J.D. Zamfirescu-Pereira

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2025-175

October 21, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-175.pdf

Generative AI systems enable creation—of text, images, code, and more—through humanlike language interfaces, promising to transform how we design and create. Yet these interfaces often lead users to apply intuitions about understanding and reasoning that then lead them astray. In this dissertation, I examine these misaligned intuitions and how they can get in the way of better leveraging unique Generative AI capabilities, such as proposing large sets of diverse alternatives. Through case studies in chatbot authoring and interactive programming, I illustrate (1) how humans approach instructing LLM behavior by drawing on human-human instructional experiences---and how those approaches can fail to serve users' goals; (2) how designing for generative AI systems is uniquely challenging; and (3) ways that computer program design tools can leverage AI's strengths while addressing these tensions, through explicit structures and proactive exploration.

Advisors: Björn Hartmann

---

BibTeX citation:

@phdthesis{Zamfirescu-Pereira:EECS-2025-175,
    Author= {Zamfirescu-Pereira, J.D.},
    Title= {The Language of Creation: Understanding and Overcoming Alignment Challenges in Designing with Generative AI},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {Oct},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-175.html},
    Number= {UCB/EECS-2025-175},
    Abstract= {Generative AI systems enable creation—of text, images, code, and more—through humanlike language interfaces, promising to transform how we design and create. Yet these interfaces often lead users to apply intuitions about understanding and reasoning that then lead them astray. In this dissertation, I examine these misaligned intuitions and how they can get in the way of better leveraging unique Generative AI capabilities, such as proposing large sets of diverse alternatives. Through case studies in chatbot authoring and interactive programming, I illustrate (1) how humans approach instructing LLM behavior by drawing on human-human instructional experiences---and how those approaches can fail to serve users' goals; (2) how designing for generative AI systems is uniquely challenging; and (3) ways that computer program design tools can leverage AI's strengths while addressing these tensions, through explicit structures and proactive exploration.},
}

EndNote citation:

%0 Thesis
%A Zamfirescu-Pereira, J.D. 
%T The Language of Creation: Understanding and Overcoming Alignment Challenges in Designing with Generative AI
%I EECS Department, University of California, Berkeley
%D 2025
%8 October 21
%@ UCB/EECS-2025-175
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-175.html
%F Zamfirescu-Pereira:EECS-2025-175