Useful individual AI contributions do not necessarily add up to better collective work. I study properties that become visible across a collection of ideas, judgments, or decisions: the breadth of possibilities explored, the variety of evaluative perspectives, and whose input can influence what happens next.
Double compression in scientific ideas and reviews
Working manuscript · In preparation. With Matt Groh and Brian Uzzi, I compare 23 proposals and 85 expert reviews from a real bioscience grant competition with outputs generated by Claude, Gemini, and GPT under the same call and evaluation criteria. Computational experiments vary prompting conditions and use embeddings, diversity and coverage measures, topic modeling, and style-normalized comparisons to examine collections of proposals and reviews.
The working manuscript identifies two forms of compression. During generation, AI proposals cover broadly similar scientific territory but concentrate around common ideas, leaving some distinctive human proposals underrepresented. During evaluation, AI panels can cover much of the semantic territory of human reviews while producing more similar judgments, especially in their criticisms.
These results motivate evaluating AI by the range of ideas and perspectives it contributes to a collective process. The study compares human and AI outputs from one competition; it does not test human scientists working with AI or measure realized innovation. Effects vary across models, prompting conditions, and measures.
Evey Huang, Matt Groh, and Brian Uzzi. Double Compression: How AI Narrows Both the Generation and the Gatekeeping of Scientific Ideas.
Whose judgments shape the response?
DIS 2024 · Honorable Mention. My earlier research on computational approaches to online harassment examines the relationship between participation and influence. Through a scoping review and qualitative analysis of 17 papers, my coauthors and I studied whether tools reflected intended beneficiaries’ identities, definitions of harm, and preferred responses.
We identified gaps in the documented connection between those needs and implemented systems. People might supply information about harm while platforms retained authority over how it was classified and addressed. This work informs my broader concern with who can shape the categories, judgments, and actions embedded in organizational AI.
Evey Huang, Abhraneel Sarma, Sohyeon Hwang, Eshwar Chandrasekharan, and Stevie Chancellor. Opportunities, Tensions, and Challenges in Computational Approaches to Addressing Online Harassment.