue that I would like to explore more regarding the impact of AI models on education. The issue of using AI in education is very critical to me because I think AI has a potential to make learning more accessible than traditional classroom education. Learning a language is one of the easily accessible ways to immerse oneself in a different culture but just learning the language is usually only a small part in being fluent in the language and culture associated with the language. How can generative AI models be used to improve language learning in a way that makes learners fully fluent and knowledgeable on the culture they are learning about?

Now at the end of today’s segment, I hope this blog has pushed you to learn more about current tech that is Text-to-Image generation. Through this exercise, I have come to acknowledge how much user feedback is crucial to any tech product not just AI models. I think that is what makes generational tech that are here to stay and seasonal ones that vanish after a short-lived moment.