Abstract Labels and the AI Hype in Computer Vision

What kinds of subjective, cultural labels are we asking AI models to assign to visual culture? How well do these work, and what biased visions of the world do they reproduce?

Situated Ground Truths and Plural Perspectives

What are we teaching AI as “truth” to emulate? How do we technically operationalize Haraway’s idea of “situated knowledge”–partial, situated perspectives–for AI?

Toxic Memes and Content Moderation

Memes and other viral cultural data can spread dangerous messages. But exactly what do we mean with online toxicity, and how are morality and normativity operationalized?

Depictions of the AI/Human Dynamic

Are we the chimpanzees of the AI era? No, but contemporary media is using evolutionary rhetoric to depict AIs as an “evolution” of human beings. Why and how?

LATENT IDENTITIES

#

Identity Labels in Latent Spaces: Generative AI  and Stereotypes

Investigating how Generative AI models learn and represent identity labels in their latent spaces, with a focus on detecting and mitigating biases and stereotypes.

Analyzing the types of prompts, concepts, and labels people use to describe their identities and comparing them with labels assigned by AI.

Latent Identities is a research direction that seeks to understand how Generative AI models learn and represent identity labels in their latent spaces. By investigating the ways individuals self-identify and the prompts and labels they use to describe their identities, researchers can uncover the elements that people consider vital but are often overlooked in AI training datasets. This information can be used to develop more inclusive and accurate representations of diverse identities, mitigating the biases and stereotypes that exist in current AI models.

​

The goal is to identify and challenge the stereotypical biases that are embedded in Generative AI models. By comparing self-definitions with labels assigned by AI, researchers can highlight the biases and inaccuracies that exist in these models. This comparative analysis will help researchers understand what AI gets “right,” what it gets “wrong,” and what it fails to consider, enabling them to develop more fair and inclusive AI systems.

​

One of the key aspects of this research is its focus on the representation of underrepresented groups. By examining the ways in which AI models learn to represent these groups, researchers can identify and challenge the visual stereotypes that are perpetuated through media usage and stock photo representations. This research can inform the development of alternative image banks and tools that are generated with citizen involvement, leading to more inclusive representations in mainstream media. Ultimately, I aim to create AI systems that promote diversity and inclusion, rather than marginalizing underrepresented communities.

TOXICITY

#

Operationalization of Toxicity and Literacy: Toxic Symbology in Memes

Developing AI models for detecting and analyzing toxicity in online platforms, with a focus on identifying patterns of toxic symbology in memes.

Advocating for digital literacy skills, particularly in relation to online platforms and social media, to promote a more positive and respectful online culture.

This research direction seeks to understand the complex and dynamic phenomenon of toxicity online. This research technically and theoretically explores the operationalization of concepts like toxicity, hatefulness, harmfulness, ethics, moderation, and extremism. By combining data-driven approaches with media studies insights, I aim to develop a deeper understanding of the ways in which AI models learn to recognize and mitigate toxic content.

​

One key area of focus for this research is the development of software that can detect and explain toxic symbology in online platforms. This includes not only identifying toxic content, but also providing insights into the underlying reasons why it is toxic. By examining cases of memes and toxic symbology, such as the extremist, racist, and other hateful memes that spread on platforms like 4chan and Reddit, we can gain a better understanding of how toxicity is perpetuated and how it can be mitigated. This research also involves experimental computer science work, including the development of Retrieval Augmented Generation (RAG) based systems for toxicity detection and explanation.

​

This research direction is also focused on the development of digital literacy skills, particularly in relation to online platforms and social media. By studying patterns of toxic symbology in memes, I aim to identify common themes and motifs that contribute to online virality. This information can be used to develop educational programs and tools that help users recognize and resist toxic content, promoting a more positive and respectful online culture.

VULNERABILITY

#

Empirical Ethics of Vulnerable Data Subjects: Using AI on Childfluencers

Using AI to analyze family vlog content, identifying patterns of potential exploitation to inform policymakers and the public about children as monetized data subjects.​

Balancing public interest with individual rights, ensuring that the use of AI is ethically sound and respectful of vulnerable data subjects, such as children.

This research direction is focused on the empirical ethics of data science when it comes to vulnerable data subjects, with a specific case study on the use of AI to analyze family vlog content for online child protection. This research aims to identify empirical patterns of potential exploitation on child influencers, using AI and other computational data science methods. The goal is to provide quantifiable evidence of overexposure and exploitation, highlighting potential harms and risks to children’s online presence.

​

By analyzing family vlog content, researchers can identify patterns of overexposure, including to harmful audiences. This information can be used to inform policymakers, advocacy groups, and the public about the normalization of children as monetized data subjects, prompting critical discussions on ethical dilemmas and potential harms. The research also aims to drive policy changes by supporting legal frameworks, such as the Illinois law requiring parents to set aside a portion of earnings for child influencers.

​

However, this research direction also raises ethical concerns, particularly with regards to the use of AI on vulnerable data subjects, including children. Part of this research direction is the development of empirical ethics frameworks to consider potential harms, such as a lack of explicit consent, emotional or social impact, normalization of surveillance, legal or platform-specific repercussions, privacy invasion, and misuse of findings.