No AI about us, without us:  Engaging Disability Communities in Meaningful Data Work to Improve AI Models

AI image generation is being rapidly adopted, despite the popular and academic literature demonstrating that its representation of people with disabilities, and other marginalized groups, is at best poor, and often offensive. As visual media shapes public perceptions and hence access to education and employment, disability communities are calling for more influence over their representation in these models. In this talk, I will discuss the design and evaluation of the Community Library Creator, which enables communities to define ‘good’ representation and embody that in a diverse, structured dataset that can be used for pre-training large image generation models.

Speaker’s Bio
I am a Sr Principal Research Manager in Equitable AI at Microsoft Research Cambridge. I co-lead the Teachable AI Experience team (TAI X) which aims to innovate new human-AI interactions that bring us to a more inclusive society. I believe strongly that we must innovate the machine learning techniques that we use in conjunction with designing new types of experiences. I hold a PhD in Computer Science from University of Cambridge (opens in new tab) and an undergraduate degree in Ethnomusicology from Barnard College, Columbia University(opens in new tab). I currently live with my family in Massachusetts.

How to Get There
Public transportation: Take the Red Line Kendal Square. Upon arrival to 1 Memorial Drive, show your ID at the desk and take the elevators to Floor M.

Re-inventing the Attention Machine
Today, algorithmic systems such as social media feeds and generative AI systems increasingly mediate human interactions and experiences. But interactions with these black-boxes reflect the worst of us due to impoverished objectives that amplify problematic content, induce algorithmic overreliance and monoculture. Existing attention-maximizing objectives optimize for engagement signals that are (at best) an indirect proxy of users’ preferences and underlying values. In this talk, I will discuss work on scaling the measurement of social AI systems to highlight the gaps in existing algorithmic objectives and propose new objectives for evaluating and steering social algorithms. First, I will show an approach for fighting the algorithmic amplification of misinformation on social media through a decentralized accuracy prompt that increases user discernment and in turn steers the algorithmic objective. Next, I will discuss how to support a wider set of human values in social media algorithms by introducing a personalized method for measuring human values in social media posts. I will then use this tool to show that X’s in-production feed algorithm is misaligned with its user’s values. I will close with implications for human creativity and content production. Together, this work underscores the promise of new forms of interactions with algorithmic systems that center human agency for prosocial outcomes.

About our speaker
Ziv Epstein is a Postdoctoral Associate at the Schwarzman College of Computing at MIT. In his research, he focuses on translating insights from design and the social sciences into the development of sociotechnical systems such as generative AI and social media platforms. Ziv has published papers in venues such as the general interest journals Nature, Science and PNAS, as well as top-tier computer science proceedings such as CHI and CSCW. His work has also received widespread media attention in outlets like the New York Times, Scientific American, and NPR..

How to Get There
Public transportation: Take the Orange Line to Ruggles or the Green Line to Northeastern University. Upon arrival to Khoury College, West Village H, take the elevator to the third floor, then turn left and proceed to Room 366.

Vibing with your AI
This talk examines how people understand and relate to conversational AI, tracing a shift from metaphor-based mental models to deliberately designed AI personalities. Drawing on years of research published across ACM CHI, CSCW, TOCHI, and CUI, it shows how users’ perceptions of voice and conversational interfaces have evolved from tools or simplistic agents to social actors with distinct “vibes.” I demonstrate how early metaphors, such as AI as a child, butler, or friend, laid the conceptual groundwork for today’s personality-driven design enabled by large language models, where traits such as agreeableness or openness can be systematically tuned. The talk concludes by surfacing both the design opportunities and ethical risks of this shift, including trust calibration, emotional manipulation, and transparency, and argues for more responsible, human-centered approaches to AI personality design.

About our speaker
Smit Desai is an Assistant Professor in the College of Arts, Media and Design (CAMD) at Northeastern University, with joint appointments in Art + Design and Communication Studies and an affiliated appointment in the Khoury College of Computer Sciences. He directs the Conversational Human–AI Interactions (CHAI) Lab, which investigates how people understand, trust, and collaborate with conversational AI—and how these systems can be designed to augment human capabilities and support meaningful interactions.

Smit’s research bridges theory and design practice: he examines users’ mental models through methods such as metaphor analysis and persona design, and applies these insights to build next-generation conversational agents. His current projects span healthcare (AI assistants for serious illness conversations in emergency departments), aging and well-being (voice-first reminiscence systems like Memory Box), and everyday collaboration (LLM-powered chatbots that adapt to user personality). Across these efforts, his work emphasizes human-centered AI principles, transparency, and the social consequences of giving machines personalities and roles.

His work has appeared in premier HCI venues including CHI, CSCW, and TOCHI. He serves on the organizing committee for the ACM SIGCHI Conversational User Interfaces Conference (2024-2026), ACM CHI (2026), and leads workshops on ethical conversational persona design at major HCI conferences. Smit received his Ph.D. from the University of Illinois Urbana-Champaign.

Naviagation: Enter the building through this gate and take left.

Nearest T station is Ruggles on Orange line and Northeastern University on Green line